To be clear, believing in an Indium Phosphide shortage is different from just being bullish on optics broadly.

Optics is an end market. Indium Phosphide is a material input.

The intensity of the material usage in the end market can either go up or down. In addition, there are also supply dynamics. Demand could go to the moon but there is no shortage if supply goes to Mars.

There are also many layers to the Indium Phosphide supply chain, as I cover in this piece:

• Indium Phosphide substrate

• MOCVD/Epitaxy machines

• Laser device fabrication

Jason's Chips

CPO Supply Chain 101

The most important concept to understand when analyzing the CPO supply chain is the power of e x p o n e n t i a l…

Read more

a month ago · 114 likes · 14 comments · Jason's Chips

Some of which will be in much greater shortage than others.

The key to… bottleneck investing (sigh) is to figure out which layer controls the gating input to the final unit produced. It could be the upstream most layer, downstream most layer, or anything in between. There is a common perception that the more upstream something is, the more of a bottleneck it is, but that relationship is not necessary.

Our agenda today is simple. We start by highlighting why Indium Phosphide demand is going hyper-exponential beyond EMLs and regular commodity CWs powered by two main products:

• Ultra-high power CW lasers for CPO

• Indium Phosphide PICs

Then, we’ll build supply and demand models for all three layers of the stack: indium phosphide substrates, MOCVD machines, and laser devices.

Finally, we’ll finish with how this translates into the financials for the most important company in each layer.

Contents

1. Ultra-High-Power CW Lasers for CPO

2. Indium Phosphide PICs

3. InP Substrate Demand Build

4. InP Substrate Supply Build

5. MOCVD Tools Supply/Demand Model

6. InP Device Supply/Demand Model

SemiAnalysis Giveaway: The first person to subscribe using the paywall on this post will be selected to receive one free month of SemiAnalysis (sent to your email).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Ultra-High-Power CW Lasers for CPO

UHP CPO lasers are a topic that I’ve covered extensively. So much so that you are probably tired of me talking about it. Welcome to my nth laser lecture.

There are two types of lasers:

1. Vertical-Cavity Surface-Emitting Lasers (VCSELs, pronounced vik-sel) emit light from vertically… from a cavity… on the surface. They use GaAs, aren’t great for long reach because they can’t emit O band light, and fail at higher data rates

2. Distributed Feedback (DFB) lasers emits light horizontally through a bunch of tiny mirrors that bounce it in a feedback loop to get a really concentrated and clean beam. They use indium phosphide! These are the primary CPO lasers.

For CPO, the DFB light source must be remote for the purpose of avoiding heat and maintaining serviceability.

The only way to ensure that the light survives the coupling loss from this journey is to make it high power.

Why not multiple lasers instead? To maintain signal integrity, the light traveling from the laser to the optical engine has to stay in a consistent orientation — consistent polarization. This means every laser requires polarization-maintaining fiber, which is expensive, and the process of attaching it is painstaking and yield-sensitive. We must minimize the number of PM fiber connections per system at all cost, which means using only one laser.

These lasers are also continuous wave (CW). DFBs can either be CW or EML.

CW lasers are lasers that just stay on. It produces a continuous stream of light, like a light bulb or flashlight. A CW laser can provide a steady beam that gets modulated (gets the signal put into it) somewhere else.

EMLs (electro-absorption modulated laser) are lasers plus a built-in light switch. EML stands for electro-absorption modulated laser. The laser creates the light, and an attached modulator rapidly changes that light on/off or high/low to encode data.

Die Sizes

Now comes the fun part: die size estimation. We know that the three main types of lasers used for optical data center optical communication are

1. Low power CWs (30-70mW, used in Sipho 1.6T transceivers)

2. EML (used in all sorts of transceivers)

3. Ultra high power CW (400mW+, for CPO)

So to gauge demand for InP, it’s time to ask: how much Indium Phosphide does each type of laser actually require?

Our base unit will be the 100G EML because it ships in the most volume today. This is the optical equivalent of Toyota Corollas or Big Macs. Or football fields.

EMLs are very low power, even lower power than low power CWs. However, they have a built-in modulator. Therefore, low power CWs are smaller than 100G EMLs.

200G EMLs are obviously larger than 100G EMLs.

Ultra high power CW lasers on the other hand, are MASSIVE. You need a physically larger gain region to generate all of that light. This makes it 5-10x larger than a 100G EML.

This massive die area consumption is what makes UHP CWs an order of magnitude more InP heavy than existing products on the market. They are the first of our two InP demand drivers.

Indium Phosphide PICs

The second InP demand driver is Indium Phosphide PICs. These are not as widely discussed as ultra-high power CWs, so let’s take a moment to understand what they are and why they are needed.

Optical communication is simply transmitting data through light. There are two ways to encode data in light (known as modulation): IMDD and coherent (no not the company).

IMDD is basically flashing a flashlight on and off. Only encoding signals in brightness.

Coherent modulation is like having one of these:

It encodes signals in brightness, phase, and amplitude.

The benefit of coherent modulation is that it makes it much easier for the receiver to detect the signal, therefore allowing the it to survive a greater insertion loss (higher link budget). This is why coherent is used for long-reach telecom and scale across.

On the other hand, it is much more expensive and complicated, as you can imagine.

Scale-out in the data center is now running into the limits of IMDD’s link budget. This is because:

1. The links are too long

2. Optical circuit switching (OCS) is being adopted. Each time light passes through an OCS, it loses 1.5 to 3 dB of signal.

Therefore, scale-out is moving to adopt a light version of Coherent known as Coherent-Lite (Yes i know it sounds like a Coherent and Lumentum merger. Who named these modulation schemes?).

Coherent light modulation requires a far more complex optical device (InP PIC) than IMDD. It makes a 100g EML look like a 2x4 Lego brick.

All of this extra stuff is to modulate phase and amplitude.

Note that not all of this is Indium Phosphide. You can have some parts of it be silicon photonics, which makes it a lot cheaper. The standard version, which has the transmit on InP and receive on SiPho, has a substrate consumption intensity of approximately 100x that of 100G EMLs. This is what we’ll be using for our models.

InP Substrate Demand Build

Next, I will share my supply and demand models for:

• indium phosphide substrates (AXTI)

• MOCVD tools (Aixtron)

• indium phosphide devices (Lumentum, Coherent, Nokia, etc.)

We start with an overall demand build for 4-inch indium phosphide wafers that informs every model.

We’ll look at some pretty charts that show the equilibrium balance/imbalance intuitively. The numbers from every use case (demand) and every supplier (supply) shall be color coded. My research is really lit.

Read more

A lot of people enjoyed it so I will be copying the format one for one because I am very uncreative.

Introduction

An OCS is just a mirror box. Light goes in, bounces off mirrors (or other stuff), then goes out in a separate direction.

The most important concept to understand when analyzing the OCS market is the power of a t t a c h r a t e.

The switch itself has a bunch of ports. Higher port count is referred to as higher radix.

Now the unit of measure for OCS shipments isn’t the number of boxes because boxes can vary in size. Big boxes cost more, small boxes cost less. Instead, it is the number of ports as the cost of a box generally scales with its port count.

Attach rate can therefore be defined as the rate at which ports are attached to XPUs. 2:1 means 2 OCS ports for every accelerator chip.

And for every OCS use case, the attach rate is rising.

On top of that, OCS itself is a new technology that is in the early-mid stage of its S-curve.

Therefore, this is an excellent market because its growth rate scales with three vectors: AI growth, adoption rate, and attach rate.

OCS growth rate = AI growth * adoption rate * attach rate

Contents

1. Switch Primer

2. OCS vs Electronic Packet Switch (EPS)

3. Spine Switch Replacement

4. Google TPU Superpod OCS

5. Scale-Up OCS

6. Types of OCS

7. OCS Market Model

8. Lumentum

9. Coherent

10. Hidden Small-Cap OCS Vendor

11. Final Thoughts: Better Than CPO?

SemiAnalysis Giveaway: The first person to subscribe using the paywall on this post will be selected to receive one free month of SemiAnalysis (sent to your email).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Switch Primer

A switch is very simple.

Data centers must be organized in a hierarchical pattern otherwise every GPU has to talk to every other GPU which means links scale at O(n^2) which would fill up the entire data center with wires pretty quickly which is generally considered to be not good.

In order to form this network, traffic must be routed, and it is switches that route the traffic.

OCS vs Electronic Packet Switch (EPS)

An Earnings Per Share (EPS) (lol) is the standard electrical switch that is currently used almost everywhere.

It receives an optical signal as an input via transceiver, converts it to an electrical signal, reads and routes that signal via switch ASIC (logic chip), converts it back to an optical signal via transceiver, and sends it to its destination.

An OCS, on the other hand, just bounces the light off of a mirror.

So think about intuitively what the trade-offs are.

If you have to convert a signal from optical to electrical and back to optical (OEO), that certainly takes time. It also takes energy. The switch itself is also active, which takes even more energy. So generally the big problems with an EPS are latency and power.

It’s not like this is a small difference btw. OCS is practically free compared to an EPS. Lumentum on their R300 switch:

The R300 has demonstrated 98% reduction in switch latency vs. an OEO packet switch solution.

The power requirements of OCS is also less than 10% of an EPS.

But obviously an EPS is better in certain ways. Otherwise, why is it the most widely used solution?

Think about what happens when the signal exits the switch system via transceiver. It gets received, then re-sent, meaning that the signal coming out the other side is a fresh beam of light. On the other hand, in OCS, the light bounces off the mirror and loses a bit of signal and isn’t re-amplified again.

In addition, intuitively speaking, the light also has to travel twice the distance because it never gets to stop at the switch. Think of it like taking a direct flight instead of a layover.

As a result the signal suffers more insertion loss which means that when the signal finally arrives at its end destination it is a lot more mangled. Specifically, if the insertion loss from distance (in EPS setup) is x and the switching loss is y, the total OCS insertion loss is 2x + y.

The other tradeoff is that moving mirrors to redirect connections takes time. Not that much time, only milliseconds. But EPS can make fresh routing decisions for every single packet. This is the problem behind the core reason OCS was not a thing before AI, which I will lay out for you now:

Traditional cloud traffic is mouse flows.

These are tiny, short-lived, and unpredictable connections (a search query and API call, a checkout). You need a switch to actually make decisions for every packet. The connections change too fast for those micro mirrors to handle.

AI workloads are elephant flows.

Training points the same exact clusters of accelerators to each other and leaves the connections alone for hours.

OCS bears would say that as we move to inference, we move back from elephant flows to mouse flows, as each individual user query is distinct. This may be true at the leaf level. However, at the spine level (where OCS sits), the network traffic is still classified as elephant flows, as the only data that passes is prefill to decode KV transfer and not the individual tokens.

Spine Switch Replacement

Now, let’s talk mirror box use cases. There are three of note that we will discuss today:

1. Spine switch replacement

2. Google TPU Superpod OCS

3. Scale-up OCS

Spine switch replacement refers to the replacement of the spine switch.

This is the perfect (and generally the canonical) use case for OCS. At the spine switch level, almost all AI workloads are elephant flows. Spine OCS is nearly free compared to a spine EPS, so free in fact that it lowers the spine capex by 30% and power consumption by 40%.

The thesis is simple: current spine OCS penetration outside of Google is under 5%. That number is gonna increase. Each EPS that gets replaced with an OCS adds to the yearly CAGR. This is an adoption rate story.

However, we still have to get around this insertion loss issue. The current link budget of 4 dB can barely take one OCS hop.

So what’s actually going to spark the adoption?

Coherent-Lite Transceivers!

Optical communication is simply transmitting data through light. There are two ways to encode data in light (known as modulation): IMDD and coherent (no not the company).

IMDD is basically flashing a flashlight on and off. Only encoding signals in brightness.

Coherent modulation is like having one of these:

It encodes signals in brightness, phase, and amplitude.

The benefit of coherent modulation is that it makes it much easier for the receiver to detect the signal, therefore allowing the it to survive a greater insertion loss (higher link budget). This is why coherent is used for long-reach telecom and scale across.

On the other hand, it is much more expensive and complicated, as you can imagine.

Therefore, scale-out is moving to adopt a light version of Coherent known as Coherent-Lite (Yes i know it sounds like a Coherent and Lumentum merger. Who named these modulation schemes?).

As these transceivers get adopted, the link budget doubles or triples, allowing for plenty of room for OCS hops. In fact, we will have so much link room that the light can handle two hops at once. Meaning it’s possible to build networks with a spine and super spine architecture, both of which are OCS.

Google TPU Superpod OCS

Google’s architecture is weird.

Nvidia (NVL72) is very simple and intuitive and easy to understand. Each rack of 72 GPUs is its own scale-up domain (one compute and memory “unit”). The scale-out fabric connects these racks together via transceivers, leaf switches, and spine switches.

Google’s interchip interconnect (ICI) is its scale-up network, but unlike NVL72, it isn’t one rack. Instead, it is a hodgepodge of 9,216 TPUs which share one gigantic memory domain but are connected with “racks” that are each a 4x4x4 cube of 64 TPUs (copper inside the cube, transceivers connecting cubes to each other). That’s why it’s called a superpod.

The pod is shaped like a 3D torus, so basically a giant chip donut.

Each one of these super pods has 48 OCS. Therefore, there is a deterministic port-to-TPU ratio of 1.5:1. So in essence, OCS scales with TPU shipments.

But it gets better. There’s an attach rate story too.

This is based on the excellent work by FundaAI.

FUNDA

Deep|AI Infra 2026: Shifting from "Brain Power" Competition to "Whole-Body" Evolution

In 2026, the focus of AI development has pivoted from chasing high benchmark scores to pursuing AI Agents capable of multi-step reasoning and autonomous action. This infrastructure arms race is undergoing a transformation akin to biological evolution. If an AI system is viewed as an evolving organism: the GPU/TPU represents the calculating brain; Memory and Storage serve as the memory carriers for experience and context; the CPU acts as the hands coordinating tasks; while Optics and Networking function as the limbs supporting systemic data flow and response sensitivity. Under the framework of the Agent Scaling Law, the core bottleneck is no longer just the FLOPS of a single chip (brain power), but rather the communication efficiency (limbs), the memory wall (memory), and the Total Cost of Ownership (TCO…

Read more

2 months ago · 74 likes · FUNDA

A 3D torus is built for an all-reduced communication pattern, which is common in training. Each chip mostly talks to its nearest neighbors, and results ripple through the network like a class of middle schoolers passing notes to each other until everyone reads it.

However, in inference for mixture-of-experts (MoE) models, the communication pattern changes to all-to-all. Every token gets dispatched to separate experts living on chips scattered anywhere in the pod.

Instead of middle schoolers passing notes to each other, it’s like every kid has to talk to every other kid at the same time. But for our purposes, let’s now use a better analogy. Think of the data center doing an all-to-all communication like a city. Every household drives to the far side of the city at once.

Now let’s try to make the problem clear. Imagine in this city there is a central intersection. When there are thousands of cars trying to cross to the other side, there may or may not be mild traffic.

In the data center, this is known as the bisection bandwidth, which becomes a critical spec when all-to-all communication is necessary.

Eventually, this problem needs to be fixed. You’ve thought of every possible solution, but nothing seems to get over the fact that in a two-dimensional plane, there’s only so much space for roads, and you just can’t fit all the cars.

So you hire a genius city planner. By virtue of having infinite tax revenue, this person decides to give everyone a helicopter. Now, there’s not a central intersection. Instead, there’s a central airspace, but because you’ve added a dimension, the bisection bandwidth is no longer a constraint at all.

Bisection bandwidth for N chips given n dimensions scales with the following equation:

bisection bandwidth ∝ N^(n−1)/n

so

N^2/3 in 3D → N^5/6 in 6D → N¹ as n→∞

This is essentially what Google is thinking of doing: moving from a 3D torus to 4D or higher. And each time you increase the dimensionality, the OCS ports per TPU mathematically increase.

This will take a while. It’s not like it’s going to happen in 2027. The patents for a 4D torus are filed, but in TPU v8, Google is sticking with a 3D torus. This is likely something on the same timeline as scale-up TPU, like 2029 to 2030 or later. But it has the potential to double the attach rate in the medium term (3:1 in a 4D torus) and in the very long term, a ceiling of 5x or more (to 10:1 ratio).

Scale-Up OCS

Optical scale-up, the act of connecting the GPUs inside of a rack with fiber instead of copper, is the crown jewel of the CPO market, but for OCS, it is only an interesting call option.

This is because, unlike CPO, scale-up OCS is completely optional!

The idea is to place OCS directly into a rack to replace NVlink switches, but the thing is there’s no law of physics forcing the regular switches to fail in the optical scale-up world. It’s merely a choice by the vendor for what kind of switch they want to use. This is why I won’t be focusing on this market or modeling it very intently.

Types of OCS

As we discussed, the two main drawbacks of optical mirror boxes are the speed of redirecting connections and the insertion loss. These two drawbacks also serve as the main trade-offs between the different types of OCS. We will discuss liquid-crystal and MEMS in this section because they are the two relevant types of OCS for today’s data center deployments. Piezoelectric and Sipho will be saved for later — you will see why.

The simplest way to think about the difference between MEMS and liquid-crystal is that for MEMS you reflect light off of stuff (mirror), and for liquid-crystal you steer light through stuff (medium). All of the spec differences stem from this.

Here is a tier list of the relevant mirror box specs.

Let’s start with insertion loss and cost.

Liquid-crystal and MEMS are very similar cost-wise, but MEMS decisively wins in insertion loss. This is probably the biggest and most important difference between the two types of OCS. MEMS has a 1.5 dB insertion loss versus 3 dB for LC. Decibels are measured on a logarithmic scale too, with 10 dB being one order of magnitude, so 3 dB means that you lose half your signal in one hop. It’s pretty bad.

From Lumentum’s OCS webinar in 2025:

This capability is embodied in the R300, which currently supports up to 300x300 ports with a typical insertion loss of 1.5 dB or less. And a note about the importance of low-loss performance –this is a critical feature, as it is one of the key technical hurdles to the introduction of OCS technology into hyperscale environments. This is seen in the requirement to support additional OCS loss within transceiver link budgets – around 4dB for FR4/DR4 and 6dB for LR4/ER4.

The core mechanism behind this difference is that, for a liquid crystal, you must first split the light, then have it pass through an active medium where some light is absorbed, and then recombine it. Both the splitting and the passing-through-a-medium create insertion loss.

Meanwhile, MEMS is literally a mirror. A literal mirror.

Liquid crystal makes big claims about their reliability versus MEMS, some of which I think are a bit overstated. For example, there is one claim that MEMS wears out after three years because of its moving parts. This fundamentally overlooks the fact that even though the parts move, they don’t slide against each other, and there is, in fact, no friction. Moving parts doesn’t mean sliding parts that wear out.

Google’s MEMS fleet has been in production for a decade with tens of thousands of units. If MEMS switches failed after three years, they would probably notice and say something about it.

However, the reliability of liquid crystal is still much better.

This is due to the system having no moving parts (mirrors) and operating with much lower voltage components.

Now let’s discuss redirection time. This is what dictates how long it takes for the switch to change your port connections.

Currently, MEMS and liquid crystal are about tied at a range of 10-100 ms (liquid crystal slightly slower actually).

