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

8 days 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

2 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

10 days 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

*written before earnings

ARM is in a very interesting position today.

I used to play baseball as a kid. I was never really that good, but I do kind of understand the sport.

Imagine your team is down by a point and it’s the ninth inning, and there are two outs. Both batters before you just hit a home run. Everyone is going crazy.

If you hit a home run, you tie the game. If you strike out, your entire team loses, and because two of your teammates just slammed it out of the park, you may or may not be feeling a little pressure.

You are ARM. Your teammates are Intel and AMD.

Well they’re not really teammates since they… you know. But I think the analogy still holds. Team CPU!

For Intel you could say their earnings sparked CPU mania. I covered them in depth here.

Basically the two big things were:

1. Intel crushed their top-line and revenue guidance by 10% via massive CPU shortage and price hikes.

2. They were the very first to spell out the agentic AI CPU to GPU ratio flip going from 1:8 in training to 1:4 in inference to parity or beyond in the agentic era.

AMD absolutely crushed it too. The headline is that they guided data center CPU to grow 70% year-over-year next quarter which was the engine of their top-line guidance beat. They also fully reiterated the CPU-to-GPU ratio. Literally the same exact 1:8 to 1:4 to parity thing that Intel mentioned.

But Lisa also made a comment that made ARM fly today.

“At our Financial Analyst Day in November, we outlined the server CPU market growing at approximately 18% annually over the next 3 to 5 years. Based on the demand signals we are seeing today and the structural increase in CPU compute requirements driven by Agentic AI, we now expect the server CPU TAM to grow at greater than 35% annually, reaching over $120 billion by 2030.”

As you can probably infer by now the point I’m trying to make is that ARM is under a ton of pressure.

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

Earnings Time!

Everything was sort of inline, which, if you compare to AI Infra peers, is kind of disappointing, but ARM has a history of doing this. Also, no one really cares about ARM’s near-term revenues or EPS anyways, given that they trade at like a gazillion times earnings.

In terms of moving the share price, what actually matters is what happens in 2075. Jk, more like what happens in the later part of this decade. The entire thesis is that CPU demand goes exponential as agentic activity goes exponential, and ARM would eventually have the majority share of this market. You could easily argue that it goes far beyond the $120 billion TAM forecasted by AMD. The more you multiply this, depending on how singularity-pilled you are, the more multiples you would get on ARM’s $9 EPS projection.

And there’s where we have some very good news. This is the one relevant signal in the earnings release.

Read more

The headline is that AI Bottleneck Co. beats and guides up, but it’s not enough to satisfy the market after their 2,000% run.

Pre-Call

Let’s look at what the print actually was.

A little disappointing. Seems like they didn’t beat buyside bogeys. This also means no random hidden pricing power anywhere despite the fact that they’re ultra supply-constrained. Operating margin is quite good.

Comparison is the thief of joy, but Bloom did beat in their actuals by 36%. Lumentum is supply-constrained. Bloom is not. This shows us that supply constraints can be as much of a problem as they are a tailwind for these companies.

This guide, though, is REALLY GOOD. If you think back to when they gave us longer-term guidance during OFC, they said that they would get to $1.25b of quarterly revenue run rate in 9-12 months (so March 2027 quarter).

They just guided to nearly a billion in quarterly revenue this quarter.

Let’s do the math. In the March quarter, they added $135m in revenue. This quarter they’re adding $185m in revenue. Their sequential absolute growth is accelerating. Now they’re only around $270m off. This means they might be on track to hit their target up to two quarters early, possibly in the September quarter.

However, for Lumentum, the vast, vast majority of their valuation is based on OCS and CPO, which are heavily concentrated in the out years. Nothing on the near-term financials of the print is actually that material for the valuation. This means that the call is far more important. It is the commentary on the call on OCS and CPO that will actually move CY27 and CY28 EPS estimates, which in turn moves the stock.

Below we talk about each of their business lines:

• EML

• Transceivers

• Scale-Across

• OCS

• Scale-Up CPO

They all oddly seem to be moving quite differently, which was interesting.

And then, of course, I give my take on whether this quarter was bullish or bearish and paste in my 200-line monster of a revenue build and income statement in Excel for people who are new subscribers and show what I updated.

Call

Read more

Still, I outperformed the index by over 2x this month with no(ish) leverage and over 3x YTD. A whole bunch of theses that I’ve heavily written about all started playing out at once. I am on massive hot streak of being right. Though I am bound to be disastrously wrong at some point due to probability.

My portfolio is going to get more and more unhinged as time passes, as I’m willing to risk losing a shit ton of money in exchange for an asymmetric pay-off in the event of the singularity.

Today we go into:

• The ten positions remaining in my portfolio and updates to my theses for them

• Which ones have gotten stronger and weaker?

• Which ones I’ve added to or reduced

• The two that I’m pondering an exit on and the one whose thesis died

• The four that I’ve exited since March

• Some hidden alpha on my positions not discussed in my other write-ups

By accessing this content, you acknowledge and agree to our terms and conditions. This publication, ESPECIALLY my portfolio, is not financial advice.

Read more

However, I also just really really love Bloom Energy as a company and think what they do is super cool, and to think that the world deserves to know why they are so cool. Enjoy!!

Contents

1. Why We Need Bloom Energy

   1. The Failure of the Grid

   2. The Failure of the Turbine

   3. The Requirements of the New Solution

2. Ten Features of Bloom Boxes

   1. Native 800V Direct Current

   2. Dynamic Load Following

   3. Modularity

   4. Absorption Chilling

   5. Carbon Capture

   6. Quick Deployment

   7. Low Latency Inference

   8. High Efficiency

   9. Data Analytics

3. The Unconstrained TAM

   1. Infinite Scalability

   2. Fuel Cells are a Technology

   3. Bloom is a Monopoly on Fuel Cells

4. Other Energy Solutions

   1. Solar & Storage

   2. Non-Scalable Renewables

   3. Small Modular Reactors & Datacenters in Space

5. Conclusion

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

Why We Need Bloom Energy

The Failure of the Grid

I won’t belabor the point here because this is already quite well known. The grid was built a century ago. It expects predictable slow-growing power demand of 1-2% a year. AI wants to double this capacity, and clearly the old infrastructure is not built for it. But the intuition here is important.

Power cannot be stored. This means that at any point in time, the total supply and the total demand of power must be perfectly equivalent. Thus, before a load (demand) is added, the grid operator must figure out how to add the equivalent supply so that the grid doesn’t break. Enter System Studies!

Imagine trying to model voltage stability, frequency response, contingency scenarios, and load flow under fault conditions for each new data center. When you do these models, you’re basically operating under the assumption of a fixed grid. But the grid is constantly changing! When grid topology is changing faster than studies can be completed, studies go stale.

This is why new interconnections take five years. The logistics of adding data centers to the grid is terrible. The system was never built for it. We need a new solution. We must bypass the grid and go behind the meter.

The Failure of the Turbine

The classical behind-the-meter solution is the gas turbine. It’s the obvious solution because it provides the base load, always on type power which data centers need. There are three main types of turbines.

Combined Cycle

Combined cycle is called combined cycle because it combines two cycles.

…

Cycle 1 converts natural gas into energy via normal turbining. Cycle two is converting the heat from cycle one into energy via steam.

The benefits of this are that it extracts the maximum amount of energy from the natural gas by not wasting the heat by-product.

The drawback is that it is very complicated and takes forever to build because you have to route more stuff around.

Therefore, this is usually the preferred gas turbine for supplying energy to the grid, but not for behind the meter. Again, big, complicated, but efficient.

Aeroderivative

An aeroderivative turbine is literally a jet engine strapped to the ground. They cost around $2,000 per kW.

This is the majority of the market. The most “normal” kind of behind-the-meter gas turbine. G.E. Vernova, Siemens, and friends make them.

However, there is also a very obvious problem. During the 1990s and early 2000s, the gas turbine oligopoly suffered a massive boom-bust cycle, almost like memory.

Therefore, the aero guys got PTSD (Post Traumatic Supply Disorder), and today, despite overwhelming demand, they refuse to build new factories and expand their capacity. This has been great for GEV stock.

But although investors love supply discipline, customers hate it because it means that they’re paying higher prices for a smaller source of supply.

It’s gotten so bad, in fact, that GEV is fully sold out through 2030. This tweet from SemiAnalysis sums it up well.

Now, customers are searching elsewhere, desperate to get their hands on anything that spins and produces electrons.

Reciprocating Engine

The “searching elsewhere” first lands at reciprocating engines, the cheaper, dirtier, and less technologically marvelous sibling of aeroderivatives.

These are literal car engines.

Although they cost less per kilowatt than aeroderivatives do, you pay for it with the insane amount of maintenance needed. They are smaller and generate less power per unit, so you need a boatload of them to power any reasonably sized data center (a 1 GW deployment requires 100-400). You basically need to set up a massive 24/7 auto repair shop on your data center campus.

Pollution

Climate change is bad!!!!

All forms of gas turbines are heavily polluting and terrible for the environment. They release NOx and SOx, which are harmful particulates that cause public health problems in the surrounding living areas. Therefore, every single deployment of gas turbines is subject to stringent permitting regulations that take an excruciatingly long time and potentially expose the developer to environmental lawsuits.

The Requirements of the New Solution

So, we need a new solution. We know that. But what characteristics should this new solution have?

Regulatory Landscape: Capital vs. Permission

Building a data center requires two things: capital and permission.

In 2024 and 2025, when AI was less developed and capable, it was capital that was scarce. Companies did not know if their CapEx would yield sufficient investment returns because all we had were chatbots. At the same time, permission was abundant because chatbots didn’t require that many data centers, and people were not afraid of losing their jobs.

Today, the situation has completely changed. AI agents are extremely capable and able to do most white-collar work. Anthropic hitting 30 billion ARR by the end of Q1 has signaled that the ROI question is effectively answered. AI compute demand is going parabolic, and now data centers are both a political and environmental issue. Today, capital is abundant. Permission is scarce.

What does this mean? We are transitioning from a regime where solutions required low capital and high permission are attractive to a regime where solutions requiring high capital and low permission are attractive. This forever changes the class of energy companies that benefit from AI capital expenditures. In order to be successful as a power equipment vendor in the future, you must help your customers get permission (FAST) and avoid regulatory stupidity.

Compute Shortage Necessitates Time to Power

Let’s think about the second-order effects of H100 prices (which were expected to decline into oblivion as technology obsoletes) rising 33% in 2026 in an unprecedented massive compute shortage.

Time to power was already important. AI cloud revenue is around $10 to $12 million per megawatt per year, meaning that having a 100MW deployment delayed by a month costs you $100 million.

Now it’s even worse. Failing to consider time to power means your (luckily secured) supply of compute can’t be used to address this shortage and collect the scarcity pricing.

Ten Features of Bloom Boxes

Bloom Energy makes fuel cells, a.k.a. Magic Energy Boxes. Natural gas goes in, magic happens, electricity comes out.

They are so cool.

Bloom’s Chief Commercial Officer tells us to think of them like a platform, like a phone where you can download many apps. With Bloom, you can utilize one of their many features any time you want, for any use case that suits you.

Yes, they cost $5k per kW in capex, which goes down to $3k after the investment tax credit, which is much more than aero derivatives or reciprocating engines. I will argue that the ten features we will discuss way, way, way more than offset this increased upfront cost.

Native 800V Direct Current

Electricity is transported long distances through the grid in AC because it is easy to step AC up or down between higher and lower voltages.

But the chips in the data center only ever use DC. Therefore, when electricity from the grid is used in data centers, it must be converted.

Today, electricity enters the data center as AC and gets converted near the rack to DC. This is usually not a problem because we use medium voltage (480V) AC and low voltage (54V) DC. The conversion process here is easier, requiring less expensive transformers and losing less energy as heat.

However, as the racks become more and more power hungry we must deliver more power to them. The fundamental power equation posits that P = I x V or power equals current times voltage. So we have two choices to increase power: increasing current or increasing voltage. The problem is current is basically the sheer volume of electrons flowing, so, to increase the current, you need to increase the width and mass of the copper cables. Now, as you can imagine, this is not scalable. To increase the power these days, we mostly rely on increasing the voltage, which is why we’ve arrived at 800V DC, the solution for the Rubin Kyber racks and all racks from then onwards.

Now, if you want to go from 480 V AC to 800 V DC, you have a problem. This conversion process is not easy; it is hard. It requires expensive transformers and loses tons of energy as heat. The ideal solution is not to intake AC into the data center; it’s to intake native DC. Well, guess what? Bloom Energy Fuel Cells produce 800 V DC natively. No gas turbine does this. Every gas turbine produces AC due to the loss of physics. Bloom will save you millions of dollars in transformer capex and energy conversion heat loss.

But you don’t need 800V DC today? No problem. Bloom boxes come with an adapter, just like your phone charger. You can take it off and switch to DC whenever you need.

Dynamic Load Following

There’s a common misconception with Bloom’s fuel cells, and it goes something like this: Gas turbines can start up in 15 minutes. Bloom’s fuel cells have to heat up for 12 hours to their 850°C internal temperature in order to create energy. Therefore, Bloom is fundamentally inflexible. It produces base-load power, but it cannot be turned off.

This is partially correct. Bloom’s fuel cells take a long time to turn on and off. However, when they are on, they can ramp from 20% to 100% output power, or anywhere in between, in seconds. This gives them a very useful feature known as dynamic load following.

Why is this useful? Training workloads have variable loads. Normally, you’re forced to build massive energy storage systems to smooth out these variable energy loads. With Bloom, all you have to do is ramp up and down.

Supercapacitors

AI workloads feature an annoying nuisance known as sub-second transients. Thousands of GPUs can simultaneously finish a compute step and wait for gradient exchange. When that happens, power draw can drop by tens of megawatts in milliseconds, then spike back up just as fast when the next compute step starts.

This causes a major problem. Electrical systems require generation and load to match in real time, or they crash. AI training workloads break this constraint violently: a 1GW cluster can swing 500MW in milliseconds, thousands of times per training run, faster than any combustion-based generator can respond. Without something to absorb these shocks, gas turbines serving AI workloads would trip offline continuously and the datacenter would never run. The standard fix is a $500M battery bank bolted onto every gigawatt of turbine deployment, sized solely to compensate for what the turbines cannot do natively.

Bloom, on the other hand, uses built in supercapacitors.

What are supercapacitors? Lithium ion batteries stores energy chemically. Supercapacitors store energy like static electricity.

This means you erase the entire battery bank. As you might imagine this comes with all sorts of capital efficiency and reliability benefits!

Modularity

Bloom boxes are modular. You can stack these things on each other like Lego bricks. Need 100 MW? Need 5 GW? No problem! Just scale up or down as needed.

This is very important because of a concept called over-provisioning. For any data center workload, you need to provide more generation capacity than is needed for reliability. And the goal is what’s known as five nines reliability, meaning the system is up 99.999% of the time.

For traditional gas turbines, because of the lack of modularity, you need to over-provision by 20 to 40% in order to achieve 5 nines reliability. For Bloom, you only need to over-provision from 0 to 8%.

Absorption Chilling

Bloom’s Fuel Cells produce a very interesting byproduct: very hot, high-quality steam. Emphasis on high quality. This is something that the management team loves to say.

What can you do with said high-quality steam? Well, data centers require cooling. Cooling actually requires energy. Normally, this comes out of your energy budget, but with said high-quality steam and a little bit of magic from PV=nRT (remember from college chemistry), you can actually use pressure to turn that steam into a liquid, then boil it again with energy and have that become a coolant.

Carbon Capture

Climate change is bad!!!

As always, Bloom is here to save the day.

