The mistaken assumption in this analysis is that all compute is created equal. The rental prices and use cases of the same hardware can vary wildly depending on how it is configured. Fragmented compute (i.e., clusters of various sizes, generally in the hundreds to low thousands of GPUs) is a commodity that is rented at market rates (these are the H100s that rent for ~$2.50/GPU-hour you mentioned). Frontier-scale coherent clusters (gigawatt-blocks of tens to hundreds of thousands of GPUs) on the other hand are a different product entirely, and are scarce. Finding unassigned frontier-scale coherent clusters available for rent before the end of 2026 is almost impossible. This is why Anthropic signed a contract to pay xAI ~$1.25B per month for access to Colossus 1 (an implied ~$7-8/GPU-hour, roughly 3x commodity rates). These frontier-scale coherent clusters are the crown jewels Meta is reportedly considering renting out.
But if Meta is selling access to its coherent clusters, doesn't that mean Meta is admitting it overbuilt? Not quite. The benefit of the frontier-scale coherent cluster is also one of its biggest drawbacks: it excels at training frontier LLMs, but it's overkill for most other tasks, and training is only one part of the model lifecycle. Outside of AI R&D, Meta's biggest need for compute comes from its recommendation systems (recsys). The compute requirements for recsys are massive, but the workload will happily run on older GPUs and fragmented clusters. In other words, commodity compute works just fine for Meta's workloads outside of LLM training. Think of it like owning a supercar and a Prius. You absolutely need the supercar if you want to race, but a Prius is a much more sensible daily driver.
So if you're Meta, and you're in between training runs, would you rather run recsys on your coherent cluster (the supercar), or run it on fragmented clusters (the Prius) at commodity rates so you can rent out the coherent cluster at 3x market rates? Obviously the latter. But what if your partner is using your Prius? Then you rent another one. This is where the CoreWeave and Nebius deals come in: Meta can offload recsys workloads onto rented commodity capacity and free up the coherent blocks to rent out at a premium.
Given all of this, I don’t think it likely that Meta slows down their capex. After all, what's better than having a supercar? Having two supercars. Or having a supercar, a moped for your kid to do DoorDash in, and a Prius for yourself. Or buying modifications to make your supercar even faster. The point is, you have a lot more optionality now that you're making extra money renting your supercar out to finance influencers on the weekends (I've heard that's a great business in Miami).
Great article. Don’t think it is fair to say Grantham is a permabear. He’s been long more than he’s been short but he did point out how implausible the everlasting AI boom story was.
The cracks are exposed indeed. You never quite mentioned whether you feel this is a top of the market or a local top. Do you let the charts and macro speak for themselves?
Good stuff Sam, I am right there with you on timing. I have been moving out into long stable dividend stocks and building cash for a pull back which is clearly on its way. With Trump on the clock and obviously going to have to buy his way out of the Iran mistake the market is facing some issues that are clearly not reflected in pricing at this time. We have some tough years ahead. I am very much enjoying your stuff and look forward to more. THX. JW.
Meta is building data centers at an unprecedented scale. Not something you do if you think the market is not going to be there.
1. Hyperion (Louisiana)
Hyperion
Located in Richland Parish, Louisiana.
Initial design around 2 GW of compute capacity.
Long-term expansion target of roughly 5 GW.
Expected to be Meta’s largest AI campus.
Campus size around 4 million square feet.
Power requirements are so large that Entergy is building multiple new power plants and major transmission infrastructure specifically for it.
To put 5 GW in perspective:
A typical large hyperscale data center might be 100–300 MW.
5 GW is roughly 15–50 times larger than many current hyperscale facilities.
2. Prometheus (Ohio)
Prometheus
Located in New Albany, Ohio.
Targeted at about 1 GW.
Scheduled to begin coming online in 2026.
Described by Meta as one of its first gigawatt-scale AI superclusters.
Other Expansion
Meta is also:
Expanding existing campuses, including El Paso, Texas.
Building additional AI clusters beyond Hyperion and Prometheus. Zuckerberg has stated Meta is constructing “several multi-GW clusters” and plans to spend hundreds of billions of dollars on AI infrastructure.
Just got on Substack, great read. I agree we are near the tail. Most of the world is long this sector (smart and dumb money) so not many buyers left if / when there’s a selloff.
Sam, great factual support here. I personally am a bit more bullish since I am 'not in the room where it happened', and I like to believe that these folks will do something to benefit themselves, even and especially at our expense. However, I think this is a very important article and its exactly why I sold ORCL a month ago, and exited many gains and just bought a bunch of NQ protection plays and PHYS.
The mistaken assumption in this analysis is that all compute is created equal. The rental prices and use cases of the same hardware can vary wildly depending on how it is configured. Fragmented compute (i.e., clusters of various sizes, generally in the hundreds to low thousands of GPUs) is a commodity that is rented at market rates (these are the H100s that rent for ~$2.50/GPU-hour you mentioned). Frontier-scale coherent clusters (gigawatt-blocks of tens to hundreds of thousands of GPUs) on the other hand are a different product entirely, and are scarce. Finding unassigned frontier-scale coherent clusters available for rent before the end of 2026 is almost impossible. This is why Anthropic signed a contract to pay xAI ~$1.25B per month for access to Colossus 1 (an implied ~$7-8/GPU-hour, roughly 3x commodity rates). These frontier-scale coherent clusters are the crown jewels Meta is reportedly considering renting out.