However, a new type of liquid crystal in development, ferroelectric LC, switches in 10-100 microseconds instead. This is a a next-generation material for Coherent. As a result, switching time is decisively won by liquid crystal.

But there’s a reason why I put redirection time in C tier.

If you even decide to use OCS, you only have to change port connections when you reconfigure your network. This happens on an order of minutes to days. Microseconds or milliseconds simply do not matter.

Finally, let’s discuss install base.

MEMS wins here for a very simple reason: Google.

Google is by far the largest employer of OCS in the world and has had a fleet of them operating for a decade. They are actually the pioneers of their technology, producing most of their switches in-house (under Palomar brand) before the massive wave of AI demand forced them to externalize. And it’s all MEMS.

Read their Mission Apollo paper if you’re interested in learning about their deployments. But notice how MEMS-centric it is.

This is S tier. This matters a lot. Proven reliability at scale is a defining requirement for data center procurement officers.

Overall, MEMS is way better.

But this doesn’t mean that liquid crystal won’t get any adoption. It is still viable technology and the specs here don’t decide adoption as a binary. If MEMS vendors’ capacity is sold out, hyperscalers will happily buy liquid crystal. Having an OCS at all is what matters more.

Below, I will share my detailed model for the OCS market with accelerator shipments, TPU shipments, external merchant share, ASP, and port ratios informed by industry checks. I will discuss Lumentum and Coherent as players in this market and how I expect it to be split between the two of them, with models for each. Finally, I will share a hidden small-cap OCS vendor.

Read more

In terms of the YTD the reason why my monthly and weekly portfolio reviews have sort of different numbers is only because of the measurement technique since I’ve been adding capital throughout the year.

(imagine you are adding capital every day to an index fund, and the index is going up, your returns on your brokerage, AKA a MoM or MOIC, would be lower than the returns of the index, AKA capital agnostic CAGR.)

Measuring it the other way would be 195%.

I honestly don’t know which is the better measure, as both are gameable, so whatever. I am up anywhere from 150-200% in less than 1/2 a year without leverage which is not bad in terms of stock picking abilities.

There is a lot of change this time. Specifically, there are 4 new positions a bunch of exits and I’ve made one company ~25% of my portfolio. Also featuring CPO discussion, a cool podcast, sell-side research, macroeconomic theory (not microeconomics this time), and the interesting stuff that happens when you try to DCF a co-location provider.

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Read more

The takeaway from all of this news aggregated together is actually massively bullish and incremental positive to optics. NOT negative!! I’ll explain why.

All it takes is some very simple logical deduction.

Contents

1. SemiAnalysis Note

2. FundaAI Note

3. What Is NPO and Why NPO

4. Does NPO Use InP & ELS

5. Lumentum at Mizuho

6. Conclusion

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

SemiAnalysis Note

The (very sad) title of their report was “Powered Down, Lights Off.”

Read more

This is the view from the top of the Taipei 101. It’s really a beautiful city.

Anyways Computex basically takes place inside of this building and there’s a whole bunch of booths and it is an absolute meat fest (very packed).

There was a ton of alpha to be found by looking at the racks and talking to the various interesting people at the booths.

We will first highlight various interesting exhibits from investors POV. We will then talk about CPO. The main takeaway I’ve had is the end-state of scale-up CPO (and it is incredibly important!)

Contents

1. Immersion Cooling

2. Fuel Cells

3. Hiwin (Robotics Supplier)

4. DRAM

5. Kyber Racks

6. Astera Labs

7. Nvidia RTX Spark (Laptops)

8. MediaTek MicroLED

9. Scale-Out CPO

10. Scale-Up CPO, Takeaways for Stonks, and Lumentum EPS

SemiAnalysis Giveaway: The first person to subscribe using the paywall on this post will be selected to receive one free month of SemiAnalysis (sent to your email).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Immersion Cooling

There is this company called QuettaFlow that literally submerges GPUs in liquid to cool them.

Isn’t that so sick?

Unfortunately, you can’t put an NVL72 rack in this liquid, as you may imagine. This is only for small chunks of compute, so basically enterprise on-prem.

The liquid isn’t water (duh). It is a non-conductive fluid that is a by-product of oil. You can plug stuff into the rack while it’s submerged. I actually touched it. It is very oily and greasy.

The benefits are that it is much cheaper (40-50%), much more compact (no IT room or raised floor), and saves a ton of electricity.

They’re private, of course, but fascinated to see where this tech can go.

Fuel Cells

Funny story.

We were at the Vertiv booth and asked the guy there about this 800VDC power module and what it does.

It is a combined transformer and rectifier that turns AC medium voltage from the grid into 800 VDC.

This is very big and expensive. I think they said like $3M ASP.

So we asked the guy, “What if you just had 800 VDC natively? Then you wouldn’t need this box, right?”

He said “Well, where would you get that?”

LOL

Jason's Chips

Bloom Energy Thesis (Free)

As a Substacker, I am obligated to say that I am releasing this for free to provide an example of what kind of research you would receive with a paid subscription…

Read more

2 months ago · 92 likes · 20 comments · Jason's Chips

Then we visited a booth for a company called Delta Electronics. They are competitor to Vertiv and Schneider. Basically, liquid cooling and stuff.

But did you know that they also have an SOFC solution?

Unlike Bloom, which uses ceramic plates, Delta licenses Ceres’ IP, which actually uses steel. Obviously specs are inferior to Bloom currently, so it’s more of an overflow capacity play.

They are currently shipping small volumes (30 to 50 MW) and installing them in Taiwan. They’re not fully independently scaled commercially, also, as the purpose of these fuel cells is more to provide a full portfolio of power solutions complementing their Vertiv-like products.

Ticker is 2308 on the Taiwan exchange.

Hiwin (Robotics Supplier)

Hiwin is a Taiwanese screw maker that supplies into robotics. They make actuators (components that help robots move). Did you know that actuators actually make up most of a robot’s BOM?

Interestingly, most of their sales are actually to North America, to companies like Tesla. This is important because a lot of this stuff comes from China, so American companies would find suppliers like this very attractive as a geopolitically safe option.

In the middle is their best component, which is a linear actuator. They are better (higher ASP, less commoditized) than the rotary ones on the sides.

I looked into them a bit further. Turns out that 50% of their sales are actually to the semis market. So yeah, here’s a company that has exposure to both robotics and semis. If you are interested, the ticker is 2049 on the Taiwan exchange. However, as you might imagine, screws are quite commoditized. They only have about 10% market share. LOL. And they trade at 40x multiple on sell-side estimates. I honestly would have expected a way lower multiple for a business like this.

DRAM

How to invest in semis 101:

1. Look at the PCB.

2. Identify what has the most surface area.

3. Buy.

One recurring theme that I noticed is how literally every single tray that you see at Computex has a shit ton of DRAM. It sounds obvious, but you really need to see it for yourself to intuitively grasp how much DRAM is in everything.

For example, this is the Vera CPU compute tray.

Those SOCAMM modules are stacked too, so it’s actually even more surface area than it appears to be.

Being bullish on CPU proliferation for agents is almost just as much bullish on DRAM. As it would drastically increase the total base of compute (accelerated or not) to which memory is attached.

Kyber Racks

Kyber racks are so cool.

GPUs go in vertically from the front. NVSwitch modules go in horizontally from the back. They meet in the middle via orthogonal midplane. Super compact! Save space!

This means you can fit 144 GPUs in one rack. It’s very beautiful and elegant. There are no cables, either.

However, the volumes at which they can actually be shipped for Rubin Ultra are constrained by the supply chain. Kyber racks require high-grade (M9) CCL (copper-clad laminate), because the per-lane rates are increasing, the layer counts per tray are increasing, and the cableless design pushes these signals through the midplanes and orthogonal backplanes, putting even more pressure on the CCL for M9 CCL. Traditional glass fiber won’t cut it. You need quartz fiber, which is much harder to manufacture and whose supply chain is much less mature.

If it does ramp, this has significant implications for future data center design. If data centers become much more compact, 800 VDC might be able to span the whole shell without needing to go through lossy transformer and rectifier steps (800 VDC cannot travel long distances). This is incredibly bullish for Bloom, as they output 800 VDC natively. Meaning that the TCO of a fuel cell becomes materially more competitive if smaller form factor racks and data centers get wider adoption. This deserves an entire article of its own.

Astera Labs

Asteria Labs makes connectivity chips, like retimers (which are chips that clean up signals and retimes it to the clock) for copper communication (CXL and PCIe).

Astera Labs is a huge Trainium supplier. The 2025 10K says that one customer was 70% of revenues. I wonder who that is!

(This is also why Astera is Gavin Baker’s biggest position. Gavin is a believer that Trainium is an even better ASIC than TPUs.)

Think of Trainium as the Amazon basics of XPUs.

SemiAnalysis

AWS Trainium3 Deep Dive | A Potential Challenger Approaching

Trainium3: A New Challenger Approaching…

Read more

7 months ago · 144 likes · 3 comments · Dylan Patel, Daniel Nishball, Wega Chu, Myron Xie, Ivan Chiam, Clara Ee, Cheang Kang Wen, Wei Zhou, Jeremie Eliahou Ontiveros, and Tanj Bennett

They are literally air-cooled and also use PCIe for scale-up like regular computers. Sometimes I am flabbergasted that compute which doesn’t use NVLink… works.

But did you know that Astera Labs actually has an LPO solution?

(An LPO is just a transceiver without a DSP. Think of it like the transceiver is to LPOs as an AEC is to passive copper.)

They don’t actually make the LPO module itself. They sell the LPO switch ASIC and the software that modulates the SerDes signal.

This can be used for PCIe scale-up!

An LPO has the same latency as passive copper DACs because there is no retiming. Compare that with an AEC, which has active data processing components. It is pretty much strictly better, as the LPO can reach up to 50m while the DAC fails within a few.

Here is an eye diagram (PAM4, you can tell because there are three eyes).

I have no opinion on this stock besides that it is a good tracker for Trainium capex.

Nvidia RTX Spark (Laptops)

Nvidia’s new agentic laptops. The only thing I have to say here is that isn’t it so weird that they are turned off and facing away from the crowd? I wanted to try them, but they don’t even work. This is just fueling more criticism of their ARM + Windows design.

MediaTek MicroLED

For some reason, MediaTek is obsessed with microLED for scale-up CPO. When I was at their booth, a ton of space was dedicated to this (very widely hated) solution while their custom ASICs weren't present at all.

Interesting.

Micro-LEDs are an alternative to lasers. A lot of photonics engineers don’t like them because laser light is basically very concentrated (coherent, highly directional) and it is also a single wavelength (linewidth anyone?) while LED light is very diffuse.

Like literally think about the LEDs in your house versus a laser that you can shine.

As a result, it is much harder for the receiver to pick up the signal, and thus you have much higher insertion loss and bit error rate. This is a pretty fundamental difference between the two technologies that is sort of glossed over in these booths.

But they do have a lot of claimed advantages. They are much more power efficient, have superior thermal resilience and reliability, and can be integrated directly onto the CMOS transceiver chip.

Micro-LEDs are very short-reach. I was at the Molex booth, which is an ODM for micro-LEDs, and they said that it only reaches up to several meters, meaning that this is a potential scale-up CPO candidate.

Scale-Out CPO

There are two types of scale-out CPO switches: InfiniBand (Quantum-X) and Ethernet (Spectrum-X).

Let’s start with the InfiniBand one.

There are 18 ELS (external laser) modules (down on the bottom of top pic) and 4 switch ASICs with 18 optical engines surrounding each (see below, optical engines circled in red).

Now for Spectrum-X. There are 16 ELS… but only one ASIC!

That one ASIC has 32 optical engines around it. So less than 1/2 optical engines total, and 2:1 OE to ELS ratio rather than 4:1.

The ELS are also liquid-cooled, unlike quantum X where the ELS are air-cooled.

Here is the higher radix one. Basically 4 switches put together.

And the Spectrum-6 rack, which is four high radix switches put together (16 normal switches). This supports eight NVL 72 racks.

Scale-Up CPO, Takeaways for Stonks, Lumentum EPS

Below the paywall, we will discuss how my models (CPO ELS/OE builds) have changed after looking at Nvidia’s reference designs, the crazy mock scale-up CPO rack I saw (and the face-slapping implications for the scale-up CPO endgame I realized), and takeaways for optical stonks.

I will be providing the answer to the question I posed in this tweet.

We will be talking a lot about ratios and numbers and assumptions because small changes to these can either blow out or tank out-year revenues for optical companies.

And I will update my Lumentum model & EPS estimates.

Read more

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

I am a wizard and absolutely nailed their earnings with my tweet.

They did, in fact, announce that their "opportunity pipeline" expanded again, while being simultaneously missing on revenues.

Revenue Misses

I will bash them in other areas, but I am actually not going to do so here. I completely buy this excuse, because if this was any other company that I cover, I would have given them the benefit of the doubt as well. This honestly doesn’t matter as much at all for whether or not their CPO shenanigans will take off.

Opportunity Pipeline

Bashing time! If I were sell-side, I would be absolutely dishing out questions to management over this. Every single mechanism for defining the opportunity pipeline is up to the discretion of management. Not a single one is objective. Not a single one relies on contracts.

What does meaningful discussions mean? What does good conviction mean? What does determining that Sivers’s technology is a good fit mean? No one knows. The management can define those however they want. This doesn’t necessarily lead to the opportunity pipeline being an inflated metric per se. Management could be even more conservative if they chose to. It just means that this is a metric that is manipulatable and should mostly be treated as noise and not signal. I’m sure if you had the right incentives, you can easily manipulate this pipeline to be an entire order of magnitude larger or smaller than it currently is.

The biggest red flag is that Vickram Vathulya does not address any forward revenue guidance.

This question is a clear invitation to share numbers, and it was dismissed with a one-sentence answer on their themes.

Their end markets are growing way faster than this BTW. CPO should be triple-digit growth.

No qualifications, no production volume orders, nothing changed.

Competition

Here they just completely conceded that they are worse than Lumentum in every way. All they needed to mention, like most companies when they’re asked this kind of question, was at least one area in which they are competitive and to talk that up as the more important trade-off. You can see Coherent doing this all the time with their 6-inch wafers, even though they are uncompetitive everywhere else, but they still have something to talk about.

The only response Sivers could give was “…but the pie is growing!” TAM doesn’t pay the bills if you can capture none of it.

Margins

This is very good. COHR GMs are barely above 40% for reference. Achieving this would actually be remarkable. Credit where credit is due.

Conclusion

This last question in the Q&A captures perfectly the essence of Sivers today.

Maybe because they’re not actually a volume supplier to the market since they haven’t qualified anywhere?

Everything else is noise until they give purchase order announcements with clear numerical amounts, and provide year-by-year revenue guidance. Until then, they’re a startup, and nothing has changed.

Both on the good side and bad, for example, the revenue miss today doesn’t matter at all, and neither does the “opportunity pipeline.” I guess that just means I’m saying that earnings don’t matter at all. LMAO.

See you next quarter!

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Contents

1. Headline Numbers and Press Release

2. Memory Downturn

3. Price Action

4. Capex, Working Capital, and Cash Flow

5. Capacity and Contribution Margin

6. Physical AI

7. Photonics

8. Overall Thoughts (and My Positioning)

Read more

It was only a few months ago when photonics was a niche, technically complex subsector of a subsector of a subsector of the AI infrastructure trade, covered by small, undiscovered, and specialized analysts. Today, it shares an investor base with 2021 NFTs and 2024 Solana meme coins.

But unlike crypto, the technology and fundamentals are still there!

So it’s truly become a fascinating contradiction where on one side there are companies with real technology and an almost perfect market position. There is some hype around them, but that hype is not nearly enough, and they still present a great opportunity (see ). On the other side there is Sivers.

Sivers is special. Before all the speculation promoted by unethical large social media accounts, there was a legit investment thesis! Credit to the several analysts that found it pre-cult.

:

Moody

Sivers Semiconductors: A Swedish Technology Dwarf Punching Far Above Its Weight in Three of Today’s Hottest Tech-Trends

When searching for the next semiconductor highflyer, most investors look to California, Taiwan, or South Korea. The fact that a Swedish company with just 125 employees, headquartered in the Stockholm suburb of Kista, is simultaneously fighting on three of the most exciting technology frontlines of our time – AI data centers, satellite internet, and auto…

Read more

4 months ago · 44 likes · 3 comments · Moody

:

Silicon Matter

Sivers Semiconductors: a $130m mcap co-packaged optics, LiDAR and SATCOM play

Intro I’ve been sitting on this piece about Sivers Semiconductors for a while, but ahead of this years’ OFC conference I believe it’s prudent to push the publish button. For those unfamiliar with Sivers, it’s a relatively asset-light and Sweden-based technology company led by CEO Vickram Vathulya, with annual revenues of $33 million and approximately 130…

Read more

4 months ago · 70 likes · Silicon Matter

:

Gaetano

$SIVE Deep Dive

Sivers Semiconductors ($SIVE) is one of the more interesting smaller names in this part of the market right now. I have received more requests on $SIVE than any other company in this sector…

Read more

3 months ago · 46 likes · 9 comments · Gaetano

Now I will pretend that Sivers is $130 million market cap again, and pitch the original thesis. It is by doing this that we can fully understand why it has broken after a 20x, unlike names like and .

We will be analyzing Sivers (and finding out why it is insanely overvalued) through a purely fundamental lens. I will not comment on the market structure or price action or ethical/legal concerns, or criticize any of the individuals that have promoted this name. I will instead save that for my tweets (lol).

Oh, by the way, for anybody asking me to short, I wish I could, but I can’t because IBKR won’t let me. My account is restricted to cash only.

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Contents

1. The Original Thesis

2. The Valuation is Absurd

3. Sivers 101: Opportunity Pipeline & NRE

4. The (Fabricated) Bull Case

5. Why Is Sivers Fabless?

6. There is No Moat

7. The Wrong Architecture

8. Conclusion

The Original Thesis

Sivers is a company with two segments: wireless and photonics.

Wireless makes up 70% of 2025 full-year revenues. Photonics (which includes both data center photonics and LIDAR) makes up the other 30%.

The company is cheap. With a $130 million market cap and around $33 million of revenue, Sivers trades at almost exactly 4x EV/revenue.

Profitability is constrained as EBITDA is currently negative, but that’s the whole point. A company that is cheap on EV/revenue with negative EBITDA that has a hidden gem segment (photonics) can show massive operating leverage if that hidden gem segment explodes.

This is precisely what sell side projects. While the legacy wireless segment is forecasted with pretty unexciting growth (5G and radar and satellite stuff), photonics, which includes both LIDAR needed for physical AI and, most importantly, CPO for data centers, carries the company to double-digit sales CAGR and their 20% EBITDA margin goal.

With revenues doubling by 2028 and EBITDA margin inflecting positive, suddenly you go from 4x EV/revenue to 2x EV/revenue and 10x EV/EBITDA. By 2028, photonics would make up half of the entire company’s sales, making this an attractive bet.

Not much has to go right here. At a market cap of $130 million, the thesis can simply rely on management delivering on their own stated goals.

The Valuation Is Absurd

The sell-side price target is less than 10% of the current market price, which I find absurdly funny. (It is really funny.)