Traditional gas turbines produce very unclean waste: CO2, NOx, SOx, and other shitty particulates. Bloom, on the other hand, produces almost purely CO2 and water vapor, which is crazy pure! What this means is that the economics of carbon capture becomes so much more favorable. In the future, there will most certainly be a climate-focused regulatory regime, and hyperscale data centers will be required to achieve net zero. With cheap carbon capture added on Bloom boxes, Bloom’s fuel cells suddenly become far more economical than any traditional gas turbine solution

In addition, there is something far more underrated here: the clean energy market. Remember, there is a market for purely clean energy: solar, wind, hydro. Data centers aren’t the only things using energy out there. Not everyone needs 5-9’s reliability. If Bloom’s Fuel Cells can become zero-carbon with just a bit more capex, it becomes a direct competitor to all forms of clean energy. As Bloom’s Chief Commercial Officer put it: “You go do your calculations of how big that market is.”

Finally, some people are big hydrogen fans and think hydrogen is the zero carbon future. Well, it turns out that with Bloom, their fuel cells deployed today can literally support hydrogen if you wanted it to; it’s fuel agnostic. The electrochemical reaction works the same for hydrogen as it does for CH4, because essentially it uses the hydrogens in CH4 anyways, so it’s kind of the same thing.

Quick Deployment

Getting behind-the-meter power takes a long time for three main reasons:

1. Manufacturing the power-producing widget at the factory. Manufacturing things takes a long time. This is pretty intuitive and simple.

2. Putting the power-producing widget onto the actual site itself. These things can be really heavy, therefore there’s probably a lot of construction required. Also, you probably have to order a bunch of electrical switchgear and transformers and stuff that go alongside it.

3. Permitting and regulatory clearance. If the regulators are up breathing down your neck, you can’t really do stuff. This slows things down.

Bloom solves all three problems:

1. Bloom is relatively capital light, and manufacturing Bloom boxes is more akin to assembly than literally bending metal.

2. Bloom’s fuel box fuel cells are quite a lot lighter than gas turbines, and therefore can just be plopped down on the ground without as much of a solid foundation.

3. As we have discussed, they are very clean; it is easy for them to get permitted.

Normally, stuff takes 18 months to go from being purchased to being deployed. Bloom claims they take 90 days, but in a recent deployment with Oracle, they did it in 55. Hooray!

Low Latency Inference

Now I’m not that big of a believer in low latency inference, but I still have to include it here because it’s important. Low latency inference must be placed close to population centers. Population centers generally aren’t a fan of the knocks and socks produced by big polluting gas turbines. In addition, they’re clunky and not modular like Bloom. Bloom is clean and modular; you can stack them on top of each other to save space, and they don’t pollute very well for putting it near people. You won’t get a lot of NIMBYism for this one.

High Efficiency

A gas turbine turns natural gas into electricity by first converting the gas into heat, the heat into mechanical energy, which then finally turns into electricity. Fuel cells, on the other hand, don’t really care about going the scenic route. They convert natural gas directly to electricity via chemical magic.

Therefore, Bloom boxes have an efficiency of around 60 to 66%, which is as good as the combined cycle gas turbines we talked about earlier that are big and complex. They are way better than reciprocating or aero derivatives, which are stuck between 35 and 45%.

Data Analytics

K.R., the CEO, says:

“We have a few trillion cell hours of field operation. Is what Bloom has. Few trillion cell hours. More than 6 billion data points come from our field. To us every single day. We are using AI. We are not only benefiting from AI on our revenue side. We are using AI to our benefit. For all this. To improve our performance every single day. Because we have a digital twin associated to every single fuel cell stack and data from the real field is coming and feeding the digital twin and making our models better and better.”

A professor once told me a story about Tesla. They were trading in the teens. This was in 2019 or 2018, before the huge run-up. He never invested in Tesla because he thought they were speculative, but his son told him that Tesla wasn’t really a car company. They were a data company, and that’s when he realized and invested. Turns out his son was right. Every single hour an electric vehicle was driving in the streets, Tesla was building the world’s most powerful and hidden data moat.

This is not different. Bloom Energy is the only player that has a fuel cell installed base. A digital twin means that there is a computer simulation of every single fuel cell that Bloom has deployed. They are able to track when any fuel cell fails in any customer deployment and fix it before it causes problems, drastically reducing the expense associated with service and increasing the revenue from that service. With operational optimization of fuel cells, based on simulations run on data collected, they can improve the stack life of their current generation. Through fleet learning and manufacturing feedback from their field telemetry, they are able to speed up manufacturing process optimization and make future fuel cell generations even better.

The Unconstrained TAM

Ok, ok, Bloom boxes are better. We get that, but how does that translate into revenues? Well, here I want to argue that Bloom’s TAM is essentially unconstrained.

The most basic premise goes like this: most high-growth AI infrastructure companies are high growth because their market is growing fast.

…

But for Bloom, it’s different. Bloom is growing fast because their technology is actively displacing an already massive market: the energy generation market. Each time the capabilities of the technology improve, new use cases are unlocked, new marginal customers switch over, and shipments grow.

Therefore, the growth is irrelevant. It doesn’t need to grow to be big. It’s already massive. This means that, theoretically, the TAM is nearly unconstrained. The stock price doesn’t really have a ceiling!

Now, on to the actual ability to capture this TAM. This requires high scalability, constantly improving technology, and dominant market share, all of which Bloom has.

Infinite Scalability

The most unintuitive thing about Bloom is just how gosh darn scalable these darn boxes are.

It really isn’t a manufacturing business. It’s an assembly business. Think of it like semi-caps. Semi-caps have fabs, yes, but those fabs don’t work with raw materials like semiconductor fabs do. They have a chain of suppliers that they get parts from, and then they just put it together. They’re an assembly house. Their moat is their IP, which means that they are just as scalable as fabless companies with capex in the range of 3% of revenue. Bloom is the exact same, building out gigawatts upon gigawatts of capacity, measured in the tens to hundreds of millions, and generating billions of revenue

Fuel Cells are a Technology

Gas turbines are an industrial product. Centuries old, already lived out the S-Curve, and very slow to change. Fuel cells are a technology. Their costs are reduced by roughly 10% to 20% each year. They literally have their own Moore’s Law-like chips, and it isn’t ending anytime soon. They’re at the bottom of their S-Curve and have yet to be adopted. When you think of fuel cells as a technology, you stop framing it as a static thing and start seeing it as a constantly improving one. You realize that for adoption it’s not a matter of if, it’s a matter of when.

If this continues, you’ll have to ask: why is Bloom limited to data centers? If they become cheaper and more elegant of a solution for producing power than even combined cycle power plants, which are used by utilities, why won’t they be used for powering the world? Why won’t utilities buy from them?

Bloom is a Monopoly on Fuel Cells

Nobody calls Bloom a monopoly today, but one day they will. Today, Bloom is seen as one entrant offering a niche solution in a crowded market. If that niche solution becomes the most popular one, fuel cells themselves become a market. When fuel cells become a market, people will ask what the market share is. Surprise, surprise, Bloom Energy has 100% market share in fuel cells, 100%! Not a single other vendor is deployed at volume.

Bloom has been making fuel cells since 2008. They have a 20-year technology head start on everyone else. 20 years of accumulated IP and manufacturing know-how. Legacy turbine manufacturers may try to pivot, but Bloom themselves aren’t scared. Here is a quote from management telling you exactly how they feel about competition. From Aman Joshi, Bloom’s Chief Commercial Officer, on the EnergySense podcast in February 2026:

“I think it’s natural as Bloom succeeds that there will be other companies who’ll try to get into the fuel cell business. Both myself and our chairman love competition because it only inspires us to become even better is how I would characterize it. Now, fuel cells have been around. The technology has been around for fifty years, but a lot of companies, you mentioned General Electric, in early two thousands had a fuel cell division. Siemens had a similar division, and so did other OEMs. They’ve all tried, invested hundreds of millions of dollars, but weren’t able to make it work and go beyond lab scale experiments to commercial grade. And that’s where Bloom’s got roughly a twenty year technology lead. So we wish everyone all the best.”

Other Energy Solutions

Let’s address some other non-gas energy solutions. The way I will frame this section is that there is a reason these aren’t discussed very often. There is a reason hyperscalers default to natural gas when going behind the meter, but we must still address the hype.

Solar & Storage

Solar and storage seems like a pretty solid alternative. The reason it feels very realistic is because it’s very scalable. Unlike wind or hydro, the amount of power generated isn’t constrained by the earth’s resources. The more solar panels you build, the more power you generate.

The main problem we run into is probably something you already know: data centers need power 24 hours a day, but storage solves the issue, right? Well, only partially. Utility-scale solar runs at 20-25% capacity factor, so 1GW of continuous baseload requires 4-5GW of nameplate capacity plus 12-16 hours of storage to bridge nighttime and 24-48 hours to bridge weather events. Remember, gas only needed around 25% overprovisioning, and Bloom’s fuel cells barely needed any?

The biggest problem actually is that solar still needs the grid for backup. We already addressed the whole reason we’re going to behind-the-meter gas in the first place is to get rid of the grid. Solar forces us to go backwards. This is why it is rarely discussed.

Non-Scalable Renewables

This part is simple. Hydro and geothermal are limited by the Earth’s resources. There’s only so much hydro in the world because there are only so many waterways in the world, and a lot of it is already being used up because hydro is a huge portion of energy generation already. Geothermal only exists in certain narrow bands of the Earth’s crust, so unless you’re going to limit your data center build out to Northern California, you’ll need to look for something else.

Wind is almost funny in how unviable it is. You have the same intermittency problem that you have for solar, except for wind it’s even worse because you need more land. Plus, wind’s intermittency is unpredictable, unlike the sun, which rises and sets at the same time every day.

SMRs & Datacenters in Space

Maybe one day we’ll get a breakthrough on these technologies that make them viable at mass scale. Today, the timelines, especially for SMRs, are completely misaligned with the window in which AI needs to be built out. 2026 to 2035 is the most key period for AI, not 2035 and beyond.

Conclusion

Despite a slightly higher upfront capex cost, Bloom’s Fuel Cells have ten advantages that make their net cost much, much lower than turbines. Bloom is set to take share in a massive, already established market by displacing legacy generation solutions, making its TAM effectively unconstrained. Bloom has an incredible ability to capture this TAM via high scalability, consistently improving technology, and a monopoly on fuel cells themselves. With no other viable alternatives appearing until the mid-2030s, I conclude that Bloom Energy is the best AI energy solution on the planet, and our transition to it will be a good one for the world at large.

Until now! These companies are in fact one of the most fascinating plays in AI infra.

There are layers to the understanding of the Bitcoin miner thesis. At the surface level, they are Bitcoin miners that have pivoted to AI infrastructure. Each time we go one level deeper, we find out something that changes the entire framing of these businesses. We flip from bull to bear and back to bull. It is truly an iceberg. Most analysis in this space only glances at the surface level, measures them by the wrong yardstick, and misses the moats and the threats that truly matter.

Today, we descend down the iceberg. We take you on a journey through each layer of the thesis, peeling back your understanding until we get to the core of what makes these Bitcoin miners important to AI.

Everyone is obsessed with physical AI bottlenecks. Meanwhile, these bitcoin miners may just hold the final choke point on the AI build out, and it has nothing to do with the physical world.

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

Contents

1. They Mine Bitcoin

2. They Are Pivoting to AI Co-Location

3. Grid Power is a Scarce Asset

4. Bloom Enters the Chat: Power is a Commodity

5. The Scarce Asset is Actually Permitting

6. D E N S I F I C A T I O N

   1. The Chip

   2. The Package

   3. The Rack

   4. The Floor

   5. The Site

   6. The Result

7. My Current Favorite Miner

They Mine Bitcoin

Bitcoin is the world’s first scam cryptocurrency (jk I am actually a crypto fan). A digital token where ownership is recorded on a decentralized ledger, spread out and validated across millions of different computers called nodes. Essentially, what it produces is a scarce asset that can be transmitted via the internet. Think of it like gold with teleportation abilities.

In order to record these transactions, miners run big datacenters full of ASICs that have literally one job: to take the transaction information in the block and hash it using the SHA-256 algorithm until the hash has a pre-determined number of trailing zeros. It is basically a lottery ticket, and whoever gets that hash first records the block and posts the transactions to earn the Bitcoin block reward.

Every four years, the Bitcoin block reward is cut in half. This is to ensure that the network only ever emits 21 million coins to preserve the scarcity of Bitcoin as an asset.

So from a miner’s perspective, this is a TERRIBLE business. Your service is fully a commodity and your pay gets cut in an exponential decay curve. The miners that survived to the present day have become experts at finding the absolute lowest cost of stranded power anywhere in the US and building data centers as cheaply as possible.

They Are Pivoting to AI Co-Location

This is the level of understanding held by most retail investors in the general public. The Bitcoin miners pivoting to AI co-location narrative took hold in mid-2025.

The intuition is intuitive. If Bitcoin miners became an expert at securing low-cost grid power and building data centers for cheap, it makes perfect sense that their skills and assets are a fit for AI.

The business model here is co-location. Hyperscalers or neoclouds bring the clusters themselves (GPUs, networking) and miners provide the grid interconnects, permitting, and data center shell. The cloud signs a lease with the miner for these services over the long term. The miner serves as the landlord, and the cloud is the tenant.

This business has a few major characteristics that are worth noting.

First is the high capex relative to revenues. Each megawatt of CITL (Critical IT Load) takes around $8 to $13 million to fully build out. Then they can earn anywhere from $1.5 to $2.5 million a year in high-margin recurring rental revenue. But because the AI co-location business is just getting started, all miners are cash flow negative as sites are all still being built out.

The second is that this business is weirdly safe compared to the other AI plays out there. Again, the intuition is intuitive. The end state of these co-locators is basically a REIT. Recurring high margin rental revenue. Semiconductors are cyclical and can be affected if overall cloud CapEx declines. Neoclouds face a fast depreciation curve and are criticized for their relatively lower margins. But miners hold long-lasting assets and face zero cyclicality.

Grid Power is a Scarce Asset

Now we have reached the level of understanding of the AI infrastructure investing space broadly.

It is quite well known that grid interconnects are incredibly hard to get these days.

SemiAnalysis

How AI Labs Are Solving the Power Crisis: The Onsite Gas Deep Dive

The Grid is Old and Tired…

Read more

4 months ago · 233 likes · 31 comments · Ajey Pandey, Jeremie Eliahou Ontiveros, and Dylan Patel

New load approvals can take five years. PJM and ERCOT get tens of gigawatts of new load requests per month and approve nearly none of them. The grid infrastructure simply isn’t built to grow quickly. It was built a century ago.

Miners, on the other hand, secured grid power during the crypto boom in 2021 and 2022, completely front-running this AI power land grab. As a result, they are some of the fastest ways to get grid power if you are a hyperscaler or Neocloud.

At this level, we understand that the grid interconnections and power, more generally, is a critical bottleneck for AI. These miners hold that asset and therefore should get re-rated as they realize that value.

Bloom Enters the Chat: Power is a Commodity

Bloom Energy is the side character cameo today.

We have previously just established that grid interconnects are scarce and power is a critical bottleneck. Well, it turns out there’s another company solving this exact problem: Bloom Energy.

I have written a very comprehensive long thesis on Bloom that covers all of these points, but here I will summarize the important attributes that make them relevant to our discussion about miners.

Bloom Energy makes fuel cells that can be manufactured in an incredibly scalable way (they have explicitly said they will never be supply-constrained on their earnings calls), can go from purchase to power in under 90 days (unlike turbines which take multiple years or the grid which takes 5+), and are now cost-competitive with the grid in “most areas” and turbines “everywhere.”

They have transition from being an alternative solution to THE solution.

Recently, Oracle replaced over 2 GW of planned turbines in their New Mexico campus for Bloom’s fuel cells.

Their rise in relevance and the resulting share price rally has been stunning. This is great for Bloom, but what does it mean for the miners?

Well, if Bloom provides fast, scalable, clean behind-the-meter power that is cost competitive with the grid, then is power really an asset? One can argue it is not. Holding X gigawatt of grid interconnect does not make a difference anymore if those same gigawatts can be accessed via fuel cell.

