AI's price is collapsing. Meta is selling spare compute, China gave away a frontier model Kimi, and chip stocks fell 18%. Five takeaways for investors.
Good article. My challenge is over 75% of all the compute is driven by arthropic and open AI. Plus you got all the circular deals between them and the hyperscalers and neo clouds . Enterprise barely constitutes 10%. For this to work that has to increase fast and be measurable and compute has to be cheaper and more effective - that is debatable . When capital dries up everything will collapse and will have empty data centers and no need for electricity.
I’d add optics alongside electricity if you’re in the camp that believes the data center buildout is going to happen regardless of who ultimately wins in AI. Faster speeds, higher bandwidth requirements, and increasing traffic across networks should accelerate network infrastructure investment and, therefore, the adoption of increasingly advanced optical technology.
I think I can add some color as to why Meta renting compute would not actually be an admission of overbuilding or excess, and does not break the scarcity story. I’ll try to explain briefly:
Not all compute is created equal. Accelerators are configured in different ways for different use cases, and reconfiguring them is not as simple as flipping a light switch because different configurations have very different requirements. Large clusters of compute (think tens of thousands of accelerators configured to act as a single supercomputer) are used extensively in AI R&D, and every single one available today is currently spoken for, making availability of these clusters without lead times stretching into 2027/2028 extremely hard to find. With this scarcity comes huge premiums, which is why Anthropic and Google are both renting from xAI at rates far above spot price. That premium may temporarily be well above what management calculates the cluster’s internal value to be as a result of near-term scarcity, meaning renting it now at inflated rates is not an admission of excess capacity, but an opportunistic decision to extract additional value from equipment they already own.
Really well written. I’d argue that there’s already been a lot of run up in AI related energy/electricity stocks like Bloom and GEV. Perhaps they’re just unfairly getting punished with this momentum unwind, but let’s see
One of the more interesting takes I've seen on the AI boom.
https://groundbreakerre.substack.com/p/the-second-derivative-why-no-one?utm_source=share&utm_medium=android&r=4p3x50
Will have a look!
Good article. My challenge is over 75% of all the compute is driven by arthropic and open AI. Plus you got all the circular deals between them and the hyperscalers and neo clouds . Enterprise barely constitutes 10%. For this to work that has to increase fast and be measurable and compute has to be cheaper and more effective - that is debatable . When capital dries up everything will collapse and will have empty data centers and no need for electricity.
Thanks! And thats some good thoughts
I’d add optics alongside electricity if you’re in the camp that believes the data center buildout is going to happen regardless of who ultimately wins in AI. Faster speeds, higher bandwidth requirements, and increasing traffic across networks should accelerate network infrastructure investment and, therefore, the adoption of increasingly advanced optical technology.
Good idea!
I think I can add some color as to why Meta renting compute would not actually be an admission of overbuilding or excess, and does not break the scarcity story. I’ll try to explain briefly:
Not all compute is created equal. Accelerators are configured in different ways for different use cases, and reconfiguring them is not as simple as flipping a light switch because different configurations have very different requirements. Large clusters of compute (think tens of thousands of accelerators configured to act as a single supercomputer) are used extensively in AI R&D, and every single one available today is currently spoken for, making availability of these clusters without lead times stretching into 2027/2028 extremely hard to find. With this scarcity comes huge premiums, which is why Anthropic and Google are both renting from xAI at rates far above spot price. That premium may temporarily be well above what management calculates the cluster’s internal value to be as a result of near-term scarcity, meaning renting it now at inflated rates is not an admission of excess capacity, but an opportunistic decision to extract additional value from equipment they already own.
Thank you Melvin!
A nice balanced take. Very well written article!!
Thanks a lot Faisal!
this is the reason I own meta and tencent. huge pool of users.
Let that sink in: Tech companies are spending more on electricity infrastructure than semiconductors!
Really well written. I’d argue that there’s already been a lot of run up in AI related energy/electricity stocks like Bloom and GEV. Perhaps they’re just unfairly getting punished with this momentum unwind, but let’s see