I think the ridiculousness of this valuation is best explained by my tweets during my retard crusade last weekend.

And the way it got to this valuation is also extremely unhealthy.

As you can imagine, if the stock is nearly 20x more expensive at a $2.3 billion valuation, the old thesis doesn’t work anymore.

This is not true for all companies. Some, like or , sell a raw material or commodity whose supply/demand scarcity dynamics have no natural ceiling. This means that one thesis (memory ASP go brr) can justify nearly any valuation.

But for other companies competing in the 10-bagger arena, the thesis must evolve as the valuation gets more stretched.

Sivers could rely on a simple CAGR back at $130 million. At $2.3 billion, they now trade at more than 230x their disclosed 2025 photonics revenue. Keep in mind that the photonic segment isn’t purely data center revenue, either. It also includes LIDAR, and they do not break down sub-segment granularity, so in reality, the actual multiple is higher. I am also only able to use trailing multiples, as the company has not disclosed a single DC photonics volume purchase order or given any DC photonics revenue guidance. They are basically a pre-qualification startup.

All management disclosed is their goal of compounding their revenue at 25 to 30% CAGR, expanding to 65% gross margin, reducing R&D to 20% of revenue, and having an EBITDA margin around 30%.

“Based on the cost structure and margin profile of the business, we would expect to reach cash flow breakeven at annual revenues of approximately $50 million-$55 million, and 65% of revenue stemming from products. This we are targeting in roughly two years’ time. Our long-term ambition is to sustain 25%-30% revenue CAGR, with fabless financial KPIs at around a 65% gross margin, R&D costs of around 20% of revenues, and an EBITDA margin around 30%. We are striving for world-class performance because we believe we have the technology to achieve it.”

This is no longer aggressive enough to save the stock.

Now, we have to evaluate whether there can be an order one or two order of magnitude step function growth from a sudden qualification breakthrough along with massive purchase orders (that can actually be manufactured with their current owned capacity and WIN allocation) that can even remotely justify the current valuation levels. This also means that Lidar and Wireless are pretty much irrelevant because the only segment that could be explosive enough to drive a fundamental breakthrough that justifies the current valuation is the massive CPO market. Other factors that people usually consider “bullish” like Chips Act funding (which is actually for their wireless segment for defense) and a Nasdaq listing (with like $16 million of inflows or something) also, don’t really matter at a $2.5 billion market cap.

If photonics breaks, the entire story breaks too.

Sivers 101: Opportunity Pipeline & NRE

Every earnings call, Sivers presents a KPI known as their Opportunity Pipeline. Their Opportunity Pipeline right now is $453 million, which is up 64% in 2025.

It sounds really big and cool because it takes the place of a number of other companies’ reports, which is very important: the backlog. However, the opportunity pipeline is not a backlog at all, not even close, and treating it as a backlog is extremely misleading.

The Opportunity Pipeline is literally everyone that they’re talking to. If they’re having a conversation, it’s part of their Opportunity Pipeline. For it to actually convert into contracted backlog, they need to move into evaluation and sampling, which is still very far away. Then get a design win and finally make it to production orders. I’ve never seen any other company in the world report the numbers this way. It makes it extremely inflated.

In my estimation, only about 10 to 20% of the Opportunity Pipeline actually converts to backlog. Which is around $68 million of their $453 million opportunity pipeline.

I started from the idea that the $453m opportunity pipeline is a five-year, non-binding funnel, not a one-year order book. If all of that were close to backlog, Sivers would already have several years of contracted revenue sitting behind a company that only did about SEK 306m / ~$30m of revenue in 2025. That does not fit the disclosures: contract liabilities were only SEK 11.6m, management says the pipeline is non-binding, and much of Photonics is still sampling/qualification. So the true committed layer has to be much smaller than the headline pipeline.

Then I sanity-checked it from the other direction: Sivers still needs enough near-term committed or near-committed work to support 2026 revenue around SEK 360-376m, plus named programs like CHIPS, Tachyon, Doosan, ALL.SPACE, LiDAR, and FWA. That gets you to a few hundred million SEK of credible revenue-converting opportunity, but not billions. Converting that into pipeline percentage: 10% of $453m is ~$45m, 15% is ~$68m, and 20% is ~$91m. That range maps to roughly 1.5-3.0x FY2025 revenue, which feels plausible for a mix of signed work, near-term NRE, product orders, and late-stage customer forecasts. Above 20-30%, the implied committed revenue starts looking too large versus what they actually disclose. Hence my practical estimate: 12-20%, with ~15% as the center.

The next Sivers term we need to discuss is NRE.

NRE means Non-Recurring Engineering Revenue. This is customer-funded R&D. The customer pays you to do R&D, and you book revenue. That’s basically what they do to get ciphers to customize their IP to the customer’s needs.

Now, most of their revenue is NRE, both wireless and photonics. That means they’re pretty much a startup at this point. This isn’t product revenue, and it’s not sustainable. Honestly, this would be a bear point at a much lower valuation, but at a valuation this high, this doesn’t even matter. They’re essentially a very frothy VC startup at this point, so NRE or not, what actually matters is if the product revenue can 100x with some massive qualification. Who cares if the tiny revenue base isn’t sustainable? That’s not what is making it a multi-billion dollar company anyways.

The (Fabricated) Bull Case

The few individuals who pump this stock focus on two main things.

This picture expresses the first one.

Basically, “omg $SIVE is in conversation with Ayar, Poet, LightMatter, and Optical Startup #463 and the ex-director-of-shits-and-giggles of one of these startups once had dinner with Matt Garman’s guinea pig, which means that they are supplying lasers to every single hyperscaler. $SIVE to the fucking moon”

This is completely and utterly wrong because, as I talked about before in the backlog section, conversations don’t mean shit without qualification and production orders. It’s not the hard part. It’s not what’s valuable. I don’t give a shit who supplies what to whom. If the product is uncompetitive, the buyer will find another supplier. If you have no customers, build a great product and the customers will come. What’s valuable is having a good product that passes qualification and performs well and then having the manufacturing expertise to ramp it at high yields and high gross margins. I can write a cold email and say I’m in conversations with Ayar Labs tomorrow if I wanted to. And not only that, the architecture of that startup has to actually be viable too (looking at you POET why do you even think you can beat hybrid bonding).

The second is possibly worse, which is that Sivers is a “chokepoint” or “bottleneck.”

(And no, a chokepoint is not different from a bottleneck. They’re the same thing.)

Sivers is literally the opposite of a bottleneck. They’re a fabless startup that people are calling a bottleneck.

Sivers hasn’t qualified anywhere. They are not choking off anyone. It’s like saying that the bottleneck for AI inference is Cerebras (lmfao). To find the real bottleneck for inference, you must look at who has majority share (Nvidia), then who owns the hard assets behind that majority share (memory and TSMC N3). The equivalent for optics here is Lumentum and AXTI.

A bottleneck must have both hard physical assets and substantial majority market share. Sivers has neither.

Why Is Sivers Fabless?

Today, Sivers is a fabless company. They have a facility in Glasgow which produces pilot volumes for R&D purposes, but in order to scale up to production volumes, they have WIN Semi as their contract manufacturer.

However, it wasn’t always this way. Turns out that in the past, Glasgow was pitched as a fab that could do up to $150 million in revenue. Yes, that is something around ten times their current photonics revenue base. It was only around the end of 2024 to the beginning of 2025 where this changed.

What happened? My best guess is yield.

Making lasers is really fucking hard. InP is brittle. Aixtron’s management said that Lumentum is their best customer in terms of their ability to use MOCVD machines to make lasers and even the best player out there only achieves 50% yields (on EML). That means one in every two dies is binned!

Manufacturing is brutal. Manufacturing is the moat, and that moat is what kept Sivers out of making their own lasers. That means that what they have left, which is just design IP, is a lot less valuable.

There is No Moat

Technology analysis will always be superior to supply chain analysis.

Here’s where I want to introduce you to a surprise. Sivers has no moat. Turns out that Lumentum also makes these laser arrays, and in fact Lumentum’s specs are even better, offering up to 100 mW per channel at 50°C, way higher than Sivers’ 65 mW per channel (with unspecified temperature).

As in classic Lumentum style, they also disclose way more other specs like spacing uniformity, channel power deviation, and reliability results. They literally do it better in every way.

This leads us to two main conclusions:

1. Laser arrays are not proprietary to Sivers. They don’t have a moat here that makes them better at making this specific technology.

2. Regular ultra-high power 400 mW lasers that send one powerful beam of light and split it, instead of using WDM and an array of lasers, are the superior default architecture.

The Wrong Architecture

Let’s zero in on the second one a bit further.

Lasers for CPO must also be placed far away from the optical engine.

Because this laser is so far away, in order for the light to reach the optical engine, it needs to survive a lot of coupling loss meaning that the light gets much less powerful every time it has to go through fiber and whatnot.

There are two ways to feed light into a CPO system.

The first is the ultra-high-power approach. You start with one very powerful continuous-wave laser outside the optical engine. That light enters the CPO package, survives the coupling loss, and then gets split into multiple lanes. This is the Lumentum-style architecture. The whole point is simple: start with so much clean light that even after losses and splitting, each lane still has enough photons to carry data.

The second approach is WDM, or wavelength division multiplexing. WDM just means using different colors of light. Instead of one very powerful laser, you use a laser array: many small lasers on one chip, each at a different wavelength. One color carries one data stream, another color carries another data stream, and so on. This is what Sivers is pitching.

WDM is a real and useful optical technique. But it solves the wrong problem for our beloved CPO application.

WDM solves a bandwidth density problem. It lets you push more data through fewer paths by stacking multiple colors together (and each data stream is one color).

UHP solves a more important problem: giving you way more starting photons.

CPO does not care about multiple colors. It fails if the light is too weak by the time it reaches the optical engine. Every connector, fiber attach, splitter, coupler, and waveguide eats optical power. If the source starts weak, adding more colors does not fix that. You just have many weak colors instead of one usable color.

In addition, Lumentum’s regular UHP architecture does not need WDM to push enough data. It can scale bandwidth through spatial parallelism: more lanes, more fibers, more waveguides, more parallel optical paths. That may be less elegant than WDM on paper, but it focuses on the one thing customers actually need first: actually having photons per lane after all that coupling loss.

WDM adds its own complexity. You need wavelength locking, thermal control, filters, tighter testing, more channel matching, and more ways for one bad channel to ruin the package. In CPO, the binding constraint appears to be optical power, reliability, and qualification.

So the conclusion is very clear!

Use many smaller lasers with different colors < < < Use much stronger lasers and split the light.

And this isn’t a Lumentum vs. Sivers thing where Lumentum wins the market with their architecture. It is that Lumentum has both architectures. They’re the leader in ultra-high power CW, and they’re also the leader in WDM laser arrays, but look at what they chose to ramp for production volume. What is their Greensboro fab going to churn out? UHP CW lasers!

Sivers is in the right market, but it is targeting the wrong architecture. Laser arrays may be useful in R&D programs, lower-loss systems, or niche WDM architectures, but the default CPO production ramp is being built around ultra-high-power external light. That is where the volume is.

Conclusion

Let’s recap. Sivers used to be an undiscovered small cap with a respectable thesis, but after a 20x return facilitated by large social media accounts (unethically and illegally) promoting the ticker, the valuation has gotten absurd. The only way to save the stock is through its optical division gaining a foothold in the CPO market, which is much less likely than what the cult is hoping for.

The cult believes Sivers will gain a foothold because they are in conversations or have remote connections to hyperscalers through various optical startups. This is wrong because what actually matters is its technology, and it is precisely technology where Sivers falls apart.

Sivers is fabless because they lack the manufacturing expertise in the very difficult process of indium phosphide fabrication. Which means they lack a moat, as they are skipping the hard part to do the easy part.

Sivers owns no exclusive technology, as Lumentum makes the same exact laser arrays, but better.

And finally, but most importantly, Sivers is targeting the wrong architecture entirely. Lumentum has a choice between UHP and WDM, and they made a choice to ramp the exact thing that is the opposite of what Sivers is doing, for volume production.

Sivers Semiconductor is the most overvalued stock in optics.

The last of my “big 5” (LITE, AIXA, INTC, BE) I haven’t started to super-intensely cover. But it was one of the first ever chip companies that I discovered and invested in!

It’s also up 10% today and 200%+ YTD and I’ve only ever added. I am proud of this one.

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Read more

However, if our end goal is to invest in AI and semiconductors are only a means to that end, I believe it’s ignorant not to look further downstream. This is exactly why I wrote my $100 trillion dollar Anthropic article.

Jason's Chips

Why Anthropic Is a $100T Company

ANTHROPIC IS THE MOST VALUABLE COMPANY IN THE WORLD, AND IT’S NOT EVEN CLOSE…

Read more

a month ago · 42 likes · 18 comments · Jason's Chips

Model labs are an even sharper expression of one’s belief in AI, because they generate the tokens themselves. Plus they capture both the upside from token value creation and from Moore’s law-driven cost reductions. I think they are the greatest business models (both in and out of AI) of all time.

But this is already well known. People are selling their kidneys for Anthropic stock. I am only saying that model going to be even more valuable than people predicted. But people are definitely already VERY excited. And when there’s that much excitement, alpha is harder to find.

But funny enough, by going just one step further downstream, to the thousands of regular businesses consuming the tokens Anthropic sells, all of that excitement ends. Here’s where you can be a true contrarian. Meaning possibly much more asymmetry.

This is the Shiller Cyclically Adjusted P/E Ratio, or CAPE.

Today CAPE is at 42x. We are in unprecedented territory that has only ever been reached at the height of the dot-com bubble. We’re even higher than in 2021, which involved a speculative frenzy and preceded a brutal year-long bear market.

Every single time the Shiller CAPE was elevated, a crash followed. It is the world’s most reliable bubble indicator. Even if we don’t crash, these high valuations can be back-solved into a terrible forward CAGR.

It creates the perfect excuse to ignore stocks on the index level, and therefore the perfect mechanism for markets to underprice the singularity.

Tokenmaxxing is about to agentmog the Shiller CAPE-cels.

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Contents

1. Microtokenomics

2. Why The “Intelligence Crisis” Won’t Happen

3. Timing and Revisions

4. The Trade

Microtokenomics

In every business or market, there is value creation and value capture. It’s a fancy way of saying consumer surplus and producer surplus, where value creation is the sum of all surplus and value capture is simply producers’ portion of it.

In AI, we can think of value creation as the value of the work that AI agents do for end consumers (how much benefit I get from tokenmaxxing) and value capture as the price charged by the model labs for inference (which also flows down to Nvidia, TSMC, the memory makers, ASML, etc.).

The higher the value capture compared to value creation, the more valuable that AI labs and semiconductors should be relative to the rest of the economy, and vice versa.

To analyze how big the value capture is compared to the value creation, we can look at microtokenomics. There’s not actually a thing called microtokenomics lmao this is just what I call looking at the value per token generated on a per-task basis.

SemiAnalysis did some amazing work on this. This is possibly one of the most impactful SemiAnalysis articles I’ve ever read.

SemiAnalysis

AI Value Capture - The Shift To Model Labs

A day in AI now feels like a year in any other industry. Model releases, software breakthroughs, and hardware improvements are compressing multi-year cycles for any other industry into weeks. Over just the past few months, agentic AI has crossed a real inflection point, driving a step-change in the value of tokens while software and hardware improvements have sharply reduced the cost of generating them…

Read more

2 months ago · 273 likes · Daniel Nishball, Dylan Patel, Cheang Kang Wen, Crystal Huang, Max Kan, Ray Wang, Myron Xie, Zane Fong, and Clara Ee

Here is the most important quote from the article. This states very plainly an idea that I’ve believed for a while but struggled to articulate.

“End users are enjoying a productivity bonanza - tasks that used to take tens of person-hours costing thousands of dollars can now be accomplished in minutes with a just a few dollars’ worth of tokens. This huge surge in revenue and margins is because the value of tokens being created is dramatically improving businesses. For example, SemiAnalysis has reached as high as $10.95 million dollar annual spend rate on Anthropic Claude tokens, but the value we derive allows us to outcompete all our competitors and gain market share.”

Later in the article, SemiAnalysis goes on to demonstrate traditional knowledge work tasks that would have taken several hours of traditional analyst labor, costing from $150 to $500, being able to be automated with a token cost of $2 to $8. They included tasks like:

• pulling 5 years of financials and emailing the results with Excel tables attached

• searching past conferences for optical networking and OCS trends

The ROI on tokens spent in these examples is over 50x.

So yeah, the ROI question is solved.

This has allowed Model Labs to expand their gross margins from the 30s and 40s all the way up to 70s and soon moving into the 80s as they take advantage of their pricing power. At the current point in time, if the ROI on tokens spent is over 50x the cost of the tokens, producer surplus is about 2% of the pie, while consumer surplus is 98%.

That orange sliver is the entire AI token supply chain that we are all invested in, everything from the Model Labs to the fabless chip designers to the foundries to the memory makers to the equipment vendors to the optical networking providers and to the power generators. All of that is 2% of the pie. The 98% gets spread out across every other firm in the economy. For every trillion dollars added to Anthropic’s market cap, $50 trillion or more should be added to the market cap of the S&P 500.

People see the ROI question mostly as a question for the AI infra value chain, but but it’s actually far more important for the economy. Negative ROI on AI means that every dollar spent reduces aggregate earnings for the S&P 500, while positive ROI means that every dollar spent adds to Aggregate Earnings.

In our case ROI is extremely positive. Every dollar of revenue for a model lab is an order of magnitude more value created for the economy (in the form of aggregate earnings). In addition, this aggregate earnings boost definitionally grows at the same exponential clip as the model lab revenues as AI penetrates deeper and deeper, meaning that earnings growth for the S&P 500 should accelerate beyond what historical norms suggest.

This is why the Shiller CAPE is flawed. Of course, stocks, which are supposed to price in future earnings, look funny based on past earnings if the future earnings are all of a sudden growing way faster than before!

Why The “Intelligence Crisis” Won’t Happen

This famous and very well-written article by Citrini single-handedly made consensus bearish on AI-impact-on-stocks.

Citrini Research

THE 2028 GLOBAL INTELLIGENCE CRISIS

Preface…

Read more

4 months ago · 9205 likes · 90 comments · Citrini and Alap Shah

Today, after seeing more empirical data, I am vastly more bullish, and the reason is a sort of a Jevons paradox but for labor.

Nathan put it well in his podcast, the AI Daily Brief.

“And what was interesting in all of this is that both the promise and the fear of AI, the promise of AI that would reduce how long it took to do your work so you could go enjoy more leisure time, and the fear of AI that would negate your value as a worker.

We’re both very far away from the lived reality of the most advanced users.

In fact, instead of finishing your work day at 3:00 PM, the more common challenge was people having to force themselves to go to bed at 3:00 AM, tearing themselves away from the next thing they could accomplish, which was always just sitting there waiting for them.”

What we see empirically is that when people use AI, they work way more rather than way less.

More quantitative evidence for this can be found in proxy metrics for the formation of small businesses. In the post-agent era, the number of domain names being registered and small business applications filed has sharply inflected.

The economic theory behind this effect is perhaps more important than the data itself.