Today, we measure Bitcoin miners through the yardstick of megawatts secured. If those megawatts are easily commoditized, they are no longer the source of pricing power, and we have to measure miners on something else. What else? Do they hold land, construction, access to fiber? All these things are not as scarce as grid connections. Hyperscalers can simply do it themselves, right?

That right there is the bear case.

The Scarce Asset is Actually Permitting

Backlash towards AI is growing. This is a technology that’s going to take your jobs AND destroy your local environment. What’s not to hate?

A growing political consensus against AI is now the base case of most analysts. We understand it well, so let’s play it forward. How does the AI build out become more difficult when backlash becomes more widespread?

It would likely be through permitting.

Permitting is the process of getting regulatory approval to build data centers. To begin a project, a data center developer needs multiple overlapping approvals across multiple jurisdictions. These include (according to Claude):

• Local land use permits and zoning approvals, where county commissions decide whether a 500 MW industrial facility belongs on a given parcel

• Building permits across structural, electrical, mechanical, and plumbing categories, each independently reviewable, each with its own timeline

• Stormwater management plans and traffic impact studies negotiated with local public works departments

• State-level air quality permits if the site has any onsite combustion or generation

• Water use permits, increasingly contested as data center cooling demands collide with agricultural and residential users

• Environmental impact assessments under state law, brutal in California and Virginia, lighter in Texas and Oklahoma

• Federal NEPA review where any federal land, funding, or permitting nexus exists

• Army Corps of Engineers wetlands permits where applicable

• FAA review for structures exceeding certain heights or near airports

• Endangered species reviews under federal and state law

Each one of these has its own timeline, own constituency, and own way of being mobilized against a project. There are calls for data center moratoria, but in most places it is likely to simply evolve into these permitting regulations becoming much stricter to serve as an effective moratorium by making it so hard to develop data centers it dissuades construction.

Getting these permits takes serious time. It could take anywhere from 18 to 60 months depending on how friendly your jurisdiction is. And this is PRE-construction. I sincerely believe this is going to become a real, substantial choke point. The funny thing is that no one is talking about it today, but I think people will start caring not too far from now.

Just like with the grid, these miners got permitted during the crypto boom long before it was a major problem. Between 2018 and 2022 was the prime time for this, where crypto mining was important and profitable enough that it made permitting these sites economical, but AI wasn’t a thing yet, so there wasn’t much competing demand nor any political scrutiny on datacenters. In fact, many localities welcomed these loads!

Now, to be clear, a permitted crypto site can’t just be converted into a permitted AI site with the push of a button, but it definitely makes it much easier. The project has gotten to the point where the marginal regulatory cost of changing or expanding the workload is much lower than it would be to build a greenfield AI site.

Miners, therefore, are one of the only available bets on regulatory backlash against AI as the asset they hold is not grid-connected megawatts, but permitted sites. It is precisely these permitted sites that become more scarce and more valuable the worse the regulatory environment gets.

D E N S I F I C A T I O N

We have arrived at the final stage. Welcome to Densification.

We have established that miners hold a critical choke point asset in permitted sites, but what happens when we play the AI build out forward? Do permitted sites become more or less valuable over time? Do they become more or less scarce compared to the other bottlenecks?

The Densification thesis argues that a permitted site becomes far more valuable over time as the amount of compute and power that can be served with one site exponentially increases through densification. This means that scaling AI won’t require building out new sites (which becomes much harder as regulatory scrutiny grows) and can instead be done through making existing sites denser. We must cram more compute into each square inch at four levels: the chip, the package, the rack, the floor, and the site.

The Chip

To cram more compute into each square inch, we first start with the chip. This is pretty easy because it is just the continuation of Moore’s law. Specifically, the most important feature here is backside power.

In the past, all the copper wiring for the chip was on the front side of the chip. This means that both the electrical signal and the power delivery had to go through the same front side. Since the power delivery has to share wiring with the signal delivery whose wiring has to be very thin, the power that reaches the chip is very degraded and the efficiency is low.

Backside power changes that. By routing straight through the backside in a dedicated power network, power delivery to the chip improves drastically and the energy efficiency goes to moon. This means we can get more compute from the same amount of power delivered to the rack.

The Package

Advanced Packaging is pretty well known and well covered at this point. If you need a primer on it, here is my short n’ casual intro.

What’s important to the Densification thesis is the roadmap beyond CoWoS. Intel’s EMIB, TSMC’s CoPoS, and glass substrates all allow the package to be expanded even further and pack more compute together. Furthermore, hybrid bonding and 3D packaging are coming. Hybrid bonding gives you orders of magnitude more connections per square millimeter and therefore massively increases the compute and memory bandwidth per package.

The Rack

Two major trends allow the rack to be densified.

The first is the increasing power draw of the rack, exemplified by NVIDIA’s Kyber Rack for Rubin. We were at the low hundreds of megawatts per rack, but now are quickly moving towards one megawatt racks. When you have each rack drawing much more compute than before while occupying the same refrigerator-sized footprint, your flops get much more concentrated.

The second is the scale-up domain. Blackwell Ultra introduced the NVL72, 72 chips NVLinked together to function as essentially one. Rubin moves this to NVL144 and NVL576, and Feynman will increase it even further.

The Floor

The floor densification is similar.

First, the Scale-up domain can go beyond one rack by using optical connections. NVL576 is actually eight Oberon racks linked together with optics. And it’s only going to get even crazier from there. As you expand the scale-up domain beyond a rack with optical, your effective compute per square foot increases even further.

Then we have the all-important 800 VDC. The fundamental power equation says that power is equal to voltage times current. To deliver more power, you need to increase 1 of the 2. Instead of increasing the current like we used to, with 800 VDC, we massively increase the voltage instead. This is great because increasing the current (which can be thought of as the total number of electrons flowing) increases the amount of copper needed, so, by going the voltage path instead, we can get more power delivered with less copper. This means we can pack the racks closer together.

The Site

Finally, we move up to the site level. How can this be densified?

Remember Bloom, right? One of Bloom’s most attractive attributes is its onsite scalability. You can stack fuel cells like Lego bricks. If you want 100 MW, 200 MW, 500 MW, a GW, simply stack more fuel cells.

While the power delivery on a given site used to be capped by the interconnect agreements, today, with Bloom, you can simply scale up the power on that site. Fuel cells are clean and easy to get permitted. This means that the power budget of a site is essentially uncapped. As the density of the floor and the rack increases and requires more power draw, more fuel cells can be added to match the requirements.

The Result

How much can we densify by 2030?

Let’s say advanced packaging and backside power can deliver five times compute per watt, high-density racks can multiply the watts per rack by another three to seven times, the optical scale up and 800 VDC can increase density of the floor by 2 to 3.5 times, and megawatts per site can go up another three to five times through leveraging Bloom Energy.

Overall compute density can increase by 100-1000x.

That’s two to three orders of magnitude more economically valuable tokens delivered to the end customer. Think about what that does to the pricing power of the holder of the permitted sites that run this compute.

My Current Favorite Miner

There is one miner that has recently made an announcement showing they are probably the first to realize this strategy, particularly on scaling up site power with Bloom (literally by 5x!). Multiplying site power capacity directly multiplies revenue for miners because they bill via dollars per megawatt. And somehow the market seems to have missed this completely. They also had a unique history which makes them quite favorable for playing out this thesis.

Read more

Aixtron is a revisions machine. I really love their management team. They give us so much color on these calls and makes things really easy for us analysts.

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

Contents

1. Earnings Presentation and Prepared Remarks

2. Interesting Call Comments

3. Opto Visibility and Revisions (!!!)

   1. The Catalyst

   2. Insane Tool Shipment Forecasts

   3. Installed Base is No Problem

   4. Epitaxial Tool Intensity

4. Capacity Math

5. Margins, Competition, Operating Leverage

Earnings Presentation and Prepared Remarks

This was basically the preliminary revenue release with a few extra tidbits attached. I won’t go into detail on the specific financials because most of it is covered in the preliminary earnings release, which I also did a review of.

The first piece of new information with specific revenue breakdown by tool and market. Opto is 51%, which is pretty expected, but what really surprised me was that the share of LED and GaN SiC power was only 17% of revenues. GaN was only 10% of order intake. This just shows you how crazy weak the power market is right now. We’re seeing a massive rally across all the power electronics names like ON Semi and Navitas because of 800 VDC, and eventually that’s going to flow to Aixtron, but they’re not talking about it and no one else is talking about it. This is definitely something to keep in mind.

The reason they’re not talking about power is because they only get orders when one of their customers needs serious levels of production. Their customers only need serious production when they have a major design win with a system integrator like Nvidia. Since the design wins haven’t happened yet, the demand isn’t there and the orders haven’t come through. All the 800 VDC talk is still theoretical.

Interesting Call Comments

First is the Malaysia site. They’re building a new greenfield fab in Malaysia, and there has been lots of speculation on why they’re doing this. Very interestingly, they said this has nothing to do with opto demand. In fact, they are not capacity constrained at all. I’ll get their actual capacity numbers later, but the only reason that they’re building a Malaysia fab is to serve their Asian customers for cheaper. That’s it.

The other thing that had a lot of speculation was on their convertible bonds. They issued 450 million euros of converts, and the way they described it on the call was low-key a little concerning. They just said it was too good of an opportunity we couldn’t pass it up, which sort of implies that they think their stock is overvalued. At the same time, when an analyst asked if it was going to be used for buybacks, they said maybe in the future but not at this share price level. Like, straight up. Kind of sad that they don’t believe in their own company as much as some of us do, but as you’ll see later on, they contradict their own bearishness with the most insanely bullish Opto forecasts. It makes for a very funny contrast.

I think it’s just the conservative Germans in them. Should be noise (but I am heavily biased due to my massive position). LOOOL

Opto Visibility and Revisions

This was the first call they really spelled out the Opto bull case, kind of the equivalent of Lumentum’s call in November 2025. Before that, it was all “demand is strong, our customers need capacity, but never ‘this is an inflection point and a regime shift.’ Today, that changed. For the first time, they said that growth continues into 2027 and beyond. The first time they mentioned and separated scale up, scale out, and scale across. They’re finally repeating what Lumentum has been telling us for half a year.

They turned into believers! This was because of a very specific and obvious catalyst that caused them to be unable to ignore the signals of explosive growth any longer. As a result, they materially increased their tool shipment forecasts. WAY higher than any of us modeled. Although they didn’t change their full-year guidance because it was already dropped in the preliminary release, I fully expect them to do so on their next earnings call.

The Catalyst

Read more

Have you ever seen a company beat consensus REVENUE estimates by 36%? They absolutely mogged these analysts lmao.

And this probably meant that they shipped between 50 and 100% more Bloom boxes than people thought, because their revenue is made up of product, service, and several other line items. A lot of the other stuff is based off of installed base. The growth measure is really product revenue, and this beat was completely driven by product revenue. So in that sense, it’s actually even bigger than the headline number suggests.

For the guide, they increased their revenue midpoint by 12.5% and their operating income midpoint by… 125%. Their operating income guide is now 50% above street consensus.

The way they’re growing quarter over quarter, I honestly severely doubt that they won’t beat their midpoint by a crazy amount. I mean, we’re already at 750 million this quarter, and the guide implies 900 million per quarter run rate. Given that they just TRIPLED product revenue year over year and are easily growing double digits sequentially. I think they’re easily above $1 billion per quarter in the second half and smash right through that guide.

Unsurprisingly, we are up 12%.

This is really incredible stuff from Bloom. Now let’s see what they drop on the call.

Read more

I added this company to my portfolio at the end of March.

Since then, they have produced great returns, but I think they are bound to go higher and without much of a ceiling on the valuation (as tail scenarios can launch it hard).

Their management team talks like a startup. In the sense that they are so cocky and comically overconfident in their own success you can’t help but also be confident with them.

It shows up in their earnings calls, the podcasts that the management team goes on, their slide presentations at technical conferences, and basically any form of communication. They use weird and funny analogies (just like I do!). It’s honestly pretty entertaining.

I don’t know exactly when, but one earnings call within the next 18 months or so will be the equivalent of Nvidia’s May 2023 blowout: numbers so far past consensus that the old company dies and a new one is born.

This is a loooong one. Get strapped in. Let’s go!

The Failure of the Grid

I won’t belabor the point here because this is already quite well known. The grid was built a century ago. It expects predictable slow-growing power demand of 1-2% a year. AI wants to double this capacity, and clearly the old infrastructure is not built for it. But the intuition here is important.

Power cannot be stored. This means that at any point in time, the total supply and the total demand of power must be perfectly equivalent. Thus, before a load (demand) is added, the grid operator must figure out how to add the equivalent supply so that the grid doesn’t break. Enter System Studies!

Imagine trying to model voltage stability, frequency response, contingency scenarios, and load flow under fault conditions for each new data center. When you do these models, you’re basically operating under the assumption of a fixed grid. But the grid is constantly changing! When grid topology is changing faster than studies can be completed, studies go stale.

This is why new interconnections take five years. The logistics of adding data centers to the grid is terrible. The system was never built for it. We need a new solution. We must bypass the grid and go behind the meter.

The Failure of the Turbine

The classical behind-the-meter solution is the gas turbine. It’s the obvious solution because it provides the base load, always on type power which data centers need. There are three main types of turbines.

Combined Cycle

Combined cycle is called combined cycle because it combines two cycles.

…

Cycle 1 converts natural gas into energy via normal turbining. Cycle two is converting the heat from cycle one into energy via steam.

The benefits of this are that it extracts the maximum amount of energy from the natural gas by not wasting the heat by-product.

The drawback is that it is very complicated and takes forever to build because you have to route more stuff around.

Therefore, this is usually the preferred gas turbine for supplying energy to the grid, but not for behind the meter. Again, big, complicated, but efficient.

Aeroderivative

An aeroderivative turbine is literally a jet engine strapped to the ground. They cost around $2,000 per kW.

This is the majority of the market. The most “normal” kind of behind-the-meter gas turbine. G.E. Vernova, Siemens, and friends make them.

However, there is also a very obvious problem. During the 1990s and early 2000s, the gas turbine oligopoly suffered a massive boom-bust cycle, almost like memory.

Therefore, the aero guys got PTSD (Post Traumatic Supply Disorder), and today, despite overwhelming demand, they refuse to build new factories and expand their capacity. This has been great for GEV stock.

But although investors love supply discipline, customers hate it because it means that they’re paying higher prices for a smaller source of supply.

It’s gotten so bad, in fact, that GEV is fully sold out through 2030. This tweet from SemiAnalysis sums it up well.

Now, customers are searching elsewhere, desperate to get their hands on anything that spins and produces electrons.

Reciprocating Engine

The “searching elsewhere” first lands at reciprocating engines, the cheaper, dirtier, and less technologically marvelous sibling of aeroderivatives.

These are literal car engines.

Although they cost less per kilowatt than aeroderivatives do, you pay for it with the insane amount of maintenance needed. They are smaller and generate less power per unit, so you need a boatload of them to power any reasonably sized data center (a 1 GW deployment requires 100-400). You basically need to set up a massive 24/7 auto repair shop on your data center campus.

Pollution

Climate change is bad!!!!

All forms of gas turbines are heavily polluting and terrible for the environment. They release NOx and SOx, which are harmful particulates that cause public health problems in the surrounding living areas. Therefore, every single deployment of gas turbines is subject to stringent permitting regulations that take an excruciatingly long time and potentially expose the developer to environmental lawsuits.

The Requirements of the New Solution

So, we need a new solution. We know that. But what characteristics should this new solution have?

Regulatory Landscape: Capital vs. Permission

Building a data center requires two things: capital and permission.

In 2024 and 2025, when AI was less developed and capable, it was capital that was scarce. Companies did not know if their CapEx would yield sufficient investment returns because all we had were chatbots. At the same time, permission was abundant because chatbots didn’t require that many data centers, and people were not afraid of losing their jobs.