But if Meta is selling access to its coherent clusters, doesn't that mean Meta is admitting it overbuilt? Not quite. The benefit of the frontier-scale coherent cluster is also one of its biggest drawbacks: it excels at training frontier LLMs, but it's overkill for most other tasks, and training is only one part of the model lifecycle. Outside of AI R&D, Meta's biggest need for compute comes from its recommendation systems (recsys). The compute requirements for recsys are massive, but the workload will happily run on older GPUs and fragmented clusters. In other words, commodity compute works just fine for Meta's workloads outside of LLM training. Think of it like owning a supercar and a Prius. You absolutely need the supercar if you want to race, but a Prius is a much more sensible daily driver.
So if you're Meta, and you're in between training runs, would you rather run recsys on your coherent cluster (the supercar), or run it on fragmented clusters (the Prius) at commodity rates so you can rent out the coherent cluster at 3x market rates? Obviously the latter. But what if your partner is using your Prius? Then you rent another one. This is where the CoreWeave and Nebius deals come in: Meta can offload recsys workloads onto rented commodity capacity and free up the coherent blocks to rent out at a premium.
Given all of this, I don’t think it likely that Meta slows down their capex. After all, what's better than having a supercar? Having two supercars. Or having a supercar, a moped for your kid to do DoorDash in, and a Prius for yourself. Or buying modifications to make your supercar even faster. The point is, you have a lot more optionality now that you're making extra money renting your supercar out to finance influencers on the weekends (I've heard that's a great business in Miami).
Great article. Don’t think it is fair to say Grantham is a permabear. He’s been long more than he’s been short but he did point out how implausible the everlasting AI boom story was.
Come on, Meta didn't announce anything. It was merely reported. The rest of your article didn't deserve that flawed first statement.
"Excess compute"... OH, BLASPHEMY!
Very good article. I still like Google, out of the bunch. Sure, the price might fall, and will someday, but not as bad as these other companies will.
The cracks are exposed indeed. You never quite mentioned whether you feel this is a top of the market or a local top. Do you let the charts and macro speak for themselves?
Good stuff Sam, I am right there with you on timing. I have been moving out into long stable dividend stocks and building cash for a pull back which is clearly on its way. With Trump on the clock and obviously going to have to buy his way out of the Iran mistake the market is facing some issues that are clearly not reflected in pricing at this time. We have some tough years ahead. I am very much enjoying your stuff and look forward to more. THX. JW.
Excellent screed here, Sam!
Meta is building data centers at an unprecedented scale. Not something you do if you think the market is not going to be there.
1. Hyperion (Louisiana)
Hyperion
Located in Richland Parish, Louisiana.
Initial design around 2 GW of compute capacity.
Long-term expansion target of roughly 5 GW.
Expected to be Meta’s largest AI campus.
Campus size around 4 million square feet.
Power requirements are so large that Entergy is building multiple new power plants and major transmission infrastructure specifically for it.
To put 5 GW in perspective:
A typical large hyperscale data center might be 100–300 MW.
5 GW is roughly 15–50 times larger than many current hyperscale facilities.
2. Prometheus (Ohio)
Prometheus
Located in New Albany, Ohio.
Targeted at about 1 GW.
Scheduled to begin coming online in 2026.
Described by Meta as one of its first gigawatt-scale AI superclusters.
Other Expansion
Meta is also:
Expanding existing campuses, including El Paso, Texas.
Building additional AI clusters beyond Hyperion and Prometheus. Zuckerberg has stated Meta is constructing “several multi-GW clusters” and plans to spend hundreds of billions of dollars on AI infrastructure.
And yet, we learn through being wrong.
Better to be wrong than invest in your own execution.
Nice work. Looking forward to reading more of your work.
https://heraldofthehudson.substack.com/p/meta-platforms-inc-meta
Just got on Substack, great read. I agree we are near the tail. Most of the world is long this sector (smart and dumb money) so not many buyers left if / when there’s a selloff.
Your analysis brilliantly highlights the toxic cycle of an AI bubble on the verge of bursting.
Hyperscalers are forecasting $690 billion in Capex for 2026 (+81%), while 95% of AI pilots have no financial impact.
The price per GPU-hour has already plummeted from $8 to $2.99, threatening debt-financed “neo-clouds” with interest rates of 9–15%.
Meta, a former giant client with a $35B commitment to CoreWeave, is now reselling its excess computing capacity, further saturating the market.
Nvidia has propped up its own demand through a closed-loop financing scheme ($30B for OpenAI, $1.5B in leases with Lambda).
This Ponzi scheme came to a screeching halt in March, exposing the sector to a chain of defaults.
Sam, great factual support here. I personally am a bit more bullish since I am 'not in the room where it happened', and I like to believe that these folks will do something to benefit themselves, even and especially at our expense. However, I think this is a very important article and its exactly why I sold ORCL a month ago, and exited many gains and just bought a bunch of NQ protection plays and PHYS.
Getting off too soon is far superior than going over the cliff.