Citrini’s original scenario went something like this: AI can do the same things that humans can do. It starts in coding, which is why software engineers get replaced first. These displaced workers stop spending, which weakens the economy and forces more businesses to cut workers and replace them with AI, further reinforcing the cycle. This continues until the drop in spending destroys the credit markets and causes a financial crisis.

The core premise of this scenario is that there is a fixed amount of work to do. The first software engineers get replaced and stay unemployed as AI has taken a portion of that fixed pie of work. This is why the empirical evidence that AI has created more work to be done than ever before is important because it shows us that the pie is not fixed. AI reduces the cost of providing many services in the economy. For example, coding, writing, and creating infographics. Many goods and services that were previously too expensive and uneconomical to produce (such as starting a software business which used to take millions of dollars of engineering talent and multiple rounds of VC financing, now only requiring a $200/month subscription to Claude) now become viable. These new goods and services present themselves in the economy as business ideas that are only possible in the age of AI. These businesses then hire more people and that then increases employment.

The end result is essentially a Jevons paradox but for labor. By making work cheaper we vastly expand the universe of economically useful work to be performed and cause the need for human labor to go up and not down.

Those first software engineers don’t get replaced because they can now do far more within the organization, or they find a new job because new businesses are created that are hiring, or they start their own businesses. The vicious cycle never begins.

There’s also another argument that humans are partially a complement to agents, rather than agents fully being a human substitute, which is argued by Dan Shipper in his essay “After Automation.”

You can read it here.

I am much less confident in this theory as we have yet to see the empirical evidence for this with much more powerful models. If this is indeed the case, AI can be treated much more like past automation technologies like spreadsheets, factories, or farming equipment which are a complement to labor (and are the examples used for the “the Luddites are wrong again” argument) rather than the theoretical substitute that is more consensus.

Timing and Revisions

The biggest risk to this trade is probably timing.

The AI build out happens before AI creates value. That is why it’s been so much easier to invest in AI infrastructure and make money than it has been to try to predict anything at the application layer or the end economic effects.

Dylan Patel has said in a podcast before that the hardest part of his job isn’t analyzing the semiconductor supply chain (supply-side) but is instead analyzing the demand side.

“I think the hardest area for us and for everyone is understanding tokenomics, economics of tokens. I think we have a really tremendously like good insight into how much it cost to run infrastructure, what the cost of tokens are, what the cost of models are, what the margins of these labs are. But the usage and adoption is what’s really difficult to model, you know, continuously, right?

We, we have these like we had like crazy in January, we had crazy estimates for February, Anthropic smashdown. How do we calibrate this model? What are the data sources for this February? We had crazy assumptions for March and then they smashed them and everyone sees the number of 10 billion and they’re like, what the fuck? How do they add 10 billion of revenue? Who is using all these tokens? Why are they using them? What are they building with them?

And then more importantly, with what they’re building with these tokens, how is that actually diffusing into the economy? And what value is that generating? Because it’s not really something that you can capture in any any GDP statistic, right?

All of the value of the tokens that I use get transformed into better information, which I then sell at a discount to what people used to sell information for relatively because and therefore that information is now making its way throughout the economy and and people are making better investment decisions or better competitive decisions.

But if they’re semi data company or data center company or hyperscaler. And now how, how much what, what is the value of this? And what is that? What is that done to the economy? It’s clearly by every subjective metric, amazing.

But where is the phantom GDP? What is the phantom GDP? How do we track the real economic because because the GDP metrics are not, you know, accurate.

If you were to say what is the GDP that Dylan Patel is making? It’s tiny compared to what the value that I think is being created. And so ultimately, what is the value being created by these tokens, not on a basis of, you know, just simple, you know, what is the knock on effect, right? What is the knock on effect of all the things that these things are doing? And I think that’s the real question and challenge that that’s hard to measure.

I think we’ve got a tremendous, you know, reading on the supply side of things. I think we’ve got a tremendous reading on even a lot of the demand side signals. But it’s it’s what is the value these tokens are generating that’s hard to quantify and measure.”

Thanks, Dylan, for the very long quote that says literally exactly what I wanted to say.

We know how much hyperscaler capex is going to be spent this year, but we can’t predict when that capex is going to turn into token revenues, let alone when those token revenues are going to contribute to productivity in the economy at large, let alone how much productivity it’s going to contribute and if that is big enough to actually move aggregate S&P 500 earnings or aggregate GDP. It’s really hard.

However, even though we can’t draw perfect causation, our job as investors is to take imperfect information and make an interpretation of it that allows us to make asymmetric bets. We can analyze the empirical data from today and come to our own conclusion on if we are too early to see the economic diffusion of AI.

And even though we are still super early in the S-curve adoption of agents, we are already seeing some anomalous data that suggests we are not too early for their macro effects.

First, 2026 EPS revisions are completely outpacing the historical norm. Funny enough the kink in the chart is quite literally right when Claude Code got widely adopted in January and February of 2026.

Second, S&P 500 margins are also increasing much faster than history would suggest, showing that the same revenues can be achieved on a much smaller cost structure, a structurally healthy sign for economy-wide ROIC.

While it is impossible to link these effects to agentic AI and it is easy for it to become another example of the classical correlation and not causation logical error, this is enough evidence for me to conclude that there is likely an effect of AI here and “likely” is good enough for a trade.

The Trade

I called out a trade in my subscriber chat not too long ago (May 8th). You know which one this is lol.

It’s performed quite well since.

It is…

Read more

Please don’t tell us that we’re going to overbuild because we double the capacity of optical widgets.

There are two layers to this exponential.

First, I think networking will grow as a share of the BOM over compute over the long run. You can see this specific value-add transition in NVIDIA’s earnings, as they missed on compute but massively beat in networking.

It’s a pretty simple concept, and I’m unsure why this isn’t talked about more in general. If we are hitting the limits of Moore’s Law and traditional dies can only be so big (sorry Cerebras), the only way that we get more powerful compute units is by linking them together. This means that the networking BOM is about proportional links:chips (inside a rack) or links:rack (inside a DC). And according to graph theory, the number of edges will always grow faster than the number of nodes.

The second layer to the exponential is the CPO copper replacement story itself. Within this layer, there are three more sub-layers. That is, we can split CPO adoption into three phases, each with half an order of magnitude more CPO content than the last.

1. Phase 0 is scale-out CPO. This actually doesn’t add any new optical content for the total market, but it does add CPO content for the CPO vendors. In Phase 0, transceivers that connect switches in the scale-out network are replaced with CPO. Though it’s a little bit confusing because the ramp happens first but happens the slowest. This does not mean that transceivers will all of a sudden disappear.

2. Phase one is inter-rack scale-up CPO. This is the art of taking multiple racks, which are usually their own scale-up domain (one coherent piece of compute to the software), and linking them together with optics to make an even larger scale-up domain. This is represented by NVIDIA’s NVL576 Oberon solution for their Vera Rubin platform.

3. Phase two is intra-rack scale-up CPO. In this final phase, we are ripping out all the copper SerDes and interconnects of a single rack and replacing it all with CPO. As you may imagine, this has by far the most CPO content out of any of the phases, but is also very far away.

Lumentum has this slide at OFC which describes it. They are very excited, as you may imagine.

The timing is also important. CPO is not going to happen as soon as the most optimistic bulls think.

Volume shipments of phase one scale-up CPO will not occur until the Feynman generation. This is despite Rubin Ultra having the NVL576 Oberon 8-rack system, as this will ship in far less volume than regular Kyber NVL144 and Oberon NVL72.

But I’ve, of course, been a very vocal CPO bull. So where do I stand on this?

I think what is underrated about CPO is, instead, its inevitability and sheer VOLUME. People simply do not understand how much CPO will be needed once it actually does ramp.

Anyways, let’s talk about all of the widgets needed for CPO and the different widget makers in the CPO supply chain.

Contents

1. InP Substrate

2. Epitaxy

3. Laser

4. Optical Engine

5. Conclusion

6. AXTI

7. Aixtron

8. Lumentum

9. Coherent

10. Soitec

11. Tower Semi

12. My Ranking (and Positions)

SemiAnalysis Giveaway: The first person to subscribe using the paywall on this post will be selected to receive one free month of SemiAnalysis (sent to your email).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

InP Substrate

InP is the pure laser raw material. Indium + Phosphide is a III-V (named after periodic table columns) compound which has a direct band gap, meaning that it can emit light, unlike silicon.

This chemical is very hard to make. You must refine high-purity indium (99.99999%) and phosphorus and then grow this compound into an ingot. It’s really hard to grow this stuff. This isn’t memory, where it’s a commodity. You gotta get qualified and develop proprietary processes (like AXT’s VGF).

InP substrates are the bottleneck to end all bottlenecks.

Jason's Chips

What is an AXTI (and why is it up 80x in 12 months)

AXTI is a funny company…

Read more

a month ago · 30 likes · 3 comments · Jason's Chips

AXTI, the poster child, actually doesn’t have majority market share. They are only the second-place player with roughly 36% share, while Sumitomo Electric holds 42% share. However, I believe AXTI will soon become the largest player here, as they have the most expandable capacity due to their ability to convert gallium arsenide capacity into indium phosphide via brownfield switcheroo.

The relevance for CPO is simple: CPO lasers are high-power CW lasers that require a much larger cavity in order to be so powerful. This means the die area is much larger, and each laser a b s o r b s much more InP substrate. And then you get these lasers to grow exponentially across three phases, each with half an order of magnitude of content increase. Today, despite having no CPO ramping at all, indium phosphide is so constrained that Lumentum is 30% behind demand in just traditional EMLs. See the problem here?

The question for this market is how big the price hikes are going to be. AXT trades at something like 30x exit 2027 ARR, which is ridiculous by any traditional measures. The management team has obviously not announced any crazy memory price hikes, but we all know that eventually they are going to raise the prices. The market is obviously discounting for some level of price hiking, but we just don’t know how much it’s going to be. Bulls will say the ASPs are going to 10x. Bears say it’s only going to 2x.

Epitaxy

Epitaxy is the next layer. In order for indium phosphide substrates to be turned into real lasers, you have to perform a process of crystal growth called epitaxy to turn them into epiwafers first. It’s an intermediate stage.

To do this, you use Aixtron’s MOCVD (metal organic chemical vapor deposition) tools. Aixtron has a proprietary process called the planetary reactor, which allows you to batch this activity. Because of this, Aixtron has a literal monopoly on this whole market. They claim “well north of 90% market share,” which is just a kind way of saying this. If you want Veeco tools, you have to throttle your throughput because you won’t have planetary reactor. Aixtron is a monopoly. You have no other choice.

Then there are the buyers of these tools and those who actually do the epitaxy-ing.

Lumentum and Coherent have this process vertically integrated, but for anyone who doesn’t, you can buy from IQE.

They are a pure-play epiwafer maker.

Laser

Once you grow the epiwafer, it goes through a more traditional compound semiconductor fab process with lithography and etch to get turned into the final laser. This is purely the domain of laser manufacturers like Lumentum and Coherent.

Since we’re only looking at the indium phosphide supply chain today, which is the vast, vast majority of CPO anyways, we are only going to consider DFB lasers and not VCSELs.

CPO lasers are high-power continuous wave lasers. This means that they have two important characteristics:

1. Being continuous wave (rather than EML)

2. Being high-power (rather than low-power)

CW lasers are lasers that just stay on. It produces a continuous stream of light, like a light bulb or flashlight. A CW laser can provide a steady beam that gets modulated (gets the signal put into it) somewhere else.

EMLs are lasers plus a built-in light switch. EML stands for electro-absorption modulated laser. The laser creates the light, and an attached modulator rapidly changes that light on/off or high/low to encode data.

Lasers for CPO must also be placed far away from the optical engine because CPO literally stands for co-packaged optics, which means the optics are co-packaged with the chip and the chip is really hot, so if the laser is right next to the chip, it overheats and breaks and stuff.

Because this laser is so far away, in order for the light to reach the optical engine, it needs to survive a lot of coupling loss. Thus, the starting power of the light must be very high. For CPO, this means generally in the range of 400 mW rather than 30 to 70 mW of traditional CW lasers for SiPho transceivers.

The higher the laser power, the harder it is to make. These lasers must meet critical thresholds on two noise specs: linewidth and RIN. This is essentially the purity of the light. I won’t go into much more detail on noise specs, but if you’re interested, I talk more about them in this article.

High-power CW lasers are currently a market almost completely dominated by Lumentum.

Optical Engine

CPO lasers are CW, which means that in order to actually use it to communicate signal and not just be a flashlight, you need external silicon photonics components to do something with the light.

The optical engine is what you can see in the above diagram of the CPO. It is made up of a photonic integrated circuit, or a PIC, and an electrical integrated circuit, or an EIC, hybrid-bonded together (using something like TSMC’s COUPE). The EIC is your standard logic chip. The PIC is a SiPho chip. It is the SiPho PIC which actually modifies and modulates the light, but both of them come in a package whenever we are thinking of CPO. Here is an article if you want to go deeper into SiPho.

Jason's Chips

Short n' Casual Intro to SiPho

SiPho is pretty cool ngl…

Read more

3 months ago · 75 likes · 2 comments · Jason's Chips

There are three layers to making an optical engine:

1. Making the SOI wafer

2. Fabbing the PIC in a foundry

3. Hybrid bonding the PIC to an EIC

SOI stands for silicon on insulator. It’s basically a silicon sandwich and needed to contain the light waves as they travel through the waveguides.

Soitec is a monopoly here. They control the majority of the IP needed to make these wafers, which they call Smart Cut. Just like with indium phosphide, this is a very difficult process.

I cover this topic a lot more deeply in this article.

Jason's Chips

The Soitec Series | Part 2: Wafer Manufacturing and the Smart Cut Moat

I wish this was a real product…

Read more

3 months ago · 55 likes · 3 comments · Jason's Chips

Tower Semi is the leading silicon photonics foundry.

Soitec provides the starting wafers that go through Tower Semi’s specialty foundry manufacturing process to become a fully fledged PIC.

They are the best at it. Globalfoundries and TSMC also play in this market, but Tower CEO has specifically flagged and quoted that they have by far the highest market share.

When Krish Sankar of TD Cowen asked Russell about market share, Russell said:

“There’s other people that claim very high market shares. I don’t see how it’s possible at all.”

LOL

Finally, the PIC and the EIC need to be hybrid-bonded to form an optical engine, and the leading platform here is TSMC’s COUPE. TSMC did something really smart here, which is that they require that the PIC is fabricated at TSMC in order to use their COUPE packaging. Now their customers have to choose between the superior PIC manufacturing of Tower Semi and the superior EIC and packaging capabilities of TSMC.

Currently, it is COUPE that has won NVIDIA CPO ramp.

Conclusion

The big takeaway you should have here is that every step in the CPO supply chain (especially the raw materials) is physical and linear. Nothing here is built to handle exponential demand. Everyone has capacity limitations.

However, refining indium phosphide substrates, building MOCVD machines, growing epiwafers, manufacturing lasers, and fabbing PICs all have different capacity dynamics: different speeds of adding shipment volume, different capital intensity, and different substitutability. It is by understanding these dynamics through a lens of perpetual supply constraints and exponential demand, you can identify which companies stand to benefit.

Below I will give my opinions on AXTI, Aixtron, Lumentum, Coherent, Soitec, and Tower Semi, give you revenue builds/EPS estimates for all of them (making the same assumptions about the total industry-wide unit shipments), and share my personal ranking based on my opinion and the positions I have.

Yes, you heard that right. That’s six models!

I also call out why AXTI could still have the most upside, how I modeled the bombshell Aixtron dropped on their last earnings call, why Coherent management speak makes me believe they are uncompetitive, and a very interesting (!!!) component ratio dynamic between Soitec/optical engines and Lumentum/ELS.

Read more

But now maybe it is time to trade on fundamentals, which will be thoroughly analyzed via this piece of writing.

Today I will be taking you on an all-inclusive journey. We start from the technical foundations of their wafer-scale engine before evaluating the business and market side of their products. Everything that you’ve ever wanted to know about Cerebras, all in one place.

Contents

1. Introduction to Dinner Plate Computing

2. The Tale of Two Cities & The Nightmare Compiler

3. Wafer-Scale Yield

4. PVT Calibration

5. Fast Inference (Tensor vs Pipeline Parallelism)

6. Memory Constraints

7. The I/O Bandwidth Problem

8. Economic Relevance

9. Financials

10. How to Think About Cerebras

11. Testing 5.3 Codex Spark

12. Bull Case

13. Conclusion

As a Substacker, I am obligated to say that this is the kind of research that comes with a paid subscription!!

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Introduction to Dinner Plate Computing

Cerebras is the well-known pioneer of dinner plate computing.

Their foundational premise is that, BIGGER IS BETTER.

Normally, when you order a wafer from TSMC, you slice up your wafer into dies. Cerebras decided to keep all of the dies intact and instead network them together into one really big chip.

These spaces between the dies are called keep-out zones because normally you keep everything out of them because they will be cut up by the diamond saws. In this case, these keep-out zones are the home for complex wiring and networking. This means that Cerebras needed to collaborate intimately with TSMC to develop proprietary IP in order to fill up these keep-out zones with networking.

Throughout their investor materials, they make a plethora of comparisons to traditional computing with a bunch of metrics of stuff that chips do (compute cores, transistors, memory bandwidth, FLOPs), and because their chip is bigger, they can do more chip things.

This can mislead retail investors, as in real life there is always a trade off. Cerebras is simply a computing architecture optimized very differently than traditional GPUs.

The Tale of Two Cities & The Nightmare Compiler

SRAM is the super-fast expensive memory at the very top of the memory hierarchy. Faster than DRAM and even HBM.

This memory is usually made with a logic process and sits very close to the compute.

In a traditional GPU, all of its cores share the same unified pool of SRAM.

However, at a wafer scale, a unified memory pool is physically impossible. Therefore, Cerebras designed each compute core to have its own distributed 48 KB of SRAM. This is very little memory (the entire WSE would only have 44 GB of memory total compared to 13.5 TB of HBM3E for the GB200 NVL72 rack).

I like to think of this as a tale of two cities.

One is a small village with a few hundred residents. In this village, there is a shared library in the middle where residents go to fetch the information that they need. It is very easy for this village to coordinate as any two residents can read the same books.

Another is a nightmarish-looking homogeneous suburbia with 900,000 identical tiny homes. Each house has its own small bookshelf. You can only read the books that are in your own house. As you might imagine, it becomes a nightmare for this civilization to coordinate and communicate any information.

A compiler is the machine that translates human (or agent ha) intent (code) into silicon action. It needs to tell the transistors what to do.

In the village with the shared library the compiler’s job is easy. In the homogeneous suburbia, it’s a nightmare as you would have to write a master schedule that guarantees person number 452,000 has the exact right piece of data placed on their tiny bookshelf at the exact right millisecond and then passes it on to person number 452,001 right as they finish using that piece of data.

Because this hardware fundamentally guarantees the compiler complexity, Cerebras can never achieve the software optimization of NVIDIA. However this does serve as a pretty intense barrier to entry for new wafer-scale competitors.

Wafer-Scale Yield

Manufacturing chips is very hard. Sometimes a speck of dust can ruin your chips. Wafer yield is simply the percentage of good dies divided by the total number of dies on the wafer.