Today, the situation has completely changed. AI agents are extremely capable and able to do most white-collar work. Anthropic hitting 30 billion ARR by the end of Q1 has signaled that the ROI question is effectively answered. AI compute demand is going parabolic, and now data centers are both a political and environmental issue. Today, capital is abundant. Permission is scarce.

What does this mean? We are transitioning from a regime where solutions required low capital and high permission are attractive to a regime where solutions requiring high capital and low permission are attractive. This forever changes the class of energy companies that benefit from AI capital expenditures. In order to be successful as a power equipment vendor in the future, you must help your customers get permission (FAST) and avoid regulatory stupidity.

Compute Shortage Necessitates Time to Power

Let’s think about the second-order effects of H100 prices (which were expected to decline into oblivion as technology obsoletes) rising 33% in 2026 in an unprecedented massive compute shortage.

Time to power was already important. AI cloud revenue is around $10 to $12 million per megawatt per year, meaning that having a 100MW deployment delayed by a month costs you $100 million.

Now it’s even worse. Failing to consider time to power means your (luckily secured) supply of compute can’t be used to address this shortage and collect the scarcity pricing.

We Will Transition and Never Return

I believe the adoption of the solution I discuss below would not be a temporary stopgap, but a secular transition.

After adoption, this is a technology that shall compound upon itself, becoming better year after year, never giving the legacy players a chance to get back in.

It has over 10 distinct advantages that will become more important over time as the requirements of AI workloads are changing in its favor.

The TAM ceiling is so large it’s irrelevant. The company has a monopoly moat. And there is a surprising lack of alternatives.

This is an extra extra long one. We have so much to talk about.

I also wrote all this by hand (found that it is easier to read and more entertaining), which is a strange thing to brag about, given that I am trying to integrate AI into every part of my life.

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

Read more

Revenue and guide are INCREDIBLE. Beating by 10%. Clearly tons of CPU price hikes happening already. As I discussed in my timely CPU article (which I will now be heavily victory lapping) we are in the very very early stages of the CPU S-curve as they were never needed before agentic AI. A few months of Claude Code and OpenClaw did this!

GM beat their prior guide by 650bps. This is pretty unholy.

I genuinely do not know why their GMs guided down without the call.

(Note from Jason after the call: Now I know. Its actually good news)

CPU price hikes are everywhere, both consumer AND AI. They are enough to offset whatever unit loss is happening due to this and memory murdering the BOM, apparently.

Here is sort of the recurring theme of Intel: Operating leverage. A big but not unheard-of beat on revenue (10%) turns into operating income that is double what was projected. Intel’s chart looks scary but remember, they used to be unprofitable and trading at 1-2x revenue!

A graph from the early days of my Substack:

This is why I like low revenue multiples and low profitability for asymmetric longs. One day they are unprofitable and the next earnings appear from thin air and start compounding at triple digits and everyone acts surprised.

Price Action

Intel surges 20% after hours and breaks its all time high, a record last set in 2000 during the dot-com bubble.

This is the semiconductor industry’s greatest comeback story.

Lip-Bu dropped the coldest quote in Intel history:

“A year ago, the conversation about Intel Corporation was about whether we could survive. Today, it is about how quickly we can add manufacturing capacity and scale our supply to meet enormous demand for our products.”

- Lip-Bu Tan, CEO, Intel Corporation

The Call

Going forward, my earnings call analysis will be reserved for the paid tier only.

Intel’s earnings call today was the most confident I have ever heard the management team in a long time. We will explore the development of 8 different themes ranging from the surprising positive developments in 18A yields to management’s newfound bullishness on CPUs for agentic AI. Finally, I will share my Intel model for the first time, including a comprehensive revenue build (splitting CCG, DCAI, external 18A-P revenue, 14A revenue, advanced packaging, and ASIC), income statement/EPS, and DCF. We will explore the path to a one trillion market cap.

Contents

1. CPUs and Agentic AI

2. Culture

3. 18A Yields

4. 14A Progress and Capex

5. Advanced Packaging

6. DCAI Product Roadmap

7. CCG Product Roadmap

8. ASIC Business

9. The Model (and Path to >$1 Trillion)

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

Read more

Like there is a certain ratio of CPU workloads to GPU workloads. But once the inference workload slowly shifts from simple feed forward to agentic loops, the mix of CPU workloads increases, eventually reaching a new equilibrium as agents mature.

This (currently consensus) story is no where near bullish enough.

It is like saying “optics will take share from copper” and just stopping there. You miss so much nuance, the kind that wakes you up to how insanely bullish a certain market is that a traditional bull case does it no justice. In the optics case, you leave out the structural growth of networking bandwidth relative to compute plus each layer of CPO (scale-out CPO, inter-rack scale-up CPO, and intra-rack CPO) being half an OOM larger than the last.

Optics has literally became a meme sector recently which is funny since it is the market I started this publication on and I could have never imagined it could have got to this point (in the best way possible).

Yes, I get it, everything is bullish these days. Every component is a bottleneck. I’m not just saying CPUs will do well on their own. It is that they will outperform GPUs (and basically all other components) relatively and by such a wide margin it may even be bearish Nvidia. (Ok I am def not bearish Nvidia but you get the point)

Because everything is a bottleneck I will actually do a bottleneck tier list at some point just to compare the relative bottleneck-ness of all the bottlenecks. I am so surprised no one has done it but I will happily be the first.

CPUs are up there in the S tier with optics.

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

Contents

1. Fingers and Brains

2. The Demand Side: Earliest in the S-Curve

3. The Supply Side: Immunity to Efficiency Gains

4. TLDR

5. Intel and Price Hikes

6. The Biggest Hidden Beneficiary

Fingers and Brains

Please direct your attention to this following analogy.

Humanity more or less has a fingers-to-brains ratio of 10:1. The brain does the thinking and then the fingers do the doing. This ratio has not really shifted in all of humanity’s history as you might imagine.

But imagine this. Human brains are getting exponentially smarter. They are also getting exponentially faster. This allows humans to do tasks more quickly and do more complex tasks than before. At the same time, they have the capability to buy more arms and fingers if they want to do more tasks. What do you think happens to the fingers-to-brains ratio?

It increases insanely fast.

On top of that, the brain is also getting way more efficient energy (food and sleep) wise. But the fingers stay the way they are. Maybe there are linear improvements as slightly better fingers are released, but according to the laws of fingers, you can never pick up an object with one finger. You can never bowl a bowling ball with two fingers. Brain investors may fear that brain count needs per task are decreased due to efficiency gains, but the fingers needed are nearly unaffected.

That is it for the silliness. The rest of the article is for alpha purposes only.

The Demand Side: Earliest in the S-Curve

Now let us turn our attention to supply and demand. Starting with demand.

Every piece of AI infrastructure sits somewhere on the S-curve. GPUs, memory, and networking were the first to inflect. Training drove the initial wave. Inference scaled it further. Test-time compute added another layer. By now, these components are well into the steep middle portion of their adoption curves, with massive installed bases, deep supply chain investment, and incremental demand from each new workload.

CPUs are still at the bottom of the curve.

GPUs serve training, inference, test-time compute, and agentic workloads. Memory stores model weights, KV caches, reasoning chains, and agent context. Networking connects all of it. For each of these components, agentic AI is the fourth or fifth demand driver layered on top of an already enormous base. It matters at the margin, but the base case was built without it.

CPUs are different. CPUs do not train models. CPUs do not run inference at competitive datacenter scale. CPUs do not power test-time compute. For the entire AI boom, the datacenter CPU has been a utility player: handling orchestration, managing I/O, running the OS. The GPU-to-CPU spend ratio has been lopsided for years because CPUs simply did not matter for the core AI workload.

Agentic AI is the first workload that changes this. When an agent calls a tool, that tool runs on a CPU. Python execution, document retrieval, URL fetching, HTML parsing, JSON serialization, bash execution: all CPU. Raj et al. (2025) in their paper “A CPU-Centric Perspective on Agentic AI” profiled five representative agentic workloads and found that CPU-bound tool processing consumed up to 90.6% of total end-to-end latency. In a RAG pipeline over a large document corpus, retrieval alone consumed 84-91% of runtime while LLM inference contributed less than half a second.

For GPUs, agentic AI is incremental demand on a mature curve. For CPUs, agentic AI is the only new demand driver that matters, hitting a market that has been underinvested for years. That is the difference between adding 10% more demand to an already-strained supply chain and igniting a category of demand that barely existed two years ago.

The Supply Side: Immunity to Efficiency Gains

Now let’s explore the supply side of the equation.

The history of AI hardware is a history of doing more with less. Mixed precision training moved the industry from FP32 to FP8, delivering 2-4x throughput gains per generation. Quantization compresses model weights from 16-bit to 4-bit. Distillation takes 70B-parameter models and produces 7B-parameter students. Sparse attention skips irrelevant tokens. These gains compound multiplicatively, and they apply to GPUs, memory, and networking alike. DeepSeek demonstrated this at scale, achieving frontier-competitive performance with dramatically less GPU compute than Western labs assumed necessary. TurboQuant compresses KV caches, fitting more concurrent users into the same HBM.

As more time passes, we need fewer units of hardware to serve the same tokens.

CPUs do not have these release valves. Consider what a CPU does in an agentic workflow: fetch a URL, parse HTML, execute Python, run nearest-neighbor search, serialize JSON. These are operational tasks, not mathematical ones. You cannot quantize an HTTP request. You cannot distill a database query. You cannot apply sparse attention to a bash script. The optimizations that do exist (better scheduling, GIL removal, NUMA-aware placement) are real but linear: 2x here, 4x there. Compare that to GPU-side compounding where quantization (4x) times distillation (10x) times sparsity (2-4x) stacks to 80-160x effective compute reduction.

Thus, CPU efficiency is slow and linear. The “DeekSeek/TurboQuant moment” threat is removed.

This asymmetry points to a conclusion about what the AI datacenter looks like in equilibrium. Every component is a bottleneck today: GPUs, HBM, networking, power. But the right question is not what is scarce now. The right question is what is the terminal bottleneck, the resource that remains scarce after five years of optimization and capital thrown at the problem. GPUs have a rich optimization roadmap and so does memory. DeepSeek proved you can do more with less GPU compute. TurboQuant proved you can do more with less KV cache memory. CPUs do not have these threats. A CPU cycle is a CPU cycle, and agentic AI needs more of them every year.

TLDR

I literally think CPUs could be a market worth trillions of dollars. If you believe agentic AI changes the world in an exponential way (which I do) it’s impossible not to see CPUs growing exponentially beyond anyone’s wildest imaginations with no DeepSeek moment to put out the fire. The agents are here. They need hands and they need fingers. If the disruption is to happen, it shall be through the margins of our great CPU companies.

That is what we shall look at next.

Intel and Price Hikes

Intel is a long-term share loser in datacenter CPUs. ARM-based custom silicon from AWS, Microsoft, Google, and now Meta is steadily eating x86 market share in cloud workloads. And even in x86 AMD continues to take share with EPYC. That much is consensus and something I agree with despite my massive long position. Too bad, agentic AI is so early in its S-curve that the out-year upside is enormous.

What is new are two near-term dynamics that could meaningfully boost Intel’s earnings right as the agentic CPU demand cycle inflects.

The first is that Intel is the only major CPU supplier that owns its own fabs. Leading-edge logic capacity at TSMC is the single most contested resource in the semiconductor industry right now. AMD, every hyperscaler, and NVIDIA are all fighting for TSMC allocation against each other. When agentic AI demand spikes CPU requirements across the industry simultaneously, everyone fighting for TSMC wafers faces a zero-sum allocation battle. Intel does not. Intel can get moderate (though not memory level) price hikes from the shortage while not suffering from the shortage themselves in terms of shipments.

The second dynamic is operating leverage on a near-zero margin base. The stock has historically traded at low single-digit revenue multiples as a result.

This is where CPU price hikes create asymmetric EPS upside. When you sell a CPU for a higher price because of a shortage, COGS does not change. The silicon was already fabbed, tested, and packaged at the same cost. Every incremental dollar of pricing flows directly to gross profit. For a company operating near breakeven, even modest pricing gains create enormous percentage moves in earnings.

How much can they hike prices? Well it turns out quite substantially. We’ll see how much of this shows up tomorrow during earnings.

Consumer CPUs have climbed roughly 5-10% since March. Server CPUs are getting hit harder, with the weight of industry commentary pointing to 10-15% increases on constrained SKUs, and some estimates suggesting cumulative 2026 hikes could reach ~30%. Intel’s channel chief Dave Guzzi confirmed constraints are hitting partners “across the board.” HSBC upgraded Intel to Buy last week with a price target of $95 (from $50), citing server CPU growth and pricing power as the primary catalyst over foundry.

Though we must note this ain’t memory.

DRAM is a commodity. When DRAM is in shortage, prices clear the market directly: we are living through a historic memory cycle with prices up multiples from trough. The shortage converts cleanly into pricing because any DRAM module is substitutable for any other. CPUs are different. They are SKU-specific, and enterprise customers are locked into specific architectures by software stacks, driver certifications, and supplier contracts. Intel's channel chief Guzzi said explicitly that CPU price increases "should not come close to" memory. Some of the CPU shortage leaks into extended lead times and outright unavailability rather than converting into pure pricing. One gaming PC executive told Nikkei that "even if we pay more, we still cannot get more." That is revenue Intel cannot capture regardless of price. So the operating leverage story is real but bounded: 10-15% price hikes on constrained SKUs, not 400% commodity repricing.

The Biggest Hidden Beneficiary

There is one company that is hilariously well positioned set to capture the majority of the upside of the steepest part of the S-curve in the out-years once agentic AI conquers the entire world.

Read more

The purpose of this article is to make one very simple point: Neoclouds are like foundries.

If you were Nvidia, do you choose your foundry based on who offers the absolute lowest price per raw wafer?

The foundries are a commodity right? They all buy their tools from the same vendors for the same price. They all share the same list of concentrated customers. The all sell the same product, no?

Hell no.

The price per wafer tells you nothing. Wafers don’t turn into Blackwells. Known-good-dies do. Every foundry conversation is dominated by yield. That’s all that matters! How hard you’re cranking the factory doesn’t pay the bills; only the factory’s actual output does.

Yet in cluster land people only care about the GPU rental price. The bull case for hyperscalers (and bear case for neos) rests on their economies of scale. They can bend the supply chain to their whim, amortize fixed cost on custom networking, and even build their own ASICs. With the lowest gross cost of compute, how can smaller, debt-fueled clouds ever compete?

My thesis is that this consensus is wrong. Neoclouds are fundamentally comparable to foundries. The gross number of GPU hours you paid for does not matter. It’s equivalent to the gross wafers processed by TSMC — a vanity input metric completely subservient to the actual output achieved. The real measure of success is goodput, the actual useful work done for training/inference and the neocloud equivalent of known-good-dies. Cloud providers that achieve the highest goodput with a certain quantity of compute, the equivalent of yield, have a sustainable competitive advantage.

In addition, achieving goodput is not easy. Just like foundry yield, it takes years of process learning and optimization. The best companies start ahead and stay ahead, leveraging a competitive advantage known as process power (from Hamilton Helmer’s 7 Powers). And the hyperscalers, with decades of running traditional web workloads, were fundamentally unprepared for the east-west dominated AI compute paradigm, giving top neoclouds a process learning head start.

One of the most important reasons I wrote the microeconomic model based on the consensus that neoclouds were a commodity is to break that foundational premise. That’s what we shall do today.

CoreWeave is a true out-of-consensus call for me. Optics/CPO was already a consensus long, I was just saying that it is bigger than people thought. Nobody is short optics companies with conviction. But because of how hated neoclouds are, if I’m right the repricing would be much more violent. The recent ~40% move means nothing to me as it is peanuts compared to what could happen. They can easily turn into the next momo trade. It’s the only long in which I hold calls and not shares. Again — debt laden, low-margin, cash flow negative is bad for safety but excellent for asymmetry.

