I guess I’m a little confused.
I hear every day that demand for compute’s insatiable in the AI era. That everybody, including and especially the hyper-scalers, needs as much compute as they can possibly access. With semi revenue rising triple-digits and component prices soaring, it’s difficult to argue the point.
And yet, many (most) of the companies insisting on that point also rent access to idle, “excess” compute to outside customers, raising this question: If it’s so damn precious and hard to come by, how do you have any to rent out, let alone so much to spare that you can build an entire business around it?
The answer goes something like this. Demand for compute, just like demand for electricity, isn’t homogenous. Scheduling varies depending on what you’re trying to accomplish (e.g., inference versus training). There are peaks and valleys for utilization, and off-peak, idle capacity can be monetized. Even conservative estimates show internal and external workloads growing exponentially, so failing to add as much capacity as possible now is to risk leaving someone in the lurch (and money on the table) later.
A tech analyst would probably say that’s not a sufficiently trenchant description, but all inside-baseball quibbling aside, that’s the gist of it. Really it is.
Setting aside the obvious pushback — i.e., that this whole enterprise rests on the possibly misguided assumption that demand for AI services will remain insatiable forever, and that the provision of those services will eventually be highly profitable — I’m not convinced that even the most competent management teams are capable of performing the juggling act described above.
On Wednesday, Meta soared nearly 9% — good for ~$130 billion in market cap — on a Bloomberg report that the company’s gearing up to challenge AWS, Azure and Google Cloud with a business that’ll sell access to AI compute and models. The new venture may even step on CoreWeave’s toes by renting so-called “raw” capacity.
On one hand, I’m actually surprised it took Mark Zuckerberg this long to stand up a cloud business. This was inevitable.
On the other hand, it feels desperately circular, consistent with the self-referential nature of what, by now, is the single-most expensive bet in the entire history of capitalism. Read this excerpt from the linked Bloomberg article:
Meta has made developing AI ‘superintelligence’ a top priority, and has committed hundreds of billions of dollars to data centers and other AI infrastructure. That investment has left investors anxious about Meta’s plans to earn a return on that spending. A cloud business offers one way to return some of that investment.
The plan, then, for generating a return on money spent to access compute involves — wait for it — renting access to compute.
And look, I’m not deliberately mischaracterizing this just to elicit reader chuckles on an otherwise slow day. Try as they surely did, Bloomberg’s journalists and copy editors couldn’t escape this black hole of circularity either. Here’s another indicative passage from Riley Griffin and Kurt Wagner:
Tech companies have committed tens of billions of dollars for data center capacity for their own needs [and] those providers have also expanded to rent the computing capacity needed to train and run AI models. It is a complex business.
Yeah, “complex” is one word for it. Critics might suggest “asinine” is at least as apt, particularly considering a lot of these compute deals involve the same 10 or so companies swapping dollars and renting each other the “same” compute.
It looks, in short, like a giant exercise in rehypothecation. When I asked ChatGPT to refute that characterization, “she” demurred, suggesting I may not fully understand all the critical nuance of the business.
Of course, a CDO structurer would’ve told me the same thing about subprime in 2006.
“Of course, a CDO structurer would’ve told me the same thing about subprime in 2006.”
I’d send a giant emoji smile on that one. Sometime around then a partner’s account went to visit LTCM out in California. When the client (a successful hedge fund manager) started asking too many sharp questions, he was told that he was not smart enough to understand what they were doing. My recollection that the speaker was Myron Scholes himself.
I’m no tech guy, but it was my understanding that Facebook’s (Meta) early leap into computing was to create the metascape where people could disappear from the world with their newly purchased Meta Quest VR Headsets and do a little online shopping in the Meta bazaars. Zuckerberg spent a fortune on that investment and has since scaled back his investment in the Metaverse. Maybe that infrastructure represents the overcapacity that others can repurpose until their data centers are completed?
Dang it, I knew I should’ve waited before commenting about Meta on the other article. Appreciate the explanation though. I can understand the comparison to electricity, but to your broader point, it sure does seem that we’ve seen this story before.
Then again, I sure wish I had the naive optimism needed to go along for the ride on one of these exponential growth stories (outside of my standard investments of course).
Zuckerberg. Wow. Zuck let Yann LeCun, a Turing Award winner, walk away so he could spend $100s of millions hiring lots of hotshot AI scientists to achieve AGI via LLMs (which LeCunn and others say is impossible w/ LLMs). And now we know why Zuck spent so much money. When his hotshots are not busy using all Meta’s compute power, they will sell it like a power company. Kind of like the spots marked for EVs outside gyms.
“It looks, in short, like a giant exercise in rehypothecation. When I asked ChatGPT to refute that characterization, “she” demurred, suggesting I may not fully understand all the critical nuance of the business.”
Don’t fall for that Walt, Chat GPT is obviously in on the whole deal!
On a side note, whenever I look at those AI circular financing diagrams that we have all seen, CoreWeave is usually somewhere near the middle with a lot of arrows (circular flows) pointing to and from it. Maybe the iceberg all the hyper-scalers are standing on is starting to shrink.
It loos to me Walt believes this train will be going quite well for couple more years! All-in?!
I would gently point out that not all compute capacity is created equal. AI workloads require a specific compute layer that demands specific GPU capacity. Whereas traditional compute workloads are dependent on CPU capacity with adjustable storage and memory. So the unused compute may not be able to support AI workloads.
Another interesting conundrum with this is customer’s incessant drive to reduce and manage compute costs. There are entire SaaS products built on helping identify wasted cloud resources and optimizing them to reduce spend.
On the Meta front, I’d also like to point out that this is an interesting strategy considering the most disruptive new entry into this space, railway, is currently building out their own data centers and their IaaC offering is far more accessible and cost effective than their competitors are now. I doubt Meta is going to be able to match what they have already built.
From The Information:
“Nvidia is promising to financially backstop young cloud providers that rent out its graphics processing units in exchange for a share of their revenues. The backstop would come in the form of promising to rent back unused GPUs if the companies can’t find AI developers to rent them.”
https://www.theinformation.com/articles/nvidia-says-will-take-cut-customers-cloud-revenues
SoftBank is establishing SB Neo to operate a U.S. neocloud business.
SB Neo will be 51% owned by SoftBank Corp. and 49% owned by SoftBank Group, with launch planned for FY27, the fiscal year ending March 2028.
The business will use SoftBank’s planned 10GW-scale energy and AI infrastructure to provide AI training and inference compute to U.S. enterprises and hyperscalers.
Throw another hat in the ring.
Softbank doesn’t have the wherewithal to fund it. So watch for some more bond issuance unless he can use some clever off-balance sheet entity.