No one can ever get close to 100% yield; it’s always somewhere between 50% to a hundred percent.

But wait a minute! Wait a minute. Cerebras is making its chip an entire wafer. If you can never get a hundred percent yield, doesn’t that mean that Cerebras can never get a functional wafer-scale engine? Every single wafer is guaranteed to have at least one defect!

Not at all. Cerebras did a very clever thing where they built in an extra redundant communication pathways between the cores. If they find that a portion of the wafer, a certain die, has a defect, the compiler simply routes around it. This is how Cerebras claims a 100% yield rate on their Wafer-Scale Engines.

PVT Calibration

PVT stands for process, voltage, and temperature variation.

Semiconductor fabrication is a highly chemical process. All the etching and depositing and stuff are all done through chemical reactions, even though in diagrams it may FEEL physical.

Because these processes are all chemical, they are never fully uniform across the entire dinner plate. Transistors near the edge may end up switching faster or running slightly hotter than the same transistors located closer to the middle.

Normally, this wouldn’t matter for a traditional GPU because the dies are cut up anyways. But for Cerebras, it becomes a problem because the entire Wafer-Scale Engine needs to be fully synchronized.

This is the exact problem that caused the industry to abandon wafer-scale computing decades ago. If you’re required to clock the speed of the entire Wafer-Scale Engine chip to the speed of the slowest transistors, the performance would be atrocious.

Cerebras, however, developed a system to dynamically recalibrate the entire wafer on the fly. They detect the temperature and voltage and frequency characteristics of every part of the wafer and then actively manage the power delivery accordingly. Basically, every single core gets exactly what they need.

This is by far one of their most important innovations. PVT calibration is a real moat for Cerebras and represents an enormous barrier to entry for Wafer-Scale competitors.

Fast Inference (Tensor vs Pipeline Parallelism)

This is our main course today. Tensor vs Pipeline Parallelism is the single trick up their sleeve which allows them to perform their main value proposition to the market: fast inference.

Tensor Parallelism

Let’s think about how a standard GPU runs inference, called tensor parallelism. Usually, the model weights are too much to fit inside of the GPU’s SRAM. It has to be chopped up and processed by multiple GPUs in parallel.

Once the separate GPUs are done, they have to share their answers via an all-reduce, where every GPU has to talk to all others. This generally happens within the same NVLink scale-up domain so we’re talking about microseconds of latency, which doesn’t sound that bad, but there are two things that actually make it pretty problematic for fast inference.

First is that this all reduce must happen for every single token the model generates as decode is sequential.

Secondly, not only does it have to happen for every token, it has to happen for every layer of the model. Transformers are made up of many attention and multi-layer perceptron layers.

So for a 100-layer model, every token generated requires 100 all-reduce operations across the tensor parallelism group. If you have a lot of tokens, that’s a lot of latency!

There is also a separate latency bottleneck caused by HBM. Each time you generate a token, you have to load the model weights from the HBM to the compute. This is completely separate from the all-reduce and the tensor parallelism approach we just discussed, but is another reason why Cerebras is able to absolutely speedmog Blackwell.

All in all, several microseconds of latency per token results in the general hundreds of tokens per second achieved by regular normal GPUs.

Pipeline Parallelism

Pipeline parallelism is like a pipeline.

Like I said the model is made up of a bunch of layers. Each time you do the math of one layer, you get an activation output, which you pass downstream of the pipeline to the next layer.

Now here’s where Cerebras comes in. Cerebras has a big ahh chip. Whereas GPUs cannot fit entire layers of a model on their SRAM, Cerebras can. Therefore, when Cerebras does matrix math, it does not need to run an all-reduce operation to combine the answers for different slices of the model layers. It also doesn’t need the HBM. All the communication happens on-chip, and on-chip communication is very speedy.

Notice I say model layers and not model. This is important. Cerebras doesn’t put the entire model on one chip. It’s still not big enough to do that. It simply strings a whole bunch of dinner plates together, with each one processing many layers, getting the activation outputs and passing that downstream to the next set of layers. Cerebras has major IO bandwidth bottlenecks (which we’ll talk about later), but that is not an issue here because the activation outputs are very small.

The result? Data literally flows down the wafers like water through a river. Smooth, buttery, and unbothered. 15 times faster inference by cutting out the all-reduce and HBM communication latency.

Memory Constraints

Remember when I told you they have 48 KB of SRAM per compute core, which is 44 GB per dinner plate? Well, turns out this is not nearly enough to do anything useful in the world of frontier models.

To serve a model, you must fit both the weights and the KV cache (context) into the SRAM. 44 GB isn’t even enough for most models’ weights; and the KV cache is way larger too. To fit a classic open source model, like one of the smaller, dumber Llama ones or the more important DeepSeek V4, you will need a single-digit number of dinner plates to hold the weights and a double-digit number of dinner plates to hold the KV cache.

Because each dinner plate has so little memory but yet is so expensive, this makes serving models using Cerebras very, very capital inefficient.

The I/O Bandwidth Problem

Now you might be asking, why don’t we just offload the KV cache to HBM outside of the wafer?

Funny enough, the picture of the chip itself should hand you an obvious clue.

BIG AREA SMALL PERIMETER

The off-chip bandwidth is 1.2 terabits per second, compared to 21 petabits per second for internal bandwidth. That means that data within the chip travels more than ten thousand times faster than data going off the chip.

Beyond that, it also does not compare favorably to GPUs. For the B200 as an example, it has 14.4 terabits per second, which is over 10 times that of the WSE-3. This is like having a massive Olympic-sized swimming pool, but the only physical way to fill it or drain it is by sucking the water through a single, thin plastic drinking straw.

This is also a major reason why their hardware failed in training. Training requires data to be constantly passed from on-WSE to external memory which it absolutely does not have the I/O bandwidth for. This is intuitive, as you can imagine the WSE’s job in training would be to mimic a massive, optically linked GPU cluster while in inference, it only needs to imitate a few servers or a rack.

But the important conclusion here is that there is no escape: both the weights and the KV cache must be held by the on-chip SRAM at all times.

Economic Relevance

What does this all mean?

Well, think about what we’ve covered.

1. We’ve established that the WSE is a fundamentally different architecture than the GPU. It requires an incredibly complex compiler, redundant communication pathways, and PVT calibration. It is very, very different.

2. We explored what makes this very different chip better via pipeline parallelism, a feature that allows Cerebras to run inference over 15 times faster than traditional GPUs by avoiding the all-reduce and HBM latency.

3. We looked at what makes this very different chip worse, namely the suffocating memory constraints and tiny I/O bandwidth.

Notice something here: the advantages are about speed and quality of the experience of the user being served by the chip, while the drawbacks are all measured in number of bits and are about the sheer cost and economic unattractiveness of such quality service.

This leads us to our conclusion. What makes this chip better is that it generates tokens faster, but what makes it worse is that it is far more expensive to generate each token. Cerebras and wafer-scale computing in general is a provider of fast, premium-priced tokens.

The business side of this company is all about exploring whether such speed is worth the extra cost. Are people actually willing to pay for such expensive but fast tokens? Will this type of token take share from the traditional NVIDIA-generated kind? Are there new use cases that require this type of token that we haven’t been able to unlock yet because the hardware didn’t exist, or is the current regime already good enough for all of our uses of AI?

Let’s find out together.

Financials

Cerebras IPO was 20 times oversubscribed at the initial $130 share price. People really want to buy this thing. Yesterday, they raised the IPO price to $160. Polymarket still has almost a 100% chance that they will close above a $50 billion market cap (that is for their basic equity value, which is only around 70% of their fully diluted, so it is way higher than what’s shown in the table).

Their 2025 revenue is miniscule and irrelevant at $510m. Currently, they have $24.6b RPO and $20b of contract value from OpenAI, which can be expanded over time, but this implies a $6.7 billion steady-state ARR.

At the assumed IPO price of $160 per share multiplied by the fully diluted share account, the implied revenue multiple to OpenAI’s ARR is 7.4x.

The First Noble Truth of Cerebras

The following is the first Noble Truth of Cerebras.

The financials are useless. Completely utterly useless. Forward estimates are useless. Growth rates are useless. Multiples are useless.

Now, why is that? Let’s think about this from first principles.

How To Think About Cerebras

Cerebras is a canonical example of an unproven technology early in its S-curve with potential for high double-digit to triple-digit growth. This high-level conceptual profile allows us to anchor to two companies that have plenty of coverage on this channel: Lumentum and Bloom Energy.

Lumentum

The 30,000ft intellectual framing of Lumentum is simple: one market (optics) is guaranteed (by the laws of physics) to replace another (copper) to achieve full dominance of a certain use case (AI networking).

Jason's Chips

The Lumentum Series | Part 3: The Genesis of Scale-Up CPO

So, I’ll introduce you all to a fun little method I use to study up on stocks…

Read more

3 months ago · 56 likes · 4 comments · Jason's Chips

Because the transition is inevitable, the leading players in the new market have a massive guaranteed growth runway. This is why optics are so hot.

At the same time, the growth is measured and measurable. It is easy to model. We can confidently predict that Lumentum has 50% year-over-year growth because we can simply use the growth rate of optics as an anchor.

Bloom Energy

Bloom is slightly different. One company provides an objectively superior solution for one single market, which has the potential to fully displace the incumbent.

Jason's Chips

Bloom Energy Thesis (Free)

As a Substacker, I am obligated to say that I am releasing this for free to provide an example of what kind of research you would receive with a paid subscription…

Read more

2 months ago · 79 likes · 19 comments · Jason's Chips

This is different from Lumentum because the market doesn’t change. Bloom plays in behind-the-meter energy, just like GE Vernova. They are direct competitors. There is no copper equivalent. There is no transition from one market to another.

Because of that, Bloom’s growth can be a little bit more unpredictable and bursty. There is no big Kager number to go off of. It is just a matter of how quickly customers adopt the new solution and how much, or if, the new solution is superior to the old one. However, growth here is still predictable. If the new solution is better, it will still slowly displace the incumbent.

Cerebras

Cerebras is completely different.

From our first article, we ended with the ultimate conclusion that Cerebras is a provider of a fundamentally different product than Nvidia: premium tokens. Premium tokens are way faster but cost way more.

Is Cerebras like Lumentum? No, there is no world in which premium tokens will inevitably displace normal tokens. They are simply two different products each with a different trade-off.

Is Cerebras like Bloom Energy? No. WSE is not a superior way to generate the same end product (electricity for Bloom, regular tokens for NVIDIA).

In both comparisons, Cerebras is actually much worse. It’s not guaranteed to take share from anyone, really. However, there is one critical distinction, and it’s very critical. Very, very critical.

The inference market is the single biggest market in the world.

AI networking and behind-the-meter power are big markets, but they’re not the biggest. Inference is, and the current incumbent is the biggest company in the world. If these premium tokens carve out a large enough slice of this overall pie, Cerebras is a very attractive investment. The only question is, will it carve up a larger left slice? Again, unlike Lumentum, it is not guaranteed to take any share at all.

Now you probably understand why I say that the financials are useless. Single-digit billions of ARR do not matter at all. If you are unprofitable and your market is the biggest market in the world, what matters much more, and really what matters at all, is the slice of the fast inference pie.

No amount of revenue will save Cerebras from offering a fundamentally unwanted solution. No valuation is too high if they serve a sizable fraction of all inference.

Therefore, the point of this article is to simply assess how likely it is the fast inference market ends up being incredibly large.

5.3 Codex Spark

5.3 Codex Spark is the custom-developed model by OpenAI, specifically made for running fast inference on Cerebras hardware. It is currently the only custom design model for Cerebras hardware, so in reality this model is one of, if not THE most important element of any Cerebras investment analysis. But for some reason, it is very under discussed.

It trades a large chunk of its performance to be smol. It features substantially lower benchmark performance, a context window that is one-eighth of the leading frontier models today, but with inference speeds of over 1,000 tokens per second.

As we established in the last post, in order to run a model of a certain size, a certain number of dinner plates must be strung together. The larger the model weights, the more wafer-scale engines are needed, and the more uneconomical it becomes. Therefore, small models are much more cost-efficient for Cerebras to run.

Now, how good is this model exactly? Here is where I found something funny.

The Cerebras IPO is one of the most hyped events of all time, but the only custom-designed model for its hardware, 5.3 Codex Spark, is one of the most un-hyped and un-loved models out there.

X Sentiment

AI lives on X, so let’s start with the X sentiment. Funny enough, I couldn’t find big accounts that posted about 5.3 codex spark just because of how underdiscussed it was, but there were enough posts for me to come to the conclusion that people generally did not like this model at all.

The main complaint is just the performance. It is a stupid model, to put it simply. Frontier intelligence unsurprisingly requires a certain amount of brute force that small models cannot muster.

My Own Testing: Equity Research

I have a little bit of a bias against benchmarks and testing that people usually do for these models. I don’t care about your score on Humanity’s last exam or whether you can build a snake game in one shot. These are not tasks that I will be doing in real life. This is a point people say a lot, but I think it is underappreciated because they are categorically different from benchmark tasks.

So, to that end, we will be doing two tests that I literally came up with by just asking myself what the next thing I have to do is. Which are equity research and coding.

In a very meta turn of events, I asked both 5.5 and 5.3 Codex Spark to look for pricing and usage data for 5.3 Codex Spark. So using Cerebras to evaluate Cerebras.

This is mainly a speed test. It is very difficult to gauge quality on research tasks because it is very subjective and could require large sample sizes all evaluated qualitatively before you can get a real sense of the difference. That is why we’re doing coding too.

But for speed, you might notice something interesting. Although Cerebras is supposed to be over ten times faster, 5.3 codex spark took 2 minutes and 19 seconds, while 5.5 took 3 minutes and 9 seconds. This means that 5.5 only took 36% more time rather than 900% more time.

This is because of two reasons. Both of which I observed first hand in my testing and is something us semiconductor people talk about all the time.

1. A large fraction, or maybe even a majority, of the time spent on doing the task was on web searches, tool calls, and JavaScript executions. Sound familiar? That’s right! CPUs strike again. All those research reports on how a large fraction of the latency in inference is CPU-bound. Well, turns out it isn’t just theory. 5.3 Codex does this really funny thing where it generates tokens in a millisecond, and then you have to wait two or three seconds for the CPU to run the tool call. You all should try it. It feels really weird, but you will immediately understand why CPUs are needed and why having fast inference isn’t all that it’s cracked up to be if you’re only speed boosting half of the workload.

2. Because the context window is so small, compaction happens very often. Every time the model has to compact its context, more latency is introduced. Even in that short two-minute task, GPT-5.3 Codex Spark had to compact its context, which is why it didn’t finish that far ahead of 5.5, which has a context window of a million tokens. Having a larger context window actually reduces latency by allowing the model to run without compaction for longer.

My Own Testing: Coding

I’m currently developing a full stack terminal for semiconductor investors, which I hope to release by the end of this month. It has a bunch of cool features like:

• A Pokédex of every single AI infra company, split up by industry and niche

• A map of all HPC co-location sites from the Bitcoin miners

• A portfolio lab for backtesting and paper trading

• An earnings calendar where you can view upcoming earnings, add events to your calendar, and download transcripts for past events

• A Taiwan revenue tracker

• A memory price tracker

• An API endpoint for AI agents to access the app’s features autonomously

• 10 different calculators

Therefore, I assigned both models to build my next feature, which is a paper trading feature for the portfolio lab.

This was not a good test for speed because 5.5 ended up pulling up the in-app browser to test as if it were a real user, while 5.3 Spark just shipped it immediately. They made completely different workflow decisions, which means that it cannot be compared apple to apples, unlike for the equity research task. Even then, GPT 5.5 took 16 minutes, while 5.3 codex Spark took 3 minutes, which is only about a 5x difference. That’s still not 10x, even with one model doing way more!

How did they perform? Let’s look at 5.5 first. GPT 5.5 successfully implemented the paper trading feature, and I was able to create my new paper portfolio called Bloomentum.

5.3 Codex Spark, on the other hand, had this result.

To see the actual finished product that 5.3 Codex Spark can create, I pasted the error back in and allowed it to fix the bug.

The quality of the work done by 5.3 Codex Spark is far worse. First of all there is another bug: the pricing quotes do not load at all. But if you just look at the UI of this tool, you can already notice the difference. It is very unintuitive and unfriendly to use. It is black and white instead of matching the colors of the app. The buy button isn’t even legible!

Additionally there’s this weird jittery effect where whenever you click a button or type something in, the elements sort of flicker.

Conclusion

Is 5.3 Codex Spark faster? Yes, without a doubt. However, is it 10 times faster like Cerebras claims? Not at all. Real-world workloads are not just pure feed forward. There are tool calls and there are context limitations. Just like doing well on the SAT doesn’t guarantee that you will be successful in a real-life career, running fast inference on a benchmark test in a lab does not guarantee that that same speedup applies in real life.

The performance difference, however, is extremely significant. There is a reason that people gravitate towards frontier-level intelligence even when cheaper open source alternatives exist. OpenAI and Anthropic’s combined ARR is not $80 billion for no reason. More intelligent models do things better. Not only do they do things better, they usually find ways to do them more efficiently, using less tokens in the process, and do it quicker as well. Not because of the sheer inference speed, but because of their superior reasoning abilities. The time it takes to correct bugs and failures, like the one seen in our test, will most likely overwhelm any gains from faster inference as the magnitude of the speed up in real life work just isn’t big enough.

And finally, one overlooked signal is the existence of 5.3 codex spark itself. OpenAI does not have an ultra-fast Cerebras option for GPT 5.5. They purposefully created a smaller weaker model optimized to run on this hardware. That’s actually pretty telling because it was a deliberate choice. They definitely thought about serving the frontier model on Cerebras first and then through testing and deliberation decided against it as it was either impossible or uneconomical.

Bull Case

Let’s be honest, these 5.3 codex spark tests were pretty bearish, so what’s the actual bull case? Well, I think I have a few.

The Undiscovered TAM

The undiscovered TAM argument simply posits that nobody really knows what fast inference could be used for, so the addressable market is completely undiscovered. Because nobody knows about these use cases, there is no market recognition of it, and because there’s no market recognition of it, it must be underpriced.

Here are three potential use cases that I can think of.

Fast Fundamental Investing

I’ve tried once or twice to trade earnings calls and catalysts and have failed every time. I come in with sheer, utter confidence in my fundamental understanding of the company and then get immediately humbled as my thinking speed is way too slow to process the information that was just presented to me and be able to connect that to what happens to the fundamentals of the company.

As you might imagine, this could be something that Cerebras is very good at. Imagine you’re a hedge fund and you have a lot of contacts in a company and a model already built. All you have to do is just give all of that to an LLM and tell it to watch out for catalysts. Once the event hits or the earnings call starts, the model is able to immediately run that inference and tweak the model and come to a conclusion. Having this fundamental-based conclusion just seconds or minutes before everyone else means that you can catch a 10 or 20% move while the rest of the market is still racking their brain around what just happened.

Embodied/Humanlike AI

What if I want an AI in my Zoom call? What if I want to talk to a model and have it interrupt me like my friends do? There is a level of speed required in humanlike back-and-forth conversation to enable idea exchange at a cadence that humans are familiar with.