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

Contents

1. Consensus = Anti-Neocloud

2. Hyperscalers = Overrated

3. GOODPUT = YIELD

   1. Pretraining

   2. Reinforcement Learning

   3. Inference

   4. Software Can’t Fix This

   5. Capital Can’t Buy It Either

4. The Model (GPU Pricing Build, Revenue Build, Income Statement, DCF)

5. CoreWeave = TSMC

   1. CoreWeave is Underearning

   2. Hyperscalers are Overearning

   3. Becoming Their True Self

Consensus = Anti-Neocloud

CoreWeave is one of the most hated stocks in the market. Not hated in the way a melting ice cube is hated, where bears quietly accumulate shorts and move on. Hated with conviction. The short thesis writes itself so cleanly that it has become consensus among generalist investors, sell-side analysts, and activist short sellers alike.

A prominent activist short seller published a report last year calling CoreWeave “an undifferentiated, heavily levered GPU rental scheme stitched together by timing and financial engineering, not lasting innovation.” No proprietary technology. No defensible IP. No edge. Microsoft’s Satya Nadella himself called the original deal “a one-time thing.” The price target: $10, implying 90% downside. The thesis in one sentence: CoreWeave is a stopgap bare metal assembler that will be displaced the moment hyperscalers finish building out their own capacity.

A major sell-side research shop arrived at the same conclusion through a more rigorous financial lens. According to their model, CoreWeave’s ROIC of roughly 10% barely covers its cost of capital, meaning every incremental dollar of investment destroys value. They estimate long-term non-GAAP operating margins of 15-17%, far below management’s 25% guidance, because depreciation alone eats roughly 50% of revenue at current asset turnover levels. Their framing is devastating in its simplicity: GPU-focused clouds are getting commoditized, CoreWeave’s tech-savvy customers can easily backward integrate, and growth at sub-WACC returns destroys shareholder value with every incremental dollar of capex.

The common thread across every bearish argument is a single foundational premise: GPU cloud is a commodity activity with no operational differentiation. The hardware is fungible. Any well-capitalized competitor can buy the same NVIDIA GPUs, rack them in a data center, and offer the same product. Customers can switch between neoclouds frictionlessly. Hyperscalers will inevitably internalize. Therefore the margins are structurally thin, the returns are structurally poor, and the debt load will crush equity holders when the music stops.

I want to be clear about what the bears get right. The debt load is real. The customer concentration is real. The capex intensity is real. The cash flow profile is genuinely ugly on a near-term basis. These are not fabricated concerns and I am not dismissing them.

But every bearish conclusion, sub-WACC returns, no pricing power, hyperscaler displacement, zero residual value, flows downstream from that single premise: that this is a commodity business. If the premise is wrong, the entire bear case collapses.

My argument is that the premise is wrong. And the reason generalist analysts and finance-focused investors cannot see it is that the evidence is not financial. It is technical. You cannot find the moat in an asset turnover ratio. You cannot find it in a depreciation schedule. You find it inside the data center, in the operational machinery that separates a cluster that works from a cluster that burns venture capital. The financial statements are a lagging indicator of something the bears have never measured.

Hyperscalers = Overrated

The bull case for hyperscalers rests on scale. They buy hardware by the container ship. They negotiate power agreements with entire municipalities. They have the lowest gross cost of compute on the planet. How can a smaller, debt-fueled neocloud possibly compete?

This framing sounds intuitive, but it gets the physics of GPU clusters completely wrong. Hyperscaler scale advantages were built over two decades of optimizing for stateless web microservices: millions of tiny, independent requests where any individual server failure is irrelevant. A GPU training cluster is the exact opposite. It is a single, tightly coupled supercomputer where thousands of GPUs must synchronize simultaneously. If one node fails, the entire job stops. Hyperscalers are retrofitting a data center paradigm built for web traffic to run massively parallel supercomputing jobs. The neoclouds built for this paradigm from day one.

This architectural mismatch shows up as hidden costs everywhere. SemiAnalysis published an exhaustive TCO framework recently:

SemiAnalysis

How Much Do GPU Clusters Really Cost?

Introduction: Rethinking the Total Cost of a GPU Cluster…

Read more

17 days ago · 98 likes · 2 comments · Jordan Nanos, Bryan Shan, Cheang Kang Wen, Daniel Nishball, and Dylan Patel

They decompose the total cluster cost far beyond the headline GPU rental rate. The hidden cost stack includes: hot NVMe storage priced as a premium add-on rather than bundled as essential plumbing, orchestration premiums for using Kubernetes or Slurm through proprietary managed services, enterprise support billed as a percentage of total cloud spend (up to $450K/month at scale just for a tiered ticketing system, versus CoreWeave’s direct Slack channel with an engineer included in the base price), mandatory paid proof-of-concept periods where customers rent entire clusters for a month ($15M+) just to tune proprietary networking protocols before training even begins, and nickel-and-dime egress and data transfer fees layered on top.

To quantify this, SemiAnalysis modeled a pretraining run on over 5,000 next-generation GPUs and assumed all providers charge the exact same headline rate per GPU-hour. The starting line was perfectly level. Over a 36-month deployment, the hyperscaler's total cost swelled to 1.10x the gold-tier neocloud's. A 10% hidden tax that never appears on the procurement spreadsheet. And this is before we account for goodput.

And this is before we account for goodput.

GOODPUT = YIELD

Throughput measures activity. Goodput measures progress. The distinction is everything.

A 4,000-GPU pretraining cluster is a single synchronized organism. Every GPU holds a piece of the neural network, and they must all communicate their calculations simultaneously. If one GPU fails, a transceiver overheats, or a memory error spikes at the wrong moment, the entire job halts. All 3,999 healthy GPUs are stranded. All compute since the last checkpoint is vaporized. Then the system spends 10-15 minutes identifying the fault, swapping the node, and reloading state, during which you are still paying for the full cluster.

How often does this happen? Individual GPUs are remarkably reliable (roughly 25,000 hours between failures on premium infrastructure). But when you combine 4,000 of them into one tightly coupled system, the cluster as a whole fails every 6.25 hours. On a lower-tier provider with worse component reliability, it fails every 3.75 hours. SemiAnalysis calculates total goodput loss of 6.14% for gold-tier providers, 10.53% for hyperscalers, and 20.91% for silver-tier (Guys is this not the same thing as yield?). One-fifth of the silver-tier’s multi-million dollar budget evaporates into recovery cycles.

CoreWeave is the sole occupant of SemiAnalysis’s ClusterMAX Platinum tier.

Every other provider, Azure, AWS, GCP, Oracle, Nebius, Crusoe, sits below them. What earns the Platinum rating is not hardware (everyone buys the same NVIDIA GPUs) but the operational machinery built through years of running the world’s most demanding clusters: automated burn-in testing against reference performance numbers, passive health checks every few seconds, weekly active diagnostics including silent data corruption detection, proprietary monitoring exporters built because standard tools were not granular enough, and a correlation engine that can automatically distinguish switch-level failures from tray-level failures by tracking error patterns across the fabric. The result is pricing power: CoreWeave commands a 10-15% premium per GPU-hour over neocloud peers, and customers pay willingly because the goodput justifies it.

This advantage manifests differently across workloads, but CoreWeave wins all three.

Pretraining

This is the classic goodput case. The blast radius of failure encompasses the entire cluster. Maximum GPU-hours between failures is the only thing that matters. This is where CoreWeave’s health check infrastructure and hot spare pools pay for themselves most directly.

Reinforcement Learning

RL neutralizes the blast radius problem (jobs are fragmented into small 64-GPU chunks, so a single failure only affects 3% of the cluster) but the hyperscaler still loses because multimodal RL requires massive hot storage (tens of petabytes in the modeled scenario) and hyperscalers nickel and dime on storage pricing and orchestration overhead. SemiAnalysis models a 1.61x TCO premium for hyperscalers on RL workloads.

Inference

This is where the leverage becomes existential. You are selling tokens at a fixed market price. Your input cost is GPU-hours. Your gross margin is the sliver in between, and it is brutally sensitive to infrastructure cost. According to SemiAnalysis benchmarks, a roughly 10% increase in the underlying cost per GPU-hour can compress gross margins by a third. On a hyperscaler, where the all-in goodput expense runs roughly 1.6x higher than a top-tier neocloud, inference startups can find themselves with negative gross margins: literally losing money on every token they sell. The more successful the product becomes, the faster the startup bleeds. Choosing the right infrastructure provider is the difference between a viable business and insolvency.

Software Can’t Fix This

The immediate counterargument: just use fault-tolerant software on cheap infrastructure. The industry is trying. Every solution trades one cost for another. Meta’s open-source TorchFT keeps training running through failures but forces GPU communication through the CPU instead of direct GPU-to-GPU RDMA, imposing a permanent performance tax on every calculation. AWS’s Checkpointless Training consumes 5% of GPU memory on every chip for redundancy. Clockwork.io’s TorchPass requires 32 idle GPUs as dedicated hot spares, doing nothing but waiting.

The best fault tolerance is simply running clean data centers with talented ops teams. Software band-aids are like fixing a burst pipe by buying more buckets. A Platinum neocloud builds a pipe that does not burst.

Capital Can’t Buy It Either

Just like how no amount of money can replicate TSMC’s node learning (see Japan’s Rapidus), the operational knowledge to turn hardware into productive clusters cannot be purchased. It is built through years of operating at scale, accumulating failure mode taxonomies, and iterating on monitoring and recovery systems cluster after cluster. This is process power. It is opaque, it compounds, and it is exactly what separates CoreWeave from everything below the Platinum tier.

Two anecdotes from SemiAnalysis’s ClusterMAX testing illustrate this. Core42, backed by effectively unlimited UAE sovereign wealth, attempted massive AMD MI300X clusters but used network cards incompatible with AMD’s standard software stack. PhD researchers spent weeks manually downloading driver archives and recompiling containers from scratch just to get the hardware to recognize itself. BitDeer secured the GB200 NVL72, the most coveted AI hardware on earth, but failed to configure the internal memory mapping protocol correctly, bottlenecking a multi-million dollar unified supercomputer to the communication speeds of a gaming PC, with health monitoring completely disabled.

The Model

Below I share my model for CoreWeave. GPU pricing build with assumptions per generation, revenue build using contracted power, mix, and software overlay, income statement and EPS figures, DCF and price targets. It’s one of my biggest and most complex models I’ve built to date.

This is a bull case scenario drawn from extrapolating current GPU shortage pricing trends (H100 rental rates mooning). The results, put plainly, will shock you.

And finally, a fifth section on how I expect this thesis to play out. Why CoreWeave is underearning, hyperscalers are overearning, when and why it will flip, and most importantly, after this current hypergrowth phase, what I expect to be CoreWeave’s final form. Let’s go.

Read more

The reason I jump around with my writeups so much is that they are usually the result of me going down epic rabbit holes. This time, I felt the urge to explore a curious puzzle…

I find it funny people don’t talk about memory MORE. So basically one could have invested $30k in Sandisk last year and be a millionaire today lmaooooo.

How did this even happen?

DRAM is a commodity too. It is tight too. In fact, DRAM is talked about way more, as HBM (and even server DRAM) is a much more obvious complement to GPUs, but Micron has performed nowhere near as well as Sandisk.

And some people still have (very credible) arguments for it still being significantly undervalued, how is that possible?

The biggest curiosity I had was the differences between the DRAM and NAND markets. I wanted to understand what caused the NAND outperformance. The fundamental differences in their market structures. And most importantly, which is the best way to play AI.

Naturally, I spent a full day on Saturday just sitting in a room voice-chatting with Claude about memory .

What I discovered is that the performance of memory is at its core based on the balance of supply/demand.

no shit.

I know it sounds obvious. But I promise this is a more important insight than at first glance. I’m saying almost everything in memory reduces this. If we understand the direction of supply and the direction of demand, we can generally predict what happens to memory.

This is especially useful in comparing the relative trajectories of DRAM and NAND because they have separation in both their supply and their demand.

That is why the first two articles about memory will be discussing the supply side and the demand side separately. And honestly the article after that (ASP likely direction and uncertainty) will just be building on our baseline understanding of supply and demand.

These memory articles won’t really cover Sandisk or any memory name in particular just because there isn’t too much differentiation in a commodity industry (duh) but I will be opining on what stock(s) I like in memory today.

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

Contents

1. Geometry and Bit Scaling

2. Tool Requirements

   1. NAND: Mostly Etch

   2. DRAM: Mostly EUV

3. Tale of Two Oligopolies

4. China

5. HBM

6. The Clear Winner of Round 1

7. TLDR

Geometry and Bit Scaling

Both DRAM and NAND are hitting fundamental capacity constraints. But the nature of those constraints, and how they got here, is different.

DRAM is a planar (flat) technology. A blanket of bits.

Made up of these little 1t1c (one transistor one capacitor) cells that store charges.

Each generation, manufacturers shrink transistors laterally, from 1-alpha to 1-beta to 1-gamma node designations. You’re packing more bits into the same wafer area by making features smaller.

NAND is a vertical 3d technology. A whole bunch of round pillar-like cells stacked on top of each other with hundreds of layers. This is why it is so much cheaper and denser per GB than DRAM but obviously comes with the bandwidth tradeoff.

To scale, rather than shrinking features laterally, manufacturers increase the number of layers: 96 layers, then 128, then 176, then 218, and now pushing toward 300+. For years, this vertical scaling was NAND’s escape hatch from the capacity wall. It allowed NAND bit capacity to grow more easily than DRAM (and therefore had a weaker supply market structure). You didn’t need to shrink features or buy expensive lithography tools. You just added more layers using existing deposition and etch equipment, and each generation delivered more bits per wafer without proportional capex increases.

That escape hatch is closing.

As layer counts push past 200, the engineering challenges compound. You’re etching memory holes through stacks of material with aspect ratios exceeding 100:1. Imagine drilling a hole 100 meters deep but only 1 meter wide, through alternating layers of different materials, and requiring nanometer-level precision at every depth. The mechanical stress on the wafer increases. Yield falls. Each incremental layer delivers less marginal cost reduction than the last. SanDisk’s own investor day data shows the cost reduction per generation from vertical scaling has fallen to approximately 0.24x, down from substantially higher levels in earlier generations.

This means NAND is converging toward DRAM’s capacity constraint model: adding bits now requires real capex, real fab buildouts, and real time. The era of cheap supply growth through “just add more layers” is ending. Both markets are now cleanroom-constrained, and both face multi-year lead times for greenfield expansion.

So TLDR, in terms of bit scaling, NAND was historically easier to scale than DRAM which meant constant gluts of new supply, but that era is ending as vertical scaling hits a wall, so NAND and DRAM capacity additions via scaling are converging.

But here is where the similarities end, and where the structural differences in supply become critical.

Tool Requirements

The manufacturing processes for DRAM and NAND are fundamentally different, and those differences create distinct bottlenecks in the equipment supply chain.

NAND: Mostly Etch and Deposition

Making a 3D NAND chip is, at its core, a vertical construction project. You start with a silicon wafer and build upward, depositing alternating layers of oxide and nitride (or polysilicon), then punching vertical holes through the entire stack to create the memory cells.

The process, simplified, works like this: First, you deposit hundreds of alternating thin films, one on top of another, using chemical vapor deposition (CVD) tools. These tools must coat each layer uniformly across the entire wafer surface while maintaining precise thickness control. Each layer is only nanometers thick, and the total stack can be several micrometers tall.

Then comes the critical step: etching. Deep reactive ion etch (DRIE) tools bore vertical channels through the entire stack. These channels will become the memory strings, the vertical columns of cells that store data. The etch must maintain near-perfect verticality through 200+ layers without the channel widening at the top or narrowing at the bottom. This is among the most demanding etch processes in semiconductor manufacturing.

After the channels are formed, additional deposition steps fill them with charge-trapping layers and electrode materials, creating the actual memory cells at each layer intersection. Then the “staircase” structure is etched at the edge of the array to allow individual electrical contact to each layer.

