Thank you for sharing your insights, this is a fantastic breakdown in my view — especially the discussion around yields, tool spend and margin impact. I recently looked at Intel by connecting how margin recovery, capex intensity and free cash flow interact across these phases, and it’s striking how different the picture looks when you follow the cash mechanics alongside the manufacturing narrative.
This is so fascinating, Jason—could you elaborate a bit more on how CPUs are becoming central to compute infrastructure as agentic workflows take center stage? To what extent does this shift mirror the structural transformation we’ve seen in memory, where HBM has moved from a commodity-like market to a high-value-add segment, and do you see CPUs playing a similar “brain” role in commanding the agentic future of AI? Also, on a related note, what about embodied and physical AI—what roles do CPUs play there, and are there untapped market opportunities? I always enjoy your insights, and I find it hilarious that you’re invested in Intel!
This is a great question. I think it comes down to how the CPU/GPU workload ratio in datacenters evolves over time. With chatbots, its mostly GPU, as there is not much compute spent other than highly parallelizable forward pass (model inference). For agentic workflows, there is now more of a mix between parallelizable forward pass and complex branching operations like spawning sandboxes, running compilers, executing tests, making API calls, managing state--the "tools" Claude Code uses that is beyond just spitting out tokens. This is all very CPU heavy, so that ratio changes in the favor of CPUs. I have not looked deeply into robotics but I would assume this trend continues. The neural network deciding how a robot should move is GPU-accelerated, but the real-time control loops, sensor fusion, safety systems, and actuator management are CPU-bound.
Thanks. After more reflection and reading I honestly think its a bit more nuanced than just yields, they have no die re-use and fungibility for chiplets unlike AMD which means they could truly be caught flat footed in a server shortage even if they have client. And there are still some fixed cost in COGS. Your posts are hilarious btw
Thank you for sharing your insights, this is a fantastic breakdown in my view — especially the discussion around yields, tool spend and margin impact. I recently looked at Intel by connecting how margin recovery, capex intensity and free cash flow interact across these phases, and it’s striking how different the picture looks when you follow the cash mechanics alongside the manufacturing narrative.
This is so fascinating, Jason—could you elaborate a bit more on how CPUs are becoming central to compute infrastructure as agentic workflows take center stage? To what extent does this shift mirror the structural transformation we’ve seen in memory, where HBM has moved from a commodity-like market to a high-value-add segment, and do you see CPUs playing a similar “brain” role in commanding the agentic future of AI? Also, on a related note, what about embodied and physical AI—what roles do CPUs play there, and are there untapped market opportunities? I always enjoy your insights, and I find it hilarious that you’re invested in Intel!
This is a great question. I think it comes down to how the CPU/GPU workload ratio in datacenters evolves over time. With chatbots, its mostly GPU, as there is not much compute spent other than highly parallelizable forward pass (model inference). For agentic workflows, there is now more of a mix between parallelizable forward pass and complex branching operations like spawning sandboxes, running compilers, executing tests, making API calls, managing state--the "tools" Claude Code uses that is beyond just spitting out tokens. This is all very CPU heavy, so that ratio changes in the favor of CPUs. I have not looked deeply into robotics but I would assume this trend continues. The neural network deciding how a robot should move is GPU-accelerated, but the real-time control loops, sensor fusion, safety systems, and actuator management are CPU-bound.
You are the GOAT man. I will do some further digging into this topic!
https://open.substack.com/pub/techitalt/p/intel-is-crashing-and-burning-as?r=5jmutn&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Yeah thats generally my take too. Though I would add I still think the upside from 14A in 28/29 is enormous.
Nice post. I thoroughly enjoy your writing style. Very readable. Looking forward to more!
Thanks, Vik 😁 this means a lot coming from you
Thanks. After more reflection and reading I honestly think its a bit more nuanced than just yields, they have no die re-use and fungibility for chiplets unlike AMD which means they could truly be caught flat footed in a server shortage even if they have client. And there are still some fixed cost in COGS. Your posts are hilarious btw