Think about the way you talk to an AI right now. It is essentially long monologue after long monologue. This is due to the inference speed limitation. You may be able to imagine the plethora of use cases that can be unlocked with this constraint gone.

Robotics is another interesting application here. Imagine if humanoid robots can talk exactly like humans do. This would be an insanely large addressable market that includes all service industries like dining, hospitality, and front desk work.

Real-Time Human Augmentation

Finally, you have real-time human augmentation, which is essentially giving humans an earpiece that can allow an LLM to help that person in real time. Think about high-stakes meetings, live negotiations, that sort of thing.

The Low Hanging Fruit

The second category is low-hanging fruit. Low-hanging fruit being architectural innovations that Cerebras can easily make, which would be a step function change in their unit economics. Because their architecture is so much less mature than traditional GPUs, it is very unoptimized and therefore can be optimized.

FP8 & FP4 Support

The first thing that they can do is to enable FP8 and FP4 support. It is quite puzzling to me why they don’t do this already. They only currently support FP16, which is like having very limited storage on your phone, but the software only permits you to store photos and videos in 4K HD.

I think it’s likely because if they enabled this support, they would have to redesign their cores. Since their original design was meant for the training market, it was never optimized in this way. Anyways, there are rumors that this should be coming to the WSE-4, and that would unlock an easy 2x improvement in memory capacity.

Hybrid Bonding

The second thing they can do is more speculative. What if they hybrid bond an entire SRAM wafer on top of their WSE? It would be quite difficult, as it involves solving a ton of engineering problems and foundry shenanigans.

If they can do this, they scale their memory by an order of magnitude. Honestly, it would make Cerebras a completely different story.

But again, it is very speculative, and there’s no indication that they’re pursuing this.

The Non-Nvidia Ecosystem

Because Nvidia acquired Grok, they are the only major accelerated computing provider to have a fast inference solution. AMD and the hyperscalers are left without one. As you can see this argument is quite simple: the only way for them to compete effectively with Nvidia’s Grok LPU racks is to partner with Cerebras. All of Cerebras’s competitors are far behind and do not have comparable commercial traction, whether it’s MatX, SambaNova, or D-Matrix. Cerebras would be the only choice.

Conclusion

Because there are fixed, unchanging real-world constraints on total latency of agentic inference, there comes a phenomenon of diminishing returns on speeding up the feed forward itself, as there is no amount of pure transformer feed forward shenanigans that you can do that will make your CPUs do web searches, run JavaScript, and call tools faster. I believe Cerebras is well past that point. In my experience, GPT 5.5 low fast latency is already far more dependent on CPUs than the GPU.

The only model that is custom made for Cerebras is a subpar performer and has very little commercial traction. This makes me believe that the technical trade-off that they made and their “premium token” value proposition that we discussed in our last article does not currently have a market fit.

Therefore, I stayed out of the IPO.

However, I will still be watching this company very closely because of the bull cases. None of them are currently unfolding today, but if they do, this entire story can flip on its head. The thing is, I am not going to speculate on them. I want to see them happen before I get in. If I miss the first 100% to guarantee that the thesis actually exists, that’s okay. Just look at the Bloom Energy or Lumentum charts. Anybody who missed the first 100% probably aren’t complaining about it.

This is despite an increasing magnitude of their beats over time, and probably is frustrating Jensen right now.

We will see if today breaks the pattern!

Here’s today’s Buy Side Consensus and Street Consensus courtesy of Jukan on X.

Same pattern as always: they guide for $78 billion, which means that they will probably beat by $2-3 billion due to historical conservatism, Sell Side models just above their guide because of Sell Side conservatism, and the stock probably dumps and takes down the whole AI infra ecosystem along with it unless they meaningfully exceed even buy-side consensus.

In terms of their rev guide, buy-side expects $90 billion. That’s a lot.

Gross margin expectation is 75% as always, and I bet it’s going to come out at 75% as always.

The Print

My initial reaction was WOW. $1 billion above buy-side on the guide and almost $1 billion above buy-side on the current quarter revenue. But then I realized it was only a $1 billion beat. Which means… nothing really happened.

The price action is really funny. We went down 3%, then flat, then up 2%, and now we’re flat again.

Two other pieces of information that were reported in the print (and we don’t have to wait until the call for closing parentheses) are:

• Nvidia said that they have not generated any revenue under their H200 program and do not know if any imports will be allowed into China (by China) at all.

• They said that Rubin is not delayed to counter all the Twitter rumors lol.

IMO, the real signal can only come from the call today. There’s a reason that nothing really interesting happens with the numbers ever, because Nvidia’s numbers are mostly a lagging indicator of AI hardware procurement. They’re also easy to estimate because Nvidia is so big. The market would rather look forward.

There have been tons of rumors on Rubin being delayed by a quarter or two. There are a few speculations on why this is.

1. Competition especially, AMD has pressured Nvidia to start upping the specs of Rubin. For example:

   • Power from 1800 W to 2300 W

   • Memory bandwidth from 13 Tb/s to 22 Tb/s

   • Increasing the HBM4 requirements to over 11 Gbps per pin

2. The increased HBM4 requirements meant that the Big Three, the Big Memory Three, had to redesign their samples and resubmit.

3. There was also a claim that Nvidia had unresolved SerDes chip interface problems and was “dialing down specs” as a response.

4. And also heat-spreader redesign?

Anyways, Nvidia is probably going to get a lot of pressure from the sell-side analysts to explain what’s going on here, so we should learn a lot today.

I have never seen a post-earnings chart as perfectly flat as this one today. It’s so perfect and so flat, it’s funny. Probably won’t be the case after the call, though!

The Call

Contents

1. New Reporting

2. Compute vs. Networking

3. Neoclouds

4. Growth Rate vs Hyperscaler Capex

5. Frontier Labs share

6. CPUs

7. Other

   1. Supply Commitments

   2. Capital Returns

   3. Groq LPX

   4. VR Ramp

Read more

His 13F showed a very scary amount of puts on literally every single large-cap semi company.

Yes, I am aware I played a part in the FUD. Citrus correctly called it out.

The lesson today is simple. OPTIONS ON 13-Fs ARE MISLEADING!!

How Options Are Reported on 13Fs

The market value row shown above is simple for shares, but not so simple for options.

For shares, it literally is just the market value at the end of the quarter.

For options, however, it is not the value of the options. Instead, it is the market value of the shares controlled by the options.

This means that the actual market value of the options is way, way smaller than the market value shown on 13F. Plus, it is completely unknowable because we do not know the expiry date and moneyness (if it is in or out of the money and by how much) of the option. These are important because the closer the expiry, the lower the market value of the option relative to the shares it controls. The more out of the money the option is, the lower the market value of the option relative to the shares it controls.

This visual really hits home. There are two main takeaways that we need to have here:

1. The option value is guaranteed to be very low compared to the reported value on the 13F. In many cases, it is near zero.

2. The uncertainty is massive. Near-term out-of-the-money options have pretty much zero market value compared to the shares they control (i.e. you can control 100 or 1,000 times your capital in shares) why long-term in-the-money options can reach 1/3 to 1/2 the value of the shares. It is an order of magnitude of uncertainty. The market value therefore is incredibly low signal.

There is a very large chance that these were one- or two-month out OTM puts to hedge Iran, in which case they are like 3% of what the 13F says they are.

Leopold’s Real 13F

So, how should we read 13Fs?

STRIP OUT ALL OF THE OPTIONS. SHARES ONLY. ONLY SHARES.

Here was Leopold’s 13F from last quarter. Meaning what he owned on December 31 2025.

As you can see, this is basically your classic super concentrated thematic equity portfolio. Bloom, Lumentum, and CoreWeave pretty much make up half of the portfolio. The rest make up the other half. This was that big Bloom bet that he took in the beginning of the year. Notice that we’re completely ignoring the Intel calls and the CoreWeave calls. They look optically very large in the December 13F, but again, we have no idea how large they are.

Here is how his holdings performed during Q1. As expected, Sandisk and Lumentum absolutely mog everyone else. The Neoclouds and miners are the laggards. Leopold has a pretty high affinity for the Neoclouds and miners and regularly adds to them, even through some rough performance over time. He gets his god-like status purely from his bets on the vertical chart crew (aka Sandisk, Lumentum, and Bloom). And honestly, that’s enough because I’ve never seen anybody successfully bet on all three. I am a Big Bloom Bull (BBB), but I came in late.

Here is his portfolio if he’d held the exact same equities through the entire quarter. His Q1 return would have been 27.3%. This is interesting because, despite a very large divergence in the performance, it looks relatively similar. At least everything outside of SanDisk. That’s really the only difference. SanDisk now takes up 14% of the portfolio, where before it was 7%. It is doubled in size. But of course, he partook in a common hedge fund activity known as “trading.”

This is his new portfolio. Notice anything different?

First of all, it is only slightly (5%) larger than the Q4 portfolio.

Second of all, both Lumentum and Coherent are gone. Tower Semi, which weren’t in the top 10, are gone as well. Every single optics play was liquidated, and because this isn’t options, this is actual shares, this is real signal.

Third of all, because everything optics is gone, even though he reduced Bloom slightly by around 20%, it still makes up around a quarter of the portfolio (bc shares appreciated). His relative conviction on Bloom hasn’t really changed. The similar thing can be said for Core Scientific: reduced very slightly, but now actually makes up a larger portion of the portfolio.

Continuing on this train of thought, anything kept or slightly added to actually now takes up a larger portion of his portfolio. This includes CoreWeave, IREN, and Applied Digital.

However, the standout here is SanDisk. Not only did it appreciate materially more than the others, but he also added to it through the quarter. It is now his second-largest position, right between CoreWeave and Bloom Energy.

Below the paywall, I will attempt to read the mind of Leopold and tell you what he is thinking. I think I have read the original Situational Awareness essay like five times, and I followed every single 13F since Q2 2025, soooo I think I’m pretty qualified.

Contents

1. Optics Exit

2. Bitcoin Miners

3. NAND

4. Energy

Read more

I don’t understand why this topic isn’t talked about more. In public equities, we look at every company, calculate their 2-3 year out revenues and earnings using detailed supply chain analysis, evaluate their competitive positioning, analyze their engineering in data sheets, build multi-tab quarter-by-quarter financial models, and apply the fairest valuation that we can.

And yet even then, we can get price charts (which is essentially the change in valuation over time where investors say “we got it wrong”) that look like this.

Thousands of market participants and investors in mid-2025 were wrong about this one company’s valuation by a factor of almost 100 times. Think about that for a moment. These shares were changing hands for almost 1% of what they are worth today. They were wrong. So wrong. Dead wrong.

And they are wrong about Anthropic today.

The 1 trillion-ish valuation is not really questioned too much. But once you start asking questions, you realize that there is completely zero chance it is anywhere close to big enough.

Today, I will justify a $100 trillion valuation for Anthropic as a fun thought exercise to show just how undervalued the model labs are. In reality, it is completely silly to claim that a company already worth nearly a trillion dollars will increase in value over 100-fold and become 20x more valuable than the current most valuable company in the world. The point of this exercise is to show that it is logically plausible. It is within the realm of possibility. If the ceiling is this high, at least we must be thinking in the wrong order of magnitude today.

Abstract

Model Labs are the greatest businesses of all time. Tokens increase in value over time as frontier models become more capable and able to automate longer time-horizon tasks, unlocking new use cases and increasing willingness to pay. Meanwhile, they decline in cost over time due to both hardware efficiency gains and algorithmic improvements in the models themselves. The combined effect of higher value at lower cost increases profit non-linearly, creating a near infinite terminal value. They are able to defend this attractive business because their moat strengthens over time via recursive self-improvement.

Anthropic’s $45b ARR represents growth never before seen in capitalism’s history, and using a Gompertz-style power law curve, we estimate its exit CY26 ARR at $300 billion, its exit CY27 ARR at $1 trillion, and exit CY28 ARR at $1.5 trillion. Yet it is absurdly still valued on trailing and current-year financials. We believe that it is a fallacy to not apply a far out forward multiple to the model labs like we do with high growth high margin public software companies like Palantir and Cloudflare. Doing so, we arrive at our $100 trillion valuation. Despite being optically unreasonable, this valuation is theoretically supported by the idea of the technological singularity, a concept commonly cited by AI theorists but that not yet entered the investment discourse.

However, retail access to this company is heavily gated and the IPO dynamics are incredibly unfavorable to retail investors. This is why I share a trade I have entered providing near 1:1 economic exposure to Anthropic to hedge against the IPO.

Contents

1. Model Labs = Greatest Businesses of All Time

   1. Increasing Value Over Time

   2. Declining Cost Over Time

   3. Strengthening Moat Over Time

2. Revenue and Valuation

   1. Estimating CY26 ARR

   2. Public Markets and Forward Numbers

   3. Estimating CY27 and CY28 ARR

   4. Margins

   5. Valuation

3. The Singularity

4. The IPO

5. How Retail Investors Can Play This

This is a longer, higher quality article than usual. Because of the time it takes me, I will only be able to do one or two of these a week. Would you prefer two of these types of higher quality articles or five articles at my usual length and depth? Earnings coverage is not impacted.

SemiAnalysis Giveaway: The first person to subscribe using the paywall on this post will be selected to receive one free month of SemiAnalysis (sent to your email).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Model Labs = Greatest Businesses of All Time

The popular media has it backwards. Model Labs are the greatest businesses of all time.

Model labs share a unique combination of characteristics that provide them with a near infinite terminal value:

1. The value of their product rises over time.

2. The cost of their input declines over time.

3. The moat strengthens over time.

As a result, the leading model lab(s) are currently on a fast track towards a monopoly position on all future intellectual and knowledge-based labor, an addressable market comprising most of global GDP.

Increasing Value Over Time

As AI models become more powerful, the value delivered by Model Labs per token rises accordingly. This is intuitive, as smarter models can do more useful things.

As the task set that models can accomplish broadens, people start discovering new ways to use models that have never been discovered before. It is only because of improving model capabilities that the agentic AI phenomenon, widely considered to be as important as the ChatGPT moment, was able to materialize.

Two of the most important examples for how more powerful models have delivered value through agents are Claude Code and OpenClaw.

Claude Code

The term vibe coding was coined by Andrej Kaparthy approximately one year ago.

At that time, when the leading model was Sonnet 3.7, vibe coding was heavily looked down upon and discouraged. I tried to build apps with it too and failed miserably. It will run into bugs and then run around in loops, building up more technical debt and fixing nothing.

Today, with Opus 4.7, Claude code is completely conquering software engineering as it is quickly moving towards 10% of global GitHub commits.

And it hasn’t stopped at just conquering coding. The only difference between an LLM and an agent is its harness: what tools it is given, what software it can access via API.

Now, most knowledge tasks can be automated via these agentic harnesses (which seem to improve just as fast as the models themselves). Agents can perform knowledge work far cheaper than humans, as shown by SemiAnalysis in one of their most recent articles.

Which also led to Doug O’Laughlin’s famous Claude Code psychosis and Anthropic’s ARR compounding at 50% month over month in 2026.

OpenClaw

A similar phenomenon happened with OpenClaw. As the model capabilities enabled longer horizon tasks, an open-source harness that enabled such work inevitably gained immense popularity.

And while OpenClaw hype has died down a bit, usage has not. Funny enough, I am the best example of this. While I talk about OpenClaw way less nowadays, I use my claws more than ever before.

Cron jobs are especially valuable. For example, a common use case in consumer SaaS is to provision a customer service agent, running once per day to respond to requests and modify the account information of real users. Without the significant improvements and model capabilities that occurred over the past year, this would not be a realistic use case, and if it wasn’t realistic, the harness would have never been developed.

Declining Cost Over Time

As the value of a token has risen, the cost to produce it has declined.

NVIDIA’s hardware enables multiple OOMs of efficiency gains in inference every generation.

This is because, with each generation, we get more flops, more memory, better software, lower precision, algorithmic improvements, and a larger scale up domain. Each effect compounds on the other.

But it gets even better. While the hardware becomes more efficient, new models are also getting trained.

This is Leopold’s graph from 2024 because I couldn’t find one more recently that expresses this same idea lol. I can assure you that nothing has changed. The cost of an Opus 3.7 or GPT-5 level model is likely 1% of what it used to be, as DeepSeek V4 Flash is basically free.

Over time, as the value provided rises and the costs drop, profit increases non-linearly.

This is different from ordinary operating leverage, which is just part of the normal course for a business as it matures and has a structural ceiling. This phenomenon for Model Labs is structural and unbounded.

I really need to emphasize this. Every year that passes, the input cost declines while the value provided rises.

There is no other business in which this phenomenon applies. It is this that differentiates the Model Lab from the semiconductor industry. In a way, our revenues are their costs, and we sadly rely on Jevons’ Paradox to drive growth (look at the line). They do not require Jevons’ Paradox at all, as they in fact benefit from technological advancement in both directions.

Strengthening Moat Over Time

The economic argument against this “perfect” business model is, of course, competition. If the profits increase non-linearly, new entrants are also incentivized non-linearly as well, which is why my next argument is that the Frontier Labs also have a moat that strengthens over time.

Over the past two years, there have been several moments where the general consensus declared frontier models a commodity. The most prescient examples are the months following the Deepseek release and the disappointment over GPT-5 back in mid-2025.

The argument is really seductive. Think about it: all it takes is a bunch of compute and a bunch of data to train a frontier model. That means that anyone with tons of capital can do it. It just becomes a very capital-intensive commodity, something like memory.

However, over the course of late 2025 and the first half of 2026, this has been proved to be completely untrue. OpenAI and Anthropic have decisively regained the lead in agentic model development, leaving even their relatively more well-capitalized competitors far behind, and are now even widening the gap.

Google Gemini briefly achieved SOTA status in November but has quickly exited the discourse as they did not have a competitive agentic coding offering. xAI and Meta used to be near the frontier in 2025 but are now significantly lagging as well. It is only Anthropic who has grown their ARR 5x in 2026. None of the other labs are even close to having the same level of commercial success. This shows that by providing a superior product, one company is rapidly gaining market share and becoming vastly more difficult to dislodge by the day.

Open Source

Now I will address the common counterargument that open source models are within striking distance of the frontier on benchmarks while being a fraction of the cost, and instead, show that the gap has been widening over time.

There are three reasons why benchmarks are imperfect:

1. Benchmarks are not representative of real-world tasks. Humanity’s Last Exam and similar benchmarks measure esoteric knowledge in a multiple choice setting not whether or not you can do an economically valuable task, like update an Excel spreadsheet with financial data pulled from a 10-K.

2. Coding benchmarks, like SWE-bench, have a dual-sided sample problem. Either the coding issues that were scraped from GitHub are not well-defined or are flawed, or the models started memorizing the answers to these benchmarks themselves.

3. Most importantly, labs are incentivized to hill climb these benchmarks in RL, therefore optimizing for a high score rather than for real-world viability.

The best measure of the capability of the model is probably just simple market share. If people are willing to pay for it, it is likely because it is creating value for them.

Using open router data, we can see the token share of the various labs and then back-solve to get the spend share based on the price of those tokens. And this is where we see Claude completely dominate.

In addition, keep in mind that this is for open router only. Openrouter only makes up about $1.5 billion of annualized token spend, which is peanuts (1/30 of Anthropic’s current $45 billion ARR).