The result: NAND manufacturing is dominated by etch and deposition tools! These two categories together account for roughly 50-60% of total NAND equipment spending. Lithography, while still important for patterning the top-level features, is less critical than in DRAM because NAND’s lateral features can be larger! The density comes from stacking, not shrinking.

Lam Research is the dominant supplier of advanced etch tools for NAND. Their Kiyo and Syndion platforms are purpose-built for the deep, high-aspect-ratio etches that 3D NAND requires. When NAND vendors spend capex, a disproportionate share flows to Lam. Applied Materials captures the deposition side, with their Endura and Centura platforms handling the hundreds of thin-film deposition steps in each NAND stack.

This tool mix matters for the supply story: NAND capex translates primarily into orders for etch and deposition equipment, which are available from multiple suppliers and can be delivered on relatively normal timelines. There is no single tool bottleneck that gates NAND expansion the way there is for DRAM.

DRAM: Mostly EUV

DRAM’s manufacturing challenge is lateral, not vertical. You’re shrinking transistors and capacitors on a 2D plane, pushing feature sizes below 15nm. At these dimensions, conventional DUV lithography cannot resolve the features. You need EUV.

EUV lithography is, without exaggeration, the most complex manufacturing technology humans have ever built. Each EUV scanner costs approximately $200-400 million. It uses a 13.5nm wavelength light source generated by vaporizing tin droplets with a high-power laser, bouncing the resulting plasma emission off multilayer mirrors in a near-vacuum chamber. The mirrors must be polished to sub-angstrom precision. The entire optical path operates in vacuum because EUV light is absorbed by air. The supply chain for these machines spans dozens of countries and hundreds of specialized suppliers.

ASML is the sole manufacturer of EUV lithography systems. There is no second source. And ASML’s production capacity is not easily scalable. They are adding perhaps a dozen machines per year to their output, and demand from logic foundries (TSMC, Samsung Foundry, Intel) competes directly with demand from DRAM fabs for those same machines.

This creates a hard bottleneck for DRAM capacity expansion that has no equivalent in NAND.

Consider the numbers: Samsung recently ordered approximately 20 EUV systems for a single new fab (P5), along with roughly 70 DUV tools. ASML’s EUV delivery slots are fully sold out through 2027, with negotiations for 2028 allocations already underway. Every EUV machine allocated to a DRAM fab is one that cannot go to a logic foundry, and vice versa.

The implication is stark: even if DRAM vendors want to add capacity aggressively, they are physically gated by EUV tool availability. This is a supply constraint that operates on a multi-year timescale and cannot be resolved by throwing money at the problem. ASML’s production ramp is measured in single-digit percentage improvements per year, not the kind of step-function increase that would relieve the bottleneck.

NAND faces no equivalent constraint. NAND fabs use DUV lithography (which is widely available from multiple suppliers) and advanced etch/deposition tools (which Lam, Applied, and others can produce at scale). The tools are expensive, but they are not scarce in the way EUV is scarce.

This asymmetry in tool availability is one of the most underappreciated structural differences between the two markets! It means DRAM supply growth is slower, more capital-intensive, and harder to accelerate than NAND supply growth, even when both markets face strong demand.

The Tale of Two Oligopolies

Supply discipline is a function of market structure. The fewer the players, the easier it is to maintain pricing power, whether through explicit coordination or simply through the rational self-interest of a small group.

DRAM has three major producers: Samsung, SK Hynix, and Micron. Together, they control approximately 94% of global DRAM bit shipments. This is a textbook oligopoly. Each player understands that aggressive capacity expansion destroys pricing for everyone, and each has learned from prior cycles where overcapacity led to catastrophic margin compression. The memory of those downturns acts as a natural governor on investment behavior.

The DRAM oligopoly is further reinforced by the HBM dynamic (which we’ll discuss in Section 5). When the highest-margin product in your portfolio (HBM) requires you to divert standard DRAM wafers, you have a built-in mechanism for constraining commodity DRAM supply without explicitly coordinating with competitors. The market structure does the discipline for you.

NAND has six significant producers: Samsung (~32% of bits shipped), Kioxia (~18%), SanDisk (~13%), SK Hynix/Solidigm (~12%), Micron (~12%), and YMTC (~7% and growing rapidly). This is a meaningfully more fragmented market.

With twice as many significant players, supply discipline becomes harder to maintain. Each vendor faces the prisoner’s dilemma: if everyone restrains capacity, prices stay high and everyone profits. But if one vendor defects and expands aggressively, they capture share while competitors suffer. The incentive to defect is stronger when there are more players, because the cost of being the one who doesn’t expand is higher.

This dynamic has played out repeatedly in NAND’s history. Prior cycles saw aggressive capacity additions from multiple vendors simultaneously, leading to oversupply, price crashes, and margin compression far more severe than anything DRAM experienced. The current cycle’s discipline is notable precisely because it’s unusual for NAND.

The question investors must answer is: is this discipline structural or temporary? If it’s structural, driven by layer scaling limits, LTAs with floor prices, and rational vendor behavior, then NAND pricing can remain elevated. If it’s temporary, a function of the current tightness that will unwind as capacity comes online, then the six-player structure makes normalization faster and more violent than in DRAM.

DRAM’s three-player structure provides a higher floor on discipline regardless of cycle conditions. Even in downturns, three players can more easily reach an implicit equilibrium on capacity and pricing than six can.

China

China’s semiconductor ambitions create different risks for DRAM and NAND, and the asymmetry is larger than most investors appreciate.

YMTC: A Real Threat to NAND

Yangtze Memory Technologies Corporation (YMTC) is not a theoretical competitor. It is a rapidly scaling NAND manufacturer that is approaching the technological frontier.

YMTC’s Wuhan Fab 3 is expected to begin full-scale operations in the second half of 2026. Once operational, YMTC is projected to surpass both SK Hynix and Micron in NAND shipment volume, becoming the world’s third-largest NAND producer. They have been steadily increasing their mix of high-complexity 200+ layer NAND across server and mobile segments, and are expected to stabilize yields on 300-layer products this year.

Perhaps most significantly, YMTC has reportedly sourced more than 50% of the equipment for one of its new fabs from domestic Chinese companies, including key tools for vertically stacking chip layers. This suggests that export controls and entity list restrictions are not preventing YMTC from building and equipping advanced NAND fabs. This is a development that has profound implications for the durability of NAND supply constraints.

The market signal that matters most: Apple is reportedly evaluating the use of YMTC products due to NAND supply constraints. If Apple, the most demanding and powerful buyer in the consumer NAND supply chain, begins qualifying YMTC, the credibility barrier falls for every other OEM. YMTC would transition from a “China-domestic supplier” to a globally qualified NAND vendor, and the effective supply base for the NAND industry would expand meaningfully.

CXMT: A Meh Threat to DRAM

ChangXin Memory Technologies (CXMT) is China’s primary DRAM entrant, but its competitive position is far weaker than YMTC’s in NAND.

CXMT only entered mass production of HBM3 in Q1 2026, and yields remain very low. Advanced DRAM requires EUV lithography, which CXMT cannot access due to export controls. Without EUV, CXMT is limited to older process nodes that cannot compete on cost or performance with Samsung, SK Hynix, or Micron’s latest products.

Moreover, whatever DRAM capacity CXMT adds is being absorbed by domestic Chinese demand, primarily Huawei. Samsung has been discontinuing legacy HBM sales to China, and Huawei cannot procure enough HBM from overseas sources. CXMT is essentially Huawei’s captive memory supplier, which means incremental CXMT capacity does not translate into incremental global DRAM supply.

The result: CXMT’s impact on the global DRAM market is contained. Its capacity is spoken for, its technology is behind, and its access to critical tools is restricted. The DRAM oligopoly’s 94% market share is not under serious threat.

YMTC’s impact on the global NAND market is growing. Its technology is competitive, its domestic equipment sourcing reduces its vulnerability to sanctions, and its potential qualification with Apple could open global markets. The NAND market’s fragmentation could increase further if YMTC achieves the scale its expansion plans imply.

HBM

The most important structural difference between DRAM and NAND supply is something that has no parallel in NAND’s history: two distinct products, serving different layers of the memory hierarchy, suddenly competing for the same manufacturing resources.

To understand why this matters, consider the memory hierarchy itself. In an AI server, HBM sits on the GPU package, providing extreme bandwidth (~5 TB/s) at extreme cost. DDR5 sits on the motherboard, providing large capacity at moderate cost. These are two separate layers with different latency, different bandwidth, and different price points. The gaps between them are not percentages or small multiples, but orders of magnitude.

Historically, these two layers were supplied by different manufacturing processes and, in some cases, different fabs. HBM was a niche product. DDR was the mainstream. They coexisted without competing for the same wafer starts.

That is no longer true.

HBM and standard DRAM are now manufactured in the same fabs, on the same process nodes, using the same EUV lithography tools, by the same three vendors. Every HBM stack requires approximately 3x more silicon per bit than standard DDR5, because of the TSV (through-silicon via) process, the yield loss from stacking, and the logic die at the base. And HBM commands dramatically higher margins, estimated at 5-10x the margin per wafer versus commodity DDR5.

The economic logic is inescapable: every DRAM vendor will preferentially allocate wafers to HBM over standard DRAM, because HBM generates far more profit per unit of scarce fab capacity. This is not a choice that requires coordination or discipline. It is the rational profit-maximizing behavior of each individual vendor, operating independently.

The consequence is a structural suppression of standard DRAM supply that operates automatically, without any vendor needing to exercise “discipline” in the traditional sense. The more HBM demand grows (driven by GPU shipments from Nvidia, AMD, and custom ASICs), the more standard DRAM supply is constrained. HBM becomes a permanent tax on DDR5 supply.

NAND has no equivalent mechanism. There is no premium NAND product that consumes 3x the silicon and commands 10x the margin, pulling wafers away from standard enterprise SSDs. High-Bandwidth Flash (HBF) is an early concept, but it is not in volume production and its economics are unproven. The “Super High IOPS SSD” that Kioxia has demonstrated is a differentiated product, but it does not cannibalize standard NAND capacity the way HBM cannibalizes standard DRAM capacity.

This is arguably the single most important structural advantage DRAM has over NAND on the supply side. DRAM supply is being constrained by an internal, self-reinforcing mechanism, the diversion of wafers to HBM, that operates independent of vendor discipline, market conditions, or competitive dynamics. NAND supply is constrained by the traditional factors of capex discipline, layer scaling limits, and demand growth. Those factors are real, but they are more fragile. They can reverse if vendors lose discipline, if layer scaling finds a second wind, or if Chinese capacity comes online faster than expected.

The Clear Winner of Round 1

The supply side of the DRAM vs. NAND question is not a draw. DRAM has structural advantages at every level:

The capacity wall affects both markets, but DRAM expansion is gated by EUV tool scarcity while NAND expansion faces no equivalent single-point bottleneck. DRAM’s three-player oligopoly enforces discipline more naturally than NAND’s six-player market. China’s DRAM entrant (CXMT) is contained by technology gaps and domestic demand absorption, while China’s NAND entrant (YMTC) is scaling rapidly and potentially entering global supply chains. And HBM creates a self-reinforcing mechanism that constrains standard DRAM supply without requiring explicit discipline, while NAND has no equivalent internal constraint.

None of this means NAND supply is loose. It is demonstrably tight today, and the layer scaling slowdown, equipment delivery bottlenecks, and LTA structures provide meaningful near-term support. But the medium-term supply outlook carries more risk in NAND than in DRAM, because NAND has more paths to supply relief and fewer structural barriers to capacity expansion.

In Part 2, we will examine the demand side, where the picture inverts: NAND’s demand story may be structurally stronger than DRAM’s, driven by the shift from ephemeral compute to persistent state in the age of agentic AI. The question investors must answer is whether demand strength can compensate for supply vulnerability, and for how long.

TLDR

DRAM wins the supply round! DRAM supply is structurally tighter than NAND. Both markets face capacity constraints, but DRAM expansion is bottlenecked by EUV lithography tools that only ASML can make, while NAND can scale with widely available etch and deposition equipment. DRAM has three major producers controlling 94% of the market. NAND has six. China's DRAM entrant (CXMT) is years behind on technology and its output gets absorbed by Huawei. China's NAND entrant (YMTC) is approaching the frontier and Apple is reportedly evaluating their products. Most importantly, HBM and standard DRAM compete for the same fabs, the same wafers, and the same EUV machines. Every wafer allocated to HBM is one that cannot make DDR5. NAND has no equivalent internal constraint on supply. The result: DRAM supply is harder to expand, better disciplined, and structurally self-constraining in ways that NAND is not.

Today I give a one minute non-consensus take on why corporate debt is actually really good. Enjoy!

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

Consensus says debt is always bad. “Net cash” and “fortress balance sheet” are the biggest green flags to most public markets investors.

Or one can cite Modigliani-Miller and argue “well actually it doesn’t matter because WACC never changes!”

The academic argument behind this is pretty simple. Debt and equity are both ways of financing a business. As stock market investors we own the equity. If the company is levered our equity is more risky. If the company is unlevered it is less risky. But it actually doesn’t matter because with the more levered business you can just own it in a portfolio with cash and for the unlevered business you can lever up yourself.

Essentially, because the investor controls their own leverage, the leverage chosen by the company doesn’t matter.

Moving from the theoretical world into the real world, the only variables left are the debt tax shield (positive for debt, WACC goes down) and the risk of bankruptcy (negative for debt, WACC goes up) which act as countervailing forces, meaning companies in a certain industry usually have an ideal level of leverage that minimizes WACC.

So that is the foundation. How will we challenge it?

I am a thematic investor. I believe in the future of AI and by extension the semiconductor industry. Many hedge funds do the same. Our jobs are to express our thematic view to the fullest extent without taking on excessive risk.

Why the excessive risk clause? Well, if I lever up my portfolio 5 to 1, I’ve just introduced massive path dependency. Those of you who read my shorting article know what I’m talking about. I could be totally right about the destination, but now because I’m levered, I’m dependent on the path, specifically that of the share price.

As all value investors tend to scream at us growth guys for, price does not always equal value. You can be right on the value, but if price takes a drawdown and you get margin called, you are fucked. This is why we must manage risk.

But see, we just talked about the investors’ ability to lever up as a reason why corporate leverage is irrelevant right?

“If the company is levered our equity is more risky. If the company is unlevered it is less risky. But it actually doesn’t matter because with the more levered business you can just own it in a portfolio with cash and for the unlevered business you can lever up yourself.”

My argument breaks this exact premise. Investor leverage is NOT the same as corporate leverage. Investor leverage introduces path dependency to the price while corporate leverage does not. For a corporation to go bankrupt, they must be unable to service their interest with cash, i.e. the value must go to zero. One is price, the other is value.

For thematic investors, this is profound. With leveraged corporations, you can get more exposure to your theme (and the direction of fundamental value) without taking on the path dependency that comes with individual leverage! Higher expected returns (provided you are correct on your thematic) with no path dependency (hold through drawdowns without fear of margin calls).

Below, I’ll share a name I’ll be entering next week (and writing up sometime) that applies this principle to the tee.

Read more

I got into shorting earlier this year and I ended up losing a few hundred bps in total. Now, I’ve stopped and closed them all to focus purely on longs. My PnL is still fantastic thanks to my longs but I thought this was a worthy learning experience.

Shorting (particularly unpaired short theses) is not the same as long-ing! They require different mindsets, different skills, and a different foundational framing. A fantastic long investor can be a terrible short investor (me) and vice versa.

And yeah I do publish short theses and will continue to do so even if I am not short anything most of the time. This is actually part of a prudent short-seller’s strategy as you’ll see in the writeup. No contradiction here.

So yes, this is an entire textbook chapter article about shorting. Just shorting and the first principles and philosophy behind it. Nothing else. Enjoy!!!