If these open source models were really at parity with a fraction of the cost, why would anybody pay Anthropic a single dollar?

Recursive Self-Improvement

So, what is the structural reason behind the widening lead of Anthropic?

I believe it is recursive self-improvement, a concept you have probably heard of somewhere in the AGI-pilled-San-fransisco-X discourse.

The simplest way to frame it is that although every company has access to frontier tokens, model labs have it the cheapest. They can consume their own tokens at cost, instead of paying 80% gross margin (a 5x markup). As the value of tokens rises over time, access to tokens becomes a larger and larger portion of the success of any firm and the input to any good or service. Model labs take advantage of it the most.

The evidence for this is multi-fold.

First, from a vibes-based perspective, Anthropic has clearly been shipping like crazy in 2026. There was that famous story where they shipped Claude CoWork in two weeks because they had Claude Code. That was back in January, and everyone was insanely impressed.

They also literally caused the entire SaaSpocalypse.

However, it’s not just vibes. They have objectively been shipping faster.

Perhaps the far more important application of recursive self-improvement is the automation of model development and AI R&D itself. This concept is heavily outlined in the “AGI maxi” literature through pieces like AI 2027 (2025), SITUATIONAL AWARENESS (2024), and The Coming Technological Singularity (1993!!).

The idea is that really smart AI will make AI research faster, which will then develop even smarter AI, and so on. This is the general concept known as the singularity. It was mostly theoretical, so discussions around it as an input to investment analysis were pretty much non-existent. But now Anthropic is actually making this happen.

You can see a clear speedup in model release cadence after Opus 4.5, widely considered to be the first truly autonomous agentic model. Mythos was also a step function above the current scaling laws, and the most important model release in this timeline (even more so than Opus 4.5), and it just so happened to occur in 2026, after the launch of Opus 4.5.

We will discuss the singularity in more detail later, but in terms of relevance to Anthropic’s moat, the way to frame it is that this is a turbocharged version of Helmer’s concept of Process Power in “The Seven Powers,” which is the general competitive advantage possessed by firms with decades of learning by doing, like TSMC’s node leadership.

With a substantially more capable model than its competitors, Anthropic will also have a substantially higher R&D productivity, enabling it to widen the model gap further, thereby widening the R&D lead further. It is a recursive cycle that strengthens their competitive moat over time, until they reach the status of a truly unbreakable monopoly.

Compute

The most credible bear case against Anthropic is that they are GPU-poor compared to OpenAI and Google.

Last year, OpenAI triggered the whole “AI bubble” bubble by signing massive deals with Oracle (5GW), AMD (6GW), NVIDIA (10GW), and Broadcom (10GW).

And Google is literally a hyperscaler. How are they even supposed to compete?

This is the vector by which my mind can be changed. I also have significant exposure to OpenAI-linked stocks because I believe that their compute advantage will help them outperform the currently low expectations.

However, if I had to bet on one of the AI labs, it would still be Anthropic.

The framework that I would use to think about the competitiveness of a frontier lab is a mix between compute and “taste.” Taste is a bit abstract, but think of it as an amalgamation of factors that affect how effectively the lab uses compute. This includes algorithmic efficiency, eval quality, knowing what to research, and general tacit process knowledge embedded in the firm.

compute x taste = effective compute

Anthropic has shown by far the most taste out of OpenAI and Google.

I believe that OpenAI, as many have properly criticized them for, is too distracted. They have too much emphasis on the consumer, pursuing ventures like acquiring TBPN and literally making a phone. (Like, what? Do you not realize that if you reach AGI first, you can make a phone way faster? Lmao)

Google also shares the same consumer-focused problem pouring resources into image, video, and world models, but it is worse because they don’t even have a competitive agentic harness product yet that competes with Claude Code and Codex. Maybe this will change soon with the Big Google Event (I/O), but they really have to pull a rabbit out of a hat to unwind six months of OpenAI and Anthropic’s lead.

Anthropic, meanwhile, has shown relentless focus on the one thing that matters: making more intelligent agentic coding models. They haven’t even produced an image generation model. Think about that! I use GPT Image Gen 2 every day, and it’s incredibly useful. Clearly there is demand out there for such a product, and Anthropic purposefully turned away that demand to maintain their focus.

It has been clear that this focus and taste has allowed Anthropic to develop more powerful models with less compute. I do not expect this trend to stop. In fact, having the most powerful models might even help them catch up in compute, as having a higher commercial market share, they would be the entity with the highest willingness to pay for a marginal petaflop.

Mythos

Now to tie this all together.

Every factor here is only working in Anthropic’s favor because they currently, according to publicly available information, have the world’s most powerful model, Claude Mythos. To be frank, if Anthropic was public, the release of Mythos should have made Anthropic stock surged by 1,000% in a single trading session, as it is the single event that significantly increased the probability of Anthropic achieving recursive self-improvement before everyone else.

On the launch of GPT 5.5, results for SWE Bench Verified and SWE Bench Pro were excluded from the model card because it significantly underperformed Claude Mythos.

However, I did say benchmarks were imperfect and real-life use matters more, but actually, this is where Mythos truly mogs.

During our testing, we found that Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so. The vulnerabilities it finds are often subtle or difficult to detect. Many of them are ten or twenty years old, with the oldest we have found so far being a now-patched 27-year-old bug in OpenBSD—an operating system known primarily for its security.

The exploits it constructs are not just run-of-the-mill stack-smashing exploits (though as we’ll show, it can do those too). In one case, Mythos Preview wrote a web browser exploit that chained together four vulnerabilities, writing a complex JIT heap spray that escaped both renderer and OS sandboxes. It autonomously obtained local privilege escalation exploits on Linux and other operating systems by exploiting subtle race conditions and KASLR-bypasses. And it autonomously wrote a remote code execution exploit on FreeBSD’s NFS server that granted full root access to unauthenticated users by splitting a 20-gadget ROP chain over multiple packets.

- Anthropic

You all have probably heard that Mythos found hundreds of zero-day vulnerabilities across the world’s most secure software infrastructure. This is the first glimpse of a model with vastly superhuman capabilities in an intellectual domain.

Importantly, this is not because Anthropic purposefully trained a cyber security model. Mythos simply inherited these capabilities due to being much more intelligent in general.

“We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy. The same improvements that make the model substantially more effective at patching vulnerabilities also make it substantially more effective at exploiting them.”

- Anthropic

Anthropic is most certainly utilizing the world’s most powerful model (that no other lab has access to) to conduct internal R&D right now. It will not be long before we see the results.

Revenue and Valuation

Estimating CY26 ARR

Currently as of May 2026, Anthropic is at around a $45 billion ARR according to sources and stuff.

They are undergoing another round of private fundraising. At their proposed new valuation of 900 billion, they would trade at 20x current ARR. This seems expensive to the general public who may view this as a hallmark of the “AI bubble.” So let’s put it into perspective.

They started the year at a $9 billion ARR. That is an annualized growth rate of (lol) 12,400% or a (very convenient) month-over-month growth rate of 50%.

Anthropic mentions their current ARR from time to time, and we can map out each of their public data points to approximate the month-over-month growth. The implied monthly growth intervals are roughly 43%, 51%, 46%, and 59%. So, it fluctuates around 50%, but doesn’t seem to be slowing down.

Let’s build a quick sensitivity table for what their ARR would be if they kept this growth rate or it slightly slowed down.

As you can see, if they kept up 50%, they would have a trillion dollars in revenue. I mean, that’s not even impossible. The data doesn’t show it slowing down and the singularity could be here anytime.

But for realistic purposes, I would assume that it slows down just based on first principles and the laws of gravity. We’re going to use something between a 20% and a 30% month-over-month growth rate for the rest of the year as our baseline assumption, which will put them at $300 billion in exit ARR. At a $1 trillion valuation, that’s only 3.3x exit 2026 ARR.

Wow, that valuation changed quickly, didn’t it?

Public Markets and Forward Numbers

Oh, another thing that we almost forgot about. Remember when I was talking about how we public investors just forget everything that we’ve learned when we go and visit the private side? Well, one of those things is the fact that we all happen to use 2027 and 2028 multiples on all of our semi stocks. When was the last time you valued a company like Lumentum on what they would earn in 2026?

So why is Anthropic any different?

This happens for a very simple reason: there is no analyst coverage on private companies.

Without “consensus” numbers to go off of, trailing metrics are the markets’ only indicator, which are, by and large, completely irrelevant. There is zero signal whatsoever in what Anthropic or OpenAI earned in 2025, just like there is zero signal in what Lumentum or Bloom Energy or AXTI earned in 2025. Yet it is the only thing visible to the public, so it is the only thing that investors anchor to.

If we apply the public markets’ approach, even partially, and use exit CY26 ARR, we already arrive at the conclusion that a one trillion valuation is absurdly cheap, as it is only a single-digit multiple, but we don’t have to stop there. Arguably, we should look out even farther for these AI labs, as they are the definition of low revenue today but high terminal value. What about exit CY27 ARR or even exit CY28 ARR? I’ve seen sell-side broker notes valuing Palantir and Snowflake on what they would earn in 2035. Shouldn’t Anthropic get that same treatment?

I say exit CY28 ARR is a happy medium. Let’s run a sensitivity on what that number could be.

Estimating CY27 and CY28 ARR

Let’s model Anthropic’s revenue as a power law. Given three data points:

• $9 billion starting ARR

• $45 billion ARR at the start of May

• $300 billion ARR at the end of 2026

we can fit this to a Gompertz-style curve where the growth rate decays exponentially over time.

Using this Wacky Ahh Equation, we can get to approximately almost exactly one trillion dollars of exit CY27 ARR.

I love it when our numbers are convenient. This is also almost exactly a tripling in revenue.

What about 2028?

And our final result is $1.5 trillion of ARR exiting 2028. Almost exactly 50% growth.

Margins

Now let’s discuss Anthropic margins.

Anthropic’s 2025 gross margin is commonly cited at 40%. Which leads many people to believe that frontier tokens are a relatively low-value-add product, and that every $1 of incremental token revenue generates only $0.40 of gross profit dollars. This is technically incorrect and very misleading to what the actual unit economics of their business is.

I would like to introduce you to a concept known as contribution margin. Cost of goods sold is generally seen as the variable costs. However, COGS actually often has fixed components. It is simply an accounting bucket meant to provide a reasonable standard for GAAP and not to be economically accurate to the actual variable cost of a business.

For Anthropic, the economically fixed portion of the cost of goods sold is essentially any unused capacity plus freebies given to users:

• reserved AWS/GCP/Azure/TPU/GPU capacity paid for whether fully used or not

• idle capacity, redundancy, peak-load headroom, and latency headroom

• low batching / lower utilization for fast interactive workloads

• subscription overuse and free-tier usage

• production infra staff, SRE, safety/moderation, routing, storage, networking

The economically correct metric is called the contribution margin. This is the very intuitive revenue minus variable costs, what a company earns for every unit of a product it sells. This is how we should evaluate token economics. According to SemiAnalysis, this is now above 70%, so I believe 75% is a fair base case for a contribution margin assumption today, expanding to 80% over the next two years.

Cost of production for token has fallen sharply because increases in accelerator pricing generation-over-generation have been more than offset by much higher throughput (tokens/sec/gpu). Average blended price per million tokens has fallen dramatically over the past few months, agentic workloads are inherently multi-turn with longer input/output ratios and higher cache hit rates, but inference margins have gone up from < 40% to > 70% in the same time frame.

- SemiAnalysis

Currently, Anthropic is GAAP unprofitable, with the model lab’s enormous cash burn being cited as a lack of favorable unit economics. However, the cash burn is actually caused by the training compute spend, aka the “R&D” of the leading labs. R&D in GAAP is expensed in the period incurred. However, economically speaking, it should really be treated more like capex, as it is an investment in a future model which pays back over many years. In R&D-heavy industries, GAAP net income can be extremely misleading as it does not account for the economics of this investment.

This is why I believe Net Income or EPS is not the right denominator on which to value the Model Labs. Instead, we should look at their contribution margin, and basically say how much is the business worth if they stopped investing today and served the models they already had?

80% of a $1.5T exit CY28 ARR is $1.2T in contribution margin income.

Valuation

What would you pay for the best business of all time? Assuming Anthropic has achieved recursive self-improvement and stays #1 (i.e. infinite moat), this is actually an accurate descriptor.

Let’s look at what people already pay today for the highest quality highest growth recurring revenue businesses: Palantir and Cloudflare.

Palantir is currently being valued at 45x 2028 earnings.

Cloudflare is valued at 90x 2028 earnings.

Now, a high double-digit out-year earnings multiple might seem extreme to most of us, but there’s actually a reason these top-quality software companies are so expensive. If you have confidence in a decade or longer runway of highly above-average growth and you run it through a DCF, you end up with incredible numbers. 30x is just not enough to reflect that. The combination of high growth, high duration, and high certainty is that rare.

Because of what I said in our above section, I believe it’s very easily arguable that Anthropic is far above the quality of these software companies. However, let’s say we just apply a multiple on the upper range of the 45x to 90x bound: 83x.

$1.2T Contribution Margin Profit * 83x multiple = $100T Valuation

Anthropic is worth one hundred trillion dollars!

The Singularity

This is the singularity.

The concept is simple. Once recursive self-improvement produces the first superintelligence, the equivalent of a near-infinite superhuman labor force enters the economy. Productivity, which is quite literally the production of goods and services, now goes through the roof because we have a country of geniuses in a data center.

An important fact people forget is that GDP growth is not set in stone. The Industrial Revolution, which added capital productivity growth to a pure labor economy, increased GDP growth from mere basis points (the growth rate of the population) to a surprisingly stable 2-3% per year.

Now, as superintelligent AI introduces the automation and growth in productivity of cognitive labor itself. I expect another step function change in GDP growth, perhaps to 5%, 10%, or even higher. As famously stated by Irving John Goode in 1965: “Thus the first ultraintelligent machine is the last invention that man need ever make.”

This is not a concept that us investors talk a lot about, but many of us privately believe in (me lmao). I don’t think this should be the case. If it is a real possibility, we should give it real consideration in investment analysis.

AI 2027 describes the post-singularity economy with the most vividness and detail of any essay about the singularity that I found, featuring special economic zones (SEZs), robots, factories, drones, and fusion.

“By the end of the year they are producing a million new robots per month. If the SEZ economy were truly autonomous, it would have a doubling time of about a year; since it can trade with the existing human economy, its doubling time is even shorter.”

“By late 2029, existing SEZs have grown overcrowded with robots and factories, so more zones are created all around the world (early investors are now trillionaires, so this is not a hard sell). Armies of drones pour out of the SEZs, accelerating manufacturing on the critical path to space exploration.”

“Robots become commonplace. But also fusion power, quantum computers, and cures for many diseases. Peter Thiel finally gets his flying car. Cities become clean and safe. Even in developing countries, poverty becomes a thing of the past, thanks to UBI and foreign aid.”

- AI 2027

What are the implications for Anthropic’s valuation?

If AI accelerates global GDP growth to 10%, that is quite literally a tenth of the global production in goods and services being added every single year by AI itself.

That is a lot of value. An unimaginable amount!

All of it is enabled by the frontier models and the labs behind them. If you are the leading monopoly AI lab, which is performing this economic acceleration, which is bringing about a robot economy and curing diseases, how much should you be worth?

I do not have 100% conviction that Anthropic will be worth $100 trillion. I don’t have conviction that any company will be worth $100 trillion. But I do have 100% unwavering confidence that the combined value of the one or multiple leading AI labs, whether it’s Anthropic, OpenAI, Google, all three, or another company we haven’t even heard of, will be worth that much once the singularity arrives.

The IPO

We spent the first half of the article discussing the why behind Anthropic and the Model Labs theoretical valuation ceiling. Now let’s move on to the how. These companies are private. What can we do to get exposure?

Cerebras taught all retail AI investors a nasty lesson. The system is not built to help us.

Cerebras began at close to $100 a share, but due to popular demand that outstripped available supply by 20x, they increased their IPO pricing to $135, then $150, then $185.

But it gets a lot worse because the IPO price isn’t the price that you can buy on your brokerage on the day the ticker starts trading when the shares actually go public. A market is made, so the actual price it starts trading at is whatever the first buyers and the first sellers agree on in the order book.

And for hot IPOs like Cerebras and definitely Anthropic or OpenAI, there is going to be a ton of buyers at the same time. Nobody’s going to want to sell. This means that the shares could open at high double-digit or triple-digit above what the IPO price initially was, as buyers relentlessly pour in.

Cerebras, after being priced at $185 for the IPO, opened at $380. Imagine being a retail investor here. You heard that you can buy Cerebras for a $25 billion valuation, maybe did your homework, realized it was a good deal, and got excited. When actually pulling up the stock to trade it on Robinhood or something, it is four times more expensive than what you expected.

I think this is almost guaranteed to happen with Anthropic and OpenAI. From here, I don’t see a path to a $2 trillion private valuation before the end of the year as unlikely. From there, the IPO price is probably slightly higher than that, and the initial trading price might already make them one of the most valuable companies in the world, on par with the Mag 7.

How Retail Investors Can Play This

Next, I will share a trade which allowed me to get dollar-for-dollar exposure to Anthropic with the least catches out of any single backdoor Anthropic play that I’ve looked at.

There are holdcos where you get a tiny fraction of the per-share value as Anthropic and there are closed-ended funds that trade at hundreds of percent of NAV. The problem with these plays is that they are incredibly capital inefficient. You can’t get exposure without fronting a significant portion of your portfolio or even taking on leverage just to have a sizable economic exposure to Anthropic. This is substantially worse than owning Anthropic stock purely, because you either have a fraction of the exposure or you must lever up significantly in which case you are completely path dependent on the share price not going down 40%.

There are perps exchanges that charge you hundreds of percent in funding rate to go long (cough cough hyperliquid). Essentially, these perps exchange funding rates act as an incentive mechanism to discourage longs and encourage shorts, to keep the price of the futures contract tethered to the oracle price. Basically, what this means is that at the current oracle price, which is the private value Anthropic mark, there are vastly, vastly more longs than shorts. Even with a conventionally extreme incentive to short, longs still vastly outnumber shorts. This is just a convoluted way of saying the price is way too low.

However, with a little bit of financial engineering, I came up with a creative way to get 1x exposure to Anthropic without these problems. No, it is not buying Anthropic on the private markets and actually owning the shares yourself, but it works reasonably well enough, and I’ve put a sizable portion (5-10%) of my portfolio into this.

Below, we will analyze the company that is the subject of this trade and do some serious quantitative analysis on the trade’s mechanics and potential payoff.

Read more

They are worth over seven or eight billion dollars once you account for the dilution. Currently, they’re at about 100 million in revenue run rate. Lol.

But okay okay, let’s forget about any and all numbers for a second.

What if I told you that there is one chemical compound that is needed for all of optics? That there is no way around it, and that it would scale exponentially once scale-up CPO hits, and you will need big massive lasers made out of this compound stuffed deeper and deeper inside of a rack.

And there’s one company that holds a substantial global share of this compound and has tons of brownfield capacity available and can convert capacity from another compound to this compound and can build greenfield once its brownfield capacity is saturated and is basically doubling capacity year over year.

Interested? That’s what I thought.