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

Contents

1. Path Dependency

2. A Good Thesis Does Not Make a Good Trade

3. Price-to-Value Convergence

4. How Long vs. Short Investors Enter a Trade

5. Paired vs. Unpaired Shorts

6. Why I Am a (Mostly) Long-Only Investor

Path Dependency

Path dependency is a concept borrowed from mathematics and physics, and it shows up anywhere the history of how a system evolved matters to the final outcome, not just the endpoints. A path-independent process only cares about where you start and where you end. A path-dependent process cares about every step in between. Gravitational potential energy is path-independent: the energy it takes to lift a book onto a shelf is the same whether you lift it straight up or walk it around the room first. Friction is path-dependent: the energy lost to dragging that book across the floor depends entirely on the route you took to get there.

In economics the term gets used to describe why certain outcomes are locked in by the sequence of events that produced them. QWERTY keyboards are the standard not because they’re optimal but because of the specific historical path of typewriter adoption. The concept generalizes to any system where the trajectory, and not just the terminal state, determines what you end up with.

Investing sits on both sides of this distinction, and the side you sit on depends entirely on whether you are long or short.

A long position is path-independent. If you buy a stock you believe is worth $200 at $100, the only thing you need to be right about is the destination. The path the stock takes to get there is almost irrelevant to your eventual P&L. It can trade down to $60 first. It can sit flat for three years. It can rally to $150, give it all back, and then finally grind to $200 in year five. None of these paths change the fact that you bought at $100 and eventually exited at $200. If anything, the adverse paths are gifts, because they let you buy more at better prices. Your cost basis improves, your future returns improve, and when convergence finally arrives you have monetized the volatility rather than suffered it. Time is on your side because the business compounds, pays dividends, buys back stock, or gets taken out. The forcing functions of corporate finance drag price toward value on a timeline that does not need to be yours.

A short position is path-dependent in exactly the way that a long position is not. If you short a stock at $100 that you believe is worth $0, you cannot simply wait for the destination. The path matters because every point along the path has the capacity to end the trade before the thesis plays out. If the stock rallies to $200 first, you are margin called, forced to cover at the top, and the $0 destination arriving six months later is a story you tell at dinner rather than a number in your P&L. If the stock sits flat for two years, borrow cost has eroded your return even if you are eventually right. If the stock grinds down slowly, you are paying to hold the position the entire time. There is no mechanism by which a worthless business is forced to trade at zero. There are only participants who decide to sell, or refuse to buy, and that decision can be deferred indefinitely.

The practical consequence is that a long thesis only requires certainty about the endpoint, while a short thesis requires certainty about the endpoint AND the trajectory. You have to be right about what the business is worth, and you have to be right about how the market gets from here to there, and you have to survive every intermediate price the market decides to print along the way. These are three separate analytical problems, and any one of them being wrong kills the trade even if the other two are correct.

This is why a correct thesis does not make a good short. A thesis is a claim about the destination. A short is a claim about the destination, the path, and the clock, all at once.

Below we discuss why a correct thesis is only the starting point of a short and not the trade itself, the structural reason price converges to value on the long side and refuses to on the short side, the entirely different entry criteria that path dependency forces on long versus short investors, the taxonomy of paired and unpaired shorts and what basis risk has to do with running a real hedge fund, and finally why all of this leads me to operate as a mostly long-only investor despite having spent considerable time building out a short watchlist I have not deployed.

A Good Thesis Does Not Make a Good Trade

There is a distinction that gets collapsed in most retail and junior-professional thinking about investing, and the cost of collapsing it is much higher on the short side than on the long side. The distinction is between a thesis and a trade.

A thesis is a claim about a business. The company is overearning. The moat is eroding. The TAM is shrinking. The accounting is aggressive. The margin structure is unsustainable. Management is incentivized to destroy value. These are all statements about the underlying enterprise and its trajectory as a going concern.

A trade is a claim about the market’s price for that business over a specific window. It is a prediction that the quoted price will move from where it is now to somewhere else, that this move will happen before borrow and opportunity cost erode the position, and that the path between those two prices will not force you out of the position before you are paid.

Longs can afford to blur this distinction because the forcing functions of corporate finance do a lot of the work for them. If the thesis is right and the business compounds, the trade tends to work eventually, even if the entry was mediocre, the timing was early, and the path was ugly. The long investor is riding a vehicle that drags price toward value on its own schedule, and the thesis and the trade converge because the mechanics of the security are designed to make them converge.

Shorts do not get this luxury. On the short side a correct thesis is a necessary but nowhere near sufficient condition for making money. You can be completely right about the business and still lose. You can be right about the direction and still lose. You can be right about the direction AND the magnitude and still lose, if the timing is wrong or the path is too volatile. The thesis is one input into the trade, and the trade has several other inputs that are doing independent work, and any of those other inputs going against you is enough to produce a losing position.

The classic examples make this vivid. Ackman was directionally correct on Herbalife in the sense that the business was doing what he said it was doing. The trade lost him roughly a billion dollars because the path and the clock did not cooperate. Einhorn was right on the accounting at several companies that nonetheless rallied for years before any of his thesis points were reflected in the price. Chanos was right on Tesla as a business analysis for most of the last decade and it is one of the most expensive correct theses in the history of short selling.

The reverse is also true, and worth noticing. You can have a mediocre or even wrong thesis and make money on a short if the trade itself is structured correctly. A stock that is merely fully valued, held by weak hands, entering a window of negative flows, with a near-term earnings catalyst that is likely to disappoint, can be a profitable short even if the underlying business is fine. The trade worked. The thesis was secondary.

This is what professional short sellers mean when they talk about “good thesis, bad trade.” They are pointing at the gap between having correctly diagnosed the company and having correctly structured a position to profit from that diagnosis. The gap is real, it is large, and it is where most amateur short sellers lose money. They do the work on the thesis, they get the thesis right, they short the stock, and they discover that being right about the company was only a small part of what the position required.

The long investor’s instinct is to treat “I am right about the business” as dispositive. On the long side, that instinct is mostly correct, because the mechanics of equity ownership will eventually reward a correct business view. On the short side, that instinct is the single most reliable way to lose money while being right.

Price-to-Value Convergence

The path dependency asymmetry has a deeper root, which is that price-to-value convergence is mechanically underwritten on the long side and entirely absent on the short side. Most investors internalize the idea that markets eventually get it right, but rarely stop to ask what actually makes that happen. When you examine the mechanisms, they all run in one direction.

On the long side, the forcing function is simply that the cash the business generates belongs to the shareholders. Whether it comes back through dividends, through buybacks, or accumulates on the balance sheet earning interest, every dollar of free cash flow is economically owned by the equity. This is not a market convention, it is the legal structure of the corporate form. Over a long enough horizon, the stock cannot trade meaningfully below the present value of the cash it will return to its owners, because if it did, a private buyer, an activist, or management itself would close the gap. The business is continuously producing value that accrues to holders, and that accrual is what pulls price toward value over time. You do not have to convince anyone of anything. You just have to own the claim on the cash.

Now apply the same question to a short. What forces a stock that is worth nothing to trade at nothing? The answer is that there is no mechanism. There is no dis-dividend that extracts cash from shareholders and hands it to shorts, no dis-buyback that dilutes holders by issuing free shares, no compounding process that erodes book value on a schedule. The only way an overvalued stock goes down is if participants decide to sell or stop buying at the current price, and beliefs can remain unupdated for years. Worse, the long-side mechanisms actively work against shorts. Buybacks remove shares from the borrow pool and push the price up. Dividends get paid through. Insider buying moves positioning against you. Every force that guarantees long convergence is a short-side headwind.

So longs are mechanically underwritten by the corporate form, while shorts are underwritten only by other participants changing their minds in a specific direction within a specific window. One is a structural bet, the other is a behavioral bet.

The same asymmetry shows up in what you do when price and value diverge. Both sides are ultimately paid in price, not value, because you cannot deposit a DCF into your brokerage account. But on the long side, if price moves against you while value holds, you can buy more at a better basis and actually monetize the gap when convergence eventually arrives. Adverse moves become a feature of the strategy. On the short side, the equivalent move is adding to a losing position at a worse basis, which is exactly what blows up short sellers. The mechanic that rewards patience on one side punishes it on the other. Longs have an embedded option that converts time and volatility into profit. Shorts have an anti-option that converts the same two inputs into ruin.

How Long vs. Short Investors Enter a Trade

Everything above forces a different entry process on each side, and the difference is sharper than most investors realize. What a good short looks like and what a good long looks like are almost unrecognizable to each other as the same activity.

A short investor cannot enter on thesis alone. Because the position is path-dependent, carries negative, and has no corporate forcing function pulling price toward value, the entry has to be underwritten to a specific risk-reward skew over a specific window. In practice this means three things. First, a target price, because the bounded upside of a short (capped at 100%) has to be weighed against a specifically modeled downside, and if the stock has more room to rally than it does to fall over the relevant horizon, the trade is bad regardless of how good the thesis is. Second, a target date, because every day the position exists is a day of paying borrow, paying dividends through, and bearing squeeze risk, so the thesis has to resolve inside a window short enough for the expected return to clear those costs. Third, catalysts, because absent a forcing function the only thing that moves price toward value is a specific event that makes other participants update their beliefs. Earnings, guide-downs, covenant breaches, regulatory actions, accounting restatements, and insider selling all qualify. Absent a catalyst, you are paying carry to wait for someone else to notice what you noticed, with no guarantee they ever will.

This is why a good short investor can maintain a watchlist of forty solid short theses and not have a single one on. The thesis being correct is a prerequisite, not a trigger. The trigger is the stock reaching a price and setup where the near-term skew is favorable. A business you believe is worth $10 trading at $25 might be an obvious short thesis and a terrible short trade, and the same name at $40 with a looming earnings catalyst might be the trade of the year. Nothing about the business has changed. What has changed is the path forward. Shorts are entered on price, on timing, and on the specific shape of the risk-reward from here, which is why the activity feels surgical even when the underlying view is broad.

A long entry is almost the opposite. It is bigger in scope and less precise in execution. The long investor is looking for a large market opportunity paired with a durable competitive advantage and a price that implies a sufficient margin of safety. Timing does not need to be specified because the forcing function described in the previous section is doing the work. Catalysts are not required because the business is compounding on its own schedule, and every quarter that passes makes the intrinsic value larger whether the market notices or not.

What replaces timing and catalysts on the long side is the size of the expected return. A long-only investor focused on multibaggers will typically not take a position unless the expected return is something like a hundred percent or more, because the abstract margin of safety in the value is what absorbs all the path risk, the timing uncertainty, and the compounding errors in the thesis. If you are right about a triple in five years, you do not need to be right about which year, which quarter, or which catalyst delivers the move. The size of the prize carries the imprecision of everything else. If you are wrong about the triple but still right about the doubling, you are fine. If you are wrong about the doubling but still right about the business compounding, you are fine. The margin of safety absorbs the error.

Shorts do not have this luxury. A short cannot be half right on the path and still make money, because the path is the trade. So shorts demand surgical entries, specific catalysts, and calibrated risk-reward, while longs demand scale of opportunity and durability of moat. Both are disciplined, but the discipline runs along completely different axes. One is precision in price and time. The other is magnitude in value and moat.

Paired vs. Unpaired Shorts

Once you accept that shorts are bets on price within a window rather than bets on value over a horizon, a useful taxonomy emerges. Not all shorts are doing the same thing. Some are taking standalone bets on a specific name going down, and others are using the short as one leg of a larger structure. The distinction matters because the entry rules, the risk-reward, and the bar for “good trade” are all different across the two.

An unpaired short is what most people picture when they hear the word. You short a single name on its own merits, hoping the stock goes down in absolute terms. Your P&L is the stock price decline net of borrow and dividends paid through. Everything in the previous sections applies in its purest form. You need a specific target price, a specific window, a catalyst, and a setup where the near-term risk-reward is genuinely skewed in your favor. The position is fully exposed to market beta, to sector beta, to factor moves, and to the reality that you can be right on the company and still lose money because the market rallied around the position. A short on a stock in isolation makes money if the stock falls. It loses money if the broader market rallies hard enough to lift the stock with it, even if the underlying thesis is playing out quarter by quarter.

A paired short is structurally different. The short is one leg of a relative trade, where you are simultaneously long another name in the same sector, factor bucket, or thematic basket. The point of the structure is to neutralize everything you do not have an opinion on so that the only exposure left is the variant perception you actually believe in. If you are long Lumentum and short Coherent, you are not betting on optics going down or up. You are betting that Lumentum outperforms Coherent. The sector beta cancels out. The factor exposures largely cancel out. What remains is your specific view on the relative quality of the two businesses, the relative defensibility of their positions in CPO, or whatever the variant perception happens to be.

The reason this matters is that paired shorts dramatically relax the surgical entry requirements that unpaired shorts demand. You no longer need the name to fall in absolute terms. You only need it to fall relative to the long leg. Path dependency is partially neutralized because both legs are riding the same market and sector waves. Timing pressure is reduced because borrow costs on a paired trade are partially offset by the carry on the long leg, and the position can sit through broad rallies and selloffs without forcing a cover. Catalysts can be longer-dated because the structure is not bleeding you out the way an unpaired short does. The thesis can be more about durable structural divergence than about a near-term price move.

The technical concept that formalizes all of this is basis risk, which is the risk that your longs and your shorts are not correlated in the way the structure assumes. Gavin Baker described it cleanly: if you are running a high-gross book, the single most important thing to manage is the correlation between your long and short legs. If you are long growth and short value, you are not running a hedge fund, you are running levered growth with extra steps and extra borrow cost. Same in reverse. The whole point of pairing is to generate long-short spread between names that are quantitatively and fundamentally similar enough that the factor exposures cancel, leaving only the alpha you actually have a view on. Pair a great semicap name against a weak semicap name and you have a real trade. Pair a great semicap name against a weak consumer staple and you do not have a hedge, you have two uncorrelated bets sharing a P&L statement.

This is also what makes high gross exposure possible at all. A book with low basis risk can be levered up because the gross exposure is not actually carrying the factor risk it appears to be carrying on paper. A book with high basis risk, levered up to the same gross, is just a directional bet wearing a hedge fund costume, and it will eventually get exposed by a factor rotation that moves the longs and shorts in opposite directions to the structure. The Sharpe ratio that hedge funds are paid to deliver comes from the spread between fundamentally similar names, not from the gross number itself.

This explains why long-short funds run their books the way they do, and why a pure short seller looks like a different animal. A long-short fund is mostly running paired or basket trades where the gross exposure is high but the net is moderate, and the discipline is in matching the long and short legs tightly enough that basis risk stays contained. A dedicated short seller does not have that luxury and has to take unpaired bets, which is why the dedicated short profession is so small, so demanding, and so difficult to do well over a full cycle.

The taxonomy has direct implications for what counts as a good entry on each side. An unpaired short demands the surgical setup described earlier, with target price, target date, and catalyst. A paired short demands a different question entirely, which is whether the relative valuation, relative quality, or relative trajectory of the two names is mispriced, and whether the pairing is tight enough on factor exposure to actually isolate that variant perception. The paired short can be entered on much weaker absolute setups because the structure is doing the work that the entry would otherwise have to do. What you give up is the convexity of an outright winner. A paired short caps your upside at the spread between the two legs, which is usually much smaller than the absolute move on a clean unpaired bet that works.

So when a short investor says they are short something, the next question worth asking is what it is paired against, if anything, and how tight the basis is between the two legs. The same nominal position can be a high-conviction surgical bet, a sector-neutral relative trade, or a factor hedge inside a basket, and the appropriate analysis changes completely depending on which one it is.

My Thoughts on AI Disruption Shorts

So you might have read my recent articles on AI disruption shorts from AI with Jason. Turns out I am actually not short any of them right now. Or well I was short but decided to exit. Doesn’t mean I don’t like the thesis — I just don’t like the trade.

The reason traces directly to everything in the previous sections. A good thesis is a prerequisite for a short, not a trigger. The trigger is the stock reaching a price and setup where the near-term risk-reward is favorable, and right now none of these names clear that bar.

Software has already compressed. The names trade at multiples that reflect a meaningful amount of the disruption thesis. HUBS, MNDY, ASAN are all well off their highs and trade at valuations where the bar to surprise to the downside is much higher than it was eighteen months ago. The pure path-dependent question of “how much room is there to fall versus rally from here” no longer skews in my favor on most of these.

Staffing and labor pass-through is in a cyclical trough. RHI revenue went from $6.4B in 2023 to $5.5B TTM. The comps over the next few quarters get progressively easier. The cycle bull case is that hiring resumes, comps flip positive, and the stock rallies on what looks like a recovery, even if the structural AI thesis is grinding away in the background. Shorting into easy comps with depressed valuations and a potential cycle turn is exactly the setup where the path can kill you for a year or two before the destination eventually arrives. The thesis can be right and the trade still loses meaningfully.

The deeper problem is that AI adoption, while genuinely fast in absolute terms, is still too small a share of revenue at any of the target companies to move near-term numbers in a way the market will price as a structural break. Cognition is replacing some functions at some clients, but not yet at a scale that shows up cleanly in next quarter’s guide. Without that kind of catalyst, the disruption thesis stays in the slow-moving structural bucket, and slow-moving structural shorts are the worst kind. They have all the carry cost and squeeze risk of a regular short with none of the near-term resolution that justifies the position.

So the entire basket sits on the watchlist, fully developed and ready to deploy, and not a dollar of capital is committed to any of it. This is the discipline that the good short investor I talked to embodies. He keeps forty live theses on the page and is short almost none of them at any given time. The thesis work is permanent infrastructure. The trades come and go based on what the price action gives you.

What would change my mind is a macro environment where the index’s risk-reward inverts. Right now the broader market still has more upside than downside on most reasonable forward looks, which means individual shorts are fighting a tailwind on top of all the other structural headwinds discussed in earlier sections. In an economic boom that pushes valuations to a point where the macro skew flips, where downside on the index becomes more likely than upside, the entire calculus changes. At that point staffing into the late-cycle peak with hiring rolling over, software priced for permanence in a multiple compression environment, and consumer intermediation in a discretionary slowdown all become trades where the path is finally working with the thesis instead of against it. The basket gets deployed not because any single thesis got better, but because the price and the macro setup finally caught up to the work.

This is the right mental model for the basket. The research was not wasted by not putting on the trades. It was the prerequisite for being able to act quickly when the entries actually appear. The mistake would have been doing the work, getting attached to the thesis, and shorting into a setup where the path was going to punish me regardless of how correct the destination ended up being.

Why I Am a (Mostly) Long-Only Investor

Two simple reasons

First, my disposition skews long. I’m a big picture, first-principles, long-term thinker and not a detail oriented, analytical, short-term one. This fits longs far better than shorts given our framework.

Second, the economics of shorting for retail investors don’t make sense. You cannot pay your bills with Sharpe ratios. Being correlated to the index is fine. In fact for someone trying to grow capital meaningfully from a small base, correlation to the index is closer to a feature than a bug, because the index has a structural upward drift that you are trying to ride and outperform, not neutralize. The whole point of running a hedged book is to deliver risk-adjusted returns to investors who care about volatility. If you are managing your own money and the goal is absolute compounding, the volatility of being net long is the price you pay for the returns you are after, and trying to hedge it away is just throwing returns away.

Third, the payoff of an hour of analysis on a long is much better than for a short. Twenty hours of work on a short might, if everything goes right, deliver a 30 percent return on the position over a six-to-twelve-month window, and most of that return comes back to you in price action that is hard to capture cleanly because of the path issues already discussed. The same twenty hours spent on a long can identify a multibagger that returns 100 to 300 percent over a few years, with no carry, no squeeze risk, no forced cover, and the option to add to it if the price moves against you in the meantime. The dollar return per hour of analytical effort is not close. And because the long return is uncapped while the short return is capped at 100 percent, the upside distribution is fundamentally different. Spending the scarce resource of research time on the side with the worse risk-reward, the worse path properties, and the lower per-hour expected return is not a defensible allocation of effort for someone in my position.

In conclusion, it is my opinion (not financial advice) that retail investors should (usually) never short.

This is a much shorter article than usual, I will try to write these to respond to catalysts quickly. Below we discuss why this relatively modest firm-level revision is actually a bigger deal for opto than I initially thought (and why I think it’s net new info) as well as a first peek at my top-down MOCVD demand model that will underpin some of my future Aixtron articles. Plus total tool capacity numbers and demand vs supply constrained.

Enjoy!

Read more

Welcome to The Grid.

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

Contents

1. Types of Generation

2. The Merit Order

3. Transmission

4. Interconnection

5. Capacity vs Spot

6. PJM

7. ERCOT

8. IPP Stocks

9. Problems With the Grid

10. Two Solutions to Overflow Demand (with Stock Picks)

Types of Generation

There are a few ways to make electrons.

Nuclear splits atoms to generate heat, which creates steam, which spins a turbine. The fuel is cheap, the output is enormous, and the plants run 24/7 for 18-24 months between refueling. The catch: they cost $10-15 billion to build and take over a decade to permit and construct. Nobody is building new ones at scale in the US. Vogtle Unit 4, the last completed plant, came online in 2024 after years of delays and billions in overruns.

Coal burns carbon to make steam. Same turbine, same idea, much cheaper to build than nuclear. But coal plants are retiring across the country due to emissions regulations and unfavorable economics against natural gas. PJM lost significant coal capacity over the past decade, and those retirements are part of the supply squeeze we’ll cover later.

Natural gas is the workhorse of the modern grid and comes in three flavors. Combined-cycle gas turbines (CCGTs) burn gas to spin a turbine, then capture the waste heat to spin a second one. They run at 55-65% efficiency and serve as the default new-build for baseload. Simple-cycle turbines burn gas once, run at 30-40% efficiency, and exist for speed: they can ramp from cold to full output in 10-20 minutes. Peakers are the extreme version. They sit idle for thousands of hours per year and only fire when the grid is desperate.

Solar converts photons directly to electricity with no moving parts. Zero marginal cost when the sun shines. The problem is obvious: it doesn’t shine at night, and output varies with clouds and seasons. Wind uses air currents to spin turbines. Similar economics (zero fuel cost, variable output) but a different intermittency profile.

Hydroelectric uses falling water. Reliable, dispatchable, zero fuel cost. Also almost fully built out in the US. There are no new rivers to dam.

The key insight: each generation type has a fundamentally different cost structure. Nuclear and coal have high fixed costs and near-zero marginal costs. Gas has moderate fixed costs and fuel-dependent marginal costs. Renewables have high fixed costs, zero marginal costs, but variable availability. This cost structure determines the dispatch order.

The Merit Order

Power demand fluctuates. Not all generation is needed all the time!

To determine who generates power when, the grid stacks generation sources in order of marginal cost; a concept called the merit order.

At the bottom sits nuclear. Always running, always cheapest per marginal MWh. Grid operators dispatch nuclear first because turning it off costs more than keeping it on.

Next come renewables. When the sun shines or wind blows, their marginal cost is effectively zero. They get dispatched whenever available, pushing other sources off the stack.

Then coal and combined-cycle gas. These form the bulk of daily generation. Coal’s marginal cost is its fuel, which has been consistently more expensive than gas in recent years. CCGTs are the flexible middle of the stack.

At the top: simple-cycle gas and peakers. Expensive per MWh, but they start fast. The grid only calls on them during peak demand. Brutally hot summer afternoons, freezing winter mornings, or unexpected generation failures.

The marginal generator (the last plant dispatched to meet demand) sets the price for the entire market in that hour. When demand is low, a CCGT sets the price at $25-40/MWh. When demand spikes, a peaker sets it at $100+ or even $5,000/MWh in extreme cases.

This dispatch hierarchy is the foundation of every energy market in the country. It explains why the same grid can produce $20 power on a mild Tuesday and $1,800 power during a winter storm.

Transmission

Generation is useless without delivery. The transmission system is the high-voltage backbone that moves electricity from power plants to population centers.

Power plants generate at relatively low voltage (11-35 kV), which gets stepped up to 115-765 kV for long-distance travel. Higher voltage means lower current for the same power, which means less energy lost as heat in the wires. At the other end, substations step voltage back down for local distribution.

Transmission lines lose roughly 5-7% of the energy they carry over long distances. That number matters. A solar farm in West Texas producing cheap power is only useful if there are enough transmission lines to carry it to Dallas or Houston. When lines are full (congested), cheap power in one region can’t reach expensive demand in another.

This creates locational marginal pricing (LMP). The same MWh of electricity can cost $20 in one part of the grid and $200 a hundred miles away, simply because the wires between them are at capacity. LMPs are the real-time manifestation of transmission constraints, and they matter enormously for datacenter siting decisions.

Building new transmission takes 7-12 years in most jurisdictions. Permitting, right-of-way acquisition, environmental review, and construction all stack up. You can’t build enough transmission to serve AI demand on AI demand’s timeline.

Interconnection

If transmission is the highway, interconnection is the on-ramp. Every new generation source or large load (like a datacenter) that wants to connect to the grid must go through an interconnection study process.

The process has three phases in most markets. First, a feasibility study determines whether the grid can physically handle the new connection. Second, a system impact study models how the new connection affects power flows across the wider grid. Third, a facilities study identifies exactly what upgrades (new transformers, switches, line extensions) are needed and what they’ll cost.

The interconnecting party pays for network upgrades. A datacenter wanting 500 MW of grid power might need $200 million in substation and transmission upgrades, with a 4-5 year construction timeline, before a single server can draw power.

PJM’s interconnection queue had over 2,500 projects totaling hundreds of GW waiting as of late 2025. The average time from application to completion exceeded four years. Many projects in the queue are speculative (developers filing applications for sites they may never build), which clogs the process for legitimate projects behind them.

This is the single biggest bottleneck for AI datacenter buildout. A hyperscaler can sign a lease, order GPUs, and break ground on a shell in months. Getting grid power delivered to that site takes years. The disconnect between the speed of AI demand and the speed of grid interconnection is the central tension in the power market today.

Capacity vs Spot Markets

Electricity markets split into two distinct mechanisms.

Spot markets trade actual electricity in real time (or day-ahead). Generators bid to supply MWh at a price, demand is forecasted, and the market clears at the marginal bid. This is the spot market. Prices move with fuel costs, weather, outages, and demand. A generator makes money by producing and selling electrons.

Capacity markets are completely different. They trade the promise to be available. A generator that clears a capacity auction gets paid a fixed amount ($/MW-day) to maintain its ability to produce power when called upon. The generator might sit idle 95% of the year and still collect capacity payments.

The logic behind capacity markets: some power plants only run a few hours per year during extreme peaks. If those plants can’t cover their fixed costs from energy market revenue alone, they’ll shut down. Without them, the grid lacks the reserves needed during heat waves or cold snaps. Capacity payments are supposed to keep those plants in business.

PJM has both spot and capacity markets. ERCOT has only an energy market. This single structural difference drives a massive divergence in outcomes, prices, and incentives. It is the most important distinction in the entire US power landscape for understanding what comes next.

PJM

My opinion on PJM is not very high. If PJM and ERCOT were companies, they would make the perfect pair trade. Well maybe you can pair trade them with the IPP names…

PJM operates the grid across 13 eastern US states, serving 67 million people. It hosts Northern Virginia (the world’s largest datacenter hub), and major hyperscaler campuses from Google, Anthropic, Amazon, and Meta.

If you ever hear about capacity auctions hitting price caps think PJM.

In the 2024/25 delivery period, capacity cost $29/MW-day. In the 2025/26 delivery period, it jumped to $270/MW-day. That is a 9.3X INCREASE in a single year. Some locations hit $450/MW-day. The subsequent 2026/27 and 2027/28 auctions both cleared at the federal price cap of $329/MW-day, because regulators imposed a ceiling after the initial shock.

Total capacity payments: roughly $16 billion annually, or about $120,000 per MW. For the average PJM household consuming 880 kWh/month, that translates to approximately $30 per month in additional costs.

Now obviously this is unsustainable and became a reason people blamed data centers for increased power prices.

The simulation problem

Capacity prices in PJM aren’t set by market participants bidding against each other in the traditional sense. They’re driven by a simulated supply-demand curve called the Variable Resource Requirement (VRR) curve. PJM builds this curve from its own internal demand and supply forecasts, using non-public models and proprietary data. The VRR curve’s shape near the clearing point determines whether capacity costs $29 or $270.

So intuitively speaking, in normal markets price is set by the market clearing of supply and demand. In PJM land they decide the price for you. Any solid scholar of economic history can probably guess how this turned out…

PJM cannot forecast datacenter load. Their own data proves this. In 2024, they cut their datacenter load forecast by 800 MW versus the prior year. In 2025, they cut it again by 1.1 GW. Two consecutive years of massive downward revisions from a central planner responsible for a $16 billion annual market.

The simulation itself converts a moderate demand increase into a 9.3x price explosion. They drastically over-forecast how much load hits the market.

Forward traders also disagree with the energy benchmark increased 12-20% in the 2028-2030 window, nothing close to 9.3x.

ERCOT

ERCOT harnesses THE POWER OF MARKETS!

The Electric Reliability Council of Texas (ERCOT) runs the Texas grid. Same AI buildout (OpenAI, Google DeepMind, Anthropic are all building massive facilities). Completely different market structure.

ERCOT has no capacity auction. No BRA. No VRR curve. No central planner simulating supply and demand to set capacity payments. Instead, it uses an energy-only market with real-time scarcity pricing via an Operating Reserve Demand Curve (ORDC).

When supply-demand balance gets tight, real-time prices spike from the normal $10-50/MWh to as high as $5,000/MWh, with additional transmission congestion adders. Gas peakers and batteries that run fewer than 100 hours per year can still pay for themselves because those hours are worth millions of dollars.

The incentive structure is fundamentally different. In PJM, generators get paid whether they perform or not. In ERCOT, generators only make real money when the grid actually needs them and they deliver.

Stable prices despite massive demand growth

ERCOT’s 2025 Long-Term Load Forecast projected 77.9 GW of potential datacenter load by 2030, more than double the prior year’s projection. However, ERCOT’s demand forecasts do not directly drive prices.

Forward prices in ERCOT have only increased 11-17% in the past year!

Winter Storm Fern

The January 2026 freeze was the real-world stress test.

ERCOT’s grid held. The Weather Watch remained precautionary. Demand ran below forecasts. No emergency procedures triggered. Real-time prices peaked around $300/MWh. Post-Uri winterization reforms proved effective under real conditions.

PJM paid $270/MW-day in capacity costs and lost 21 GW of generation. Fifteen percent of the cleared fleet went offline due to frozen equipment and fuel delivery failures. The Department of Energy issued emergency orders under Section 202(c) of the Federal Power Act, authorizing bypass of environmental limits and access to roughly 35 GW of backup generation (including datacenter and industrial backup power that would have been ineligible for the BRA). System-wide average prices hit $700/MWh. Virginia’s datacenter-heavy Dominion zone spiked to $1,800/MWh.

Lower cost, better outcome. ERCOT’s market discipline delivered what PJM’s $16 billion in capacity payments could not.

Below the paywall we discuss a couple of popular IPP (power producer) stocks, why it was a popular trade in 2024, and why they’ve stopped working recently. I made a comps table for all the IPPs on their different types of capacity (nuclear, CCGT, peaker, solar, etc.), converted them into CCGT-equivalents to show who is the relative cheapest. Then we move on to more intriguing and granular discussion about the problems with the grid when it comes to powering AI. Finally, we end with the real deal. Two solutions that actually solve the grid’s constraints, essentially “overflow valves” that are a call option on AI energy demand and the stocks I actually own for each.

IPP Stocks

Read more