And that’s what the market also thinks too, as they have returned a cool 7x since the start of this year.

And their 12-month return?

So yeah, they confuse the shit out of me forever, but it is finally time to look at what they actually do. Funny enough, the fact that they have kept running while the general sentiment on X and Substack was extreme skepticism and bearishness probably means that the bull case is underrated. This company makes SNDK look like a stock you would buy after reading The Intelligent Investor by Ben Graham.

Deep coverage on blue chips refers to LITE, BE, INTC, big memory vendors, etc.

Thematic work includes:

• industry/tech overviews

• investment maps comparing all of the players in an industry

• economic style thinking and theory for AI infra

• macro

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Introduction

The agenda for today is simple. We will first talk about the fundamentals of the company and then review their most recent earnings call to see how they’re doing.

AXTI makes indium phosphide substrates, which are the raw materials needed to make data center lasers, such as EMLs (transceivers) and high-power CWs (CPO).

Which happens to be the number one bottleneck in all of AI at the moment. AXTI is the bottleneck of all bottlenecks. The bottleneck final boss.

The spot price has ripped, but I think it undersells how bad this constraint has gotten. The chart isn’t even as atrocious as memory today. Let me just say that there is a reason Lumentum is now 30% behind demand on EMLs and probably over 50%, (depending on how you interpret management’s commentary) behind demand on their scale-across laser components.

For an introduction on why indium phosphide is needed for lasers in the first place, please see my Sipho introduction. The gist of it is that there is basically only one semiconductor chemical in the world that’s able to produce light at the required wavelength for long-haul data center communications, and that is indium phosphide. Silicon can’t produce light at all. Gallium arsenide is the other choice, but it will increasingly be relegated to niche shorter-reach use cases. I will skip over a lot of the technical stuff for this article today because I want to mainly be focused on answering the question of why the hell this stock is up so much instead of why this chemical over this other chemical and how the chemicals are made.

Jason's Chips

Short n' Casual Intro to SiPho

SiPho is pretty cool ngl…

Read more

3 months ago · 75 likes · 2 comments · Jason's Chips

AXTI controls roughly a third of this market and has by far the most vertically integrated supply chain out of everyone who makes indium phosphide. Their operations are primarily based in China through their subsidiary Beijing Tongmei, and they control a wide swath of raw materials companies all across the indium phosphide supply chain. They control everything including extracting the stuff, refining the stuff, and making the stuff super pure. This gives them cost and speed advantages.

They process the stuff like this:

1. Tongmei/JVs source and refine high-purity indium and phosphorus.

2. AXT combines them into InP source material.

3. They grow a single-crystal InP boule/ingot using proprietary VGF, or Vertical Gradient Freeze.

4. They dope it depending on spec, for example sulfur-doped n-type for laser substrates or iron-doped semi-insulating material for other uses.

5. They slice, lap, etch, polish, and inspect the wafers for diameter, resistivity, orientation, thickness, and low defect density.

And last but not least, their entire destiny is dictated by export licenses, lol.

Earnings

Now, on to the earnings review.

Read more

Anyways, both of them had massive beats and are looking great. Let’s review the earnings for both today, as we can learn a lot from them in terms of both individual company performance, but more importantly, read throughs to their broader ecosystems (neoclouds for Nebius and optics for Tower).

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Headlines

Nebius

Nebius is the smaller (only revenue wise as their market cap flipped CRWV) sibling of CoreWeave and the favorite child of investors.

They are better at serving the long-tail of AI demand instead of only signing big deals and are executing MUCH better on several fronts (and thus have the Super Ultra Premium Valuation). I am biased towards CoreWeave due to my positioning so this is objective. Very rarely do I make a call that contradicts share price momentum.

First, they reiterated their $8b and $3.2b guidance for YE26 ARR and revenue respectively (compared to CoreWeave’s $18.5b and $12.5b). They guided their EBITDA margins to 40% for 2026 (vs 60% for CoreWeave due to scale). They are sold out for 2026 capacity as well.

Now on to how they’re executing really well.

Nebius significantly raised their contracted power guideline to 4 GW. This is an entire gigawatt more than what they gave for their Q4 earnings release. Despite having a smaller (less than 50%) base than CoreWeave today they’ve contracted more power. As a read through of Neoclouds in general, it shows that contracting power is really never the issue anymore, as both public Neoclouds surprised to the upside on their contracted power. Nebius also tends to self-build a lot more, and they have a big 1.2 GW site coming online in Pennsylvania.

Next is their increase in capex from a midpoint of $18 billion to a midpoint of $22.5 billion. Unlike for CoreWeave, this is due to higher projected 2027 revenue rather than cost increases, though they did cite around 3-5% of cost pressure for 2026 from component pricing (memory lol).

For financing, they raised $4b from converts and $2b from equity in the quarter, and are now tapping into debt as they can sign asset-backed loans with their much larger contracts, like the new $27 billion deal with Meta.

Tower Semi

Tower had a pretty sizeable beat on the Q2 guide.

But the real story is that they signed $1.3 billion of Sipho revenue for 2027. This is a massive revision upwards from sell-side and SemiAnalysis projections. It is actually quite close to my own projections from my original long thesis ($1.44b) before I got peer pressured to revise it down lmao.

On the call, they said that they are backed by even more contractual revenue in 2028. Pretty crazy stuff.

They are continuing on their goal to 5x Q4 2025 shipments by year-end 2026 and are on track to execute their $920 million investment plan. Today, their fabs are running mostly clustered around 60 to 80% utilization, which means that they’ll most likely hit capacity limits soon and join Lumentum in the sold-out club.

Jason's Chips

Lumentum Earnings Review | Being Sold Out is Now A Problem

It’s LITE day (again)! Let’s go…

Read more

2 months ago · 46 likes · 5 comments · Jason's Chips

Call Takeaways & Read-Through

Now on to the meat of our analysis, which is on their Q&A and the read-throughs that we can glean for the rest of their peers.

For Nebius, I talk about the effects of stronger GPU pricing, timing of their capacity additions at their Pennsylvania campus, and the structure of their Meta contract.

We read this through to CoreWeave and Neoclouds in general for what this means for future contract structures (with my own speculation) and whether or not there is a real cap on capacity. I share how this changes my view of and positioning on CoreWeave.

For Tower Semi, we discuss their interesting comments on their content in CPO versus plugables and a little bit of reading of management’s tone, clarify the $1.3 billion commitment with more granular details, unpack their market share comments, discuss pricing power, and talk about their positioning against TSMC COUPE.

We then read this through to the rest of the optics ecosystem and give one lesson I learned that we can take away and apply to how we view our favorite names.

Nebius

Read more

Yesterday, we covered Cerebras’ technology (dinner plate computing), which is a prerequisite to this article as we build upon the main conclusion from that piece.

Today, we will be going over the financial and business side of things and actually analyze Cerebras as an investment. This will include extensive testing of an actual model running on their hardware for real-world agentic inference tasks. Exciting right?

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

Financials

Cerebras IPO was 20 times oversubscribed at the initial $130 share price. People really want to buy this thing. Yesterday, they raised the IPO price to $160. Polymarket still has almost a 100% chance that they will close above a $50 billion market cap (that is for their basic equity value, which is only around 70% of their fully diluted, so it is way higher than what’s shown in the table).

Their 2025 revenue is miniscule and irrelevant at $510m. Currently, they have $24.6b RPO and $20b of contract value from OpenAI, which can be expanded over time, but this implies a $6.7 billion steady-state ARR.

At the assumed IPO price of $160 per share multiplied by the fully diluted share account, the implied revenue multiple to OpenAI’s ARR is 7.4x.

The First Noble Truth of Cerebras

The following is the first Noble Truth of Cerebras.

The financials are useless. Completely utterly useless. Forward estimates are useless. Growth rates are useless. Multiples are useless.

Now, why is that? Let’s think about this from first principles.

How To Think About Cerebras

Cerebras is a canonical example of an unproven technology early in its S-curve with potential for high double-digit to triple-digit growth. This high-level conceptual profile allows us to anchor to two companies that have plenty of coverage on this channel: Lumentum and Bloom Energy.

Lumentum

The 30,000ft intellectual framing of Lumentum is simple: one market (optics) is guaranteed (by the laws of physics) to replace another (copper) to achieve full dominance of a certain use case (AI networking).

Jason's Chips

The Lumentum Series | Part 3: The Genesis of Scale-Up CPO

So, I’ll introduce you all to a fun little method I use to study up on stocks…

Read more

3 months ago · 56 likes · 4 comments · Jason's Chips

Because the transition is inevitable, the leading players in the new market have a massive guaranteed growth runway. This is why optics are so hot.

At the same time, the growth is measured and measurable. It is easy to model. We can confidently predict that Lumentum has 50% year-over-year growth because we can simply use the growth rate of optics as an anchor.

Bloom Energy

Bloom is slightly different. One company provides an objectively superior solution for one single market, which has the potential to fully displace the incumbent.

Jason's Chips

Bloom Energy Thesis (Free)

As a Substacker, I am obligated to say that I am releasing this for free to provide an example of what kind of research you would receive with a paid subscription…

Read more

2 months ago · 79 likes · 19 comments · Jason's Chips

This is different from Lumentum because the market doesn’t change. Bloom plays in behind-the-meter energy, just like GE Vernova. They are direct competitors. There is no copper equivalent. There is no transition from one market to another.

Because of that, Bloom’s growth can be a little bit more unpredictable and bursty. There is no big Kager number to go off of. It is just a matter of how quickly customers adopt the new solution and how much, or if, the new solution is superior to the old one. However, growth here is still predictable. If the new solution is better, it will still slowly displace the incumbent.

Cerebras

Cerebras is completely different.

From our first article, we ended with the ultimate conclusion that Cerebras is a provider of a fundamentally different product than Nvidia: premium tokens. Premium tokens are way faster but cost way more.

Is Cerebras like Lumentum? No, there is no world in which premium tokens will inevitably displace normal tokens. They are simply two different products each with a different trade-off.

Is Cerebras like Bloom Energy? No. WSE is not a superior way to generate the same end product (electricity for Bloom, regular tokens for NVIDIA).

In both comparisons, Cerebras is actually much worse. It’s not guaranteed to take share from anyone, really. However, there is one critical distinction, and it’s very critical. Very, very critical.

The inference market is the single biggest market in the world.

AI networking and behind-the-meter power are big markets, but they’re not the biggest. Inference is, and the current incumbent is the biggest company in the world. If these premium tokens carve out a large enough slice of this overall pie, Cerebras is a very attractive investment. The only question is, will it carve up a larger left slice? Again, unlike Lumentum, it is not guaranteed to take any share at all.

Now you probably understand why I say that the financials are useless. Single-digit billions of ARR do not matter at all. If you are unprofitable and your market is the biggest market in the world, what matters much more, and really what matters at all, is the slice of the fast inference pie.

No amount of revenue will save Cerebras from offering a fundamentally unwanted solution. No valuation is too high if they serve a sizable fraction of all inference.

Therefore, the point of this article is to simply assess how likely it is the fast inference market ends up being incredibly large.

Abstract

To assess how likely it is that the fast inference market ends up being large, we must observe and test how performant and useful fast inference is today.

There is significant alpha in the testing. Just by observing the everyday experience of using the model, we can literally feel the components that bottleneck inference (hint one of the market’s current favorite themes) and reason about whether Cerebras is positioned to solve it.

For my testing, I completely disregard classic benchmarks that are unrepresentative of actual work and instead focus on two real-world tasks that I perform every day: equity research and coding. For equity research, we ask the model a classic retrieval question, and for coding, we have it develop a feature in a (real!!) live application and compare them side by side.

Through these two tests we observe what actually bottlenecks inference speed on a day-to-day basis and link this to our silicon world. We evaluate 5.3 codex Spark on both its real vs claimed speedup and the performance relative to the frontier.

We synthesized the results from our tests into a conclusion about the speed and the performance quality of 5.3 Codex Spark and Cerebras hardware.

Then I underline three potential bull cases:

1. The undiscovered TAM, where nobody knows what fast inference could be used for, so we are underpricing it today. I describe three potential market fits:

   • Fast fundamental investing

   • Embodied and human-like AI

   • Real-time human augmentation

2. The low-hanging fruit: architectural innovations that Cerebras can easily make, which would be a step function change in their unit economics. These include FP8/FP4 support and hybrid bonding.

3. Why they are the only choice for the non-NVIDIA ecosystem.

Most importantly, I culminate both yesterday’s and today’s research into a final conclusion and what I personally am doing at IPO.

Contents

1. 5.3 Codex Spark

   1. X Sentiment

   2. My Own Testing: Equity Research

   3. My Own Testing: Coding

   4. Conclusion

2. Bull Case

   1. The Undiscovered TAM

   2. The Low Hanging Fruit

   3. The Non-Nvidia Ecosystem

3. Conclusion and What I’m Doing at IPO

5.3 Codex Spark

5.3 Codex Spark is the custom-developed model by OpenAI, specifically made for running fast inference on Cerebras hardware. It is currently the only custom design model for Cerebras hardware, so in reality this model is one of, if not THE most important element of any Cerebras investment analysis. But for some reason, it is very under discussed.

It trades a large chunk of its performance to be smol. It features substantially lower benchmark performance, a context window that is one-eighth of the leading frontier models today, but with inference speeds of over 1,000 tokens per second.

As we established in the last post, in order to run a model of a certain size, a certain number of dinner plates must be strung together. The larger the model weights, the more wafer-scale engines are needed, and the more uneconomical it becomes. Therefore, small models are much more cost-efficient for Cerebras to run.

Now, how good is this model exactly? Here is where I found something funny.

Read more

Over the next two days we shall do an analytical analysis of Cerebras in advance of the IPO. Today’s article focuses on their technology and tomorrow’s will be on stock-related stuff (business, market relevance, and financials).

Introduction to Dinner Plate Computing

Cerebras is the well-known pioneer of dinner plate computing.

Their foundational premise is that, BIGGER IS BETTER.

Normally, when you order a wafer from TSMC, you slice up your wafer into dies. Cerebras decided to keep all of the dies intact and instead network them together into one really big chip.

These spaces between the dies are called keep-out zones because normally you keep everything out of them because they will be cut up by the diamond saws. In this case, these keep-out zones are the home for complex wiring and networking. This means that Cerebras needed to collaborate intimately with TSMC to develop proprietary IP in order to fill up these keep-out zones with networking.

Throughout their investor materials, they make a plethora of comparisons to traditional computing with a bunch of metrics of stuff that chips do (compute cores, transistors, memory bandwidth, FLOPs), and because their chip is bigger, they can do more chip things.

This can mislead retail investors, as in real life there is always a trade off. Cerebras is simply a computing architecture optimized very differently than traditional GPUs.

Abstract

Today’s article focuses on the science and technology behind the Cerebras Wafer-Scale Engine.

First, I discuss the Cerebras distributed SRAM problem, and how this leads to extreme compiler complexity and inefficiency. I make the comparison to NVIDIA’s GB200 clear through an analogy: the tale of two cities.

Next, we find out how Cerebras claims a 100% yield rate on their wafer-scale engines, despite wafers being nearly guaranteed to have a defect.

Then we dissect PVT (process, voltage, temperature) calibration, one of Cerebras’s most important innovations, one that solves a problem which prevented the entire wafer-scale industry from taking off for decades.

After that, we talk about the science behind the one mechanism which enables their 15x inference speedup: pipeline parallelism. We explore how a normal GPU runs inference on a transformer (tensor parallelism), walking you through what happens inside a rack per-token and per-model-layer, before contrasting it with the Cerebras WSE approach (pipeline parallelism) and how data literally flows down the wafers like water through a river.

Following which we detail the two key constraints their architecture has, first being severe memory constraints and second being the limited I/O bandwidth due to very large area relative to perimeter.

And finally, we go one step beyond most other technical research out there. We underline why all this technological analysis is economically relevant for their product. I will walk you through the single conclusion that follows from our research, which forms the foundational premise for all of our economic and business-related analysis coming tomorrow.

Importantly, I make it all fun and understandable.

These two Cerebranalysis articles are a must-read. Get strapped in. This stuff is crazy technical but technical is fun. Let’s go!

Contents

1. The Tale of Two Cities & The Nightmare Compiler

2. Wafer-Scale Yield

3. PVT Calibration

4. Fast Inference (Tensor vs Pipeline Parallelism)

5. Memory Constraints

6. The I/O Bandwidth Problem

7. The Economically Relevant Conclusion

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

The Tale of Two Cities & The Nightmare Compiler

Read more

The cool thing about CoreWeave is that their earnings are super high signal, unlike a lot of other AI infra plays.

This is because of two main reasons:

1. CoreWeave just gives you a ton of visibility. On the Q4 call, they guide it for 2026 full-year revenue, 2027 revenue, exit ARR for 2026 and 2027, and the active power exiting each year. You could literally draw a line on the graph with management statements.

2. CoreWeave’s demand itself is a leading indicator for AI economics. Other companies that are upstream usually have much more predictability and certainty and therefore are lagging indicators instead of leading ones. While Memory, Optics, and Semi-Caps generally have revenue dynamics stemming from supply chain shenanigans, CoreWeave’s topline and backlog are completely dependent on end user demand for AI, which makes them much more unpredictable and high signal.

I made a nice tier list for all of CoreWeave’s key KPIs that are announced during earnings every quarter.

First is the S-tier. This includes the backlog, FY26 and FY27 revenue guide, and ARR Exiting FY26 and FY27.

The reason these are S-tier is simple: it’s literally revenue. Top-line growth is a majority of what CoreWeave is valued on, and it is the very first leading indicator for downstream end-user AI demand.

After announcing the Meta, Anthropic, and Jane Street deals, sellside pencilled in around $100 billion of backlog, which is exactly in line with what they announced. No surprise there. But the all-important revenue guide and exit ARR are still yet to be announced as they are on the call, not the print.

Second is the A tier. This includes near-term revenue from this quarter and next quarter, active and contracted power, which are derivatives of future revenue, and finally operating margins, as those are the most important margin number for CoreWeave. Here, they had a pretty solid top line beat, probably in line with by side whisper, though, while having it real beat on contracted power, which is a good forward-looking signal. The more important metrics are announced on the call.

Finally, we have the B tier, which are just nice to haves or second derivatives of A and S tier KPIs. You might notice gross margin is here. Why is that? Well, gross margin is very dependent on depreciation and revenue timing. As a result, the more often cited margin metric for CoreWeave is the operating margin, and gross margin usually takes a back seat to that.

And as expected, the stock does precisely nothing before the call. The call is where the real KPIs slide in.

The Call

Subscribe now

By accessing this content, you acknowledge and agree to our terms and conditions. This research is not financial advice.

RIP

The two big headline misses were:

1. The Full-Year Revenue Guide, which was $12.5 billion (reiterated from last quarter’s earnings call) instead of $13.5 billion expected by consensus.

2. Q2 revenue guide, which was $2.5 billion versus $2.8 billion expected.

This is definitely a negative result, no doubt. I am disappointed. Below we get into:

1. The reasons behind the miss. Including capacity bring up speed, memory pricing, and margins/timing.

2. Three mitigants to the negative result.

3. My new confidence level in CoreWeave, whether or not my long-term thesis changed at all, and if I’m doing anything with my positioning.

Read more