And then we package it, throw a pretty wrapper around the package—perhaps a posh frock, create derivatives and hoo boy, there is money to be made!
Just kidding. While Collateralized Compute Obligations are almost alliterative, a major selling point, that's not where this story is going. Yet.
From Bloomberg, September 26:
The AI Boom Needs a Market for Compute
Just as oil futures and spectrum auctions unlocked waves of investment, turning compute into a tradable commodity will be needed to fuel AI’s next stage.
There’s money — really big money — in selling compute, the processing power driving the AI revolution.
Hyperscaler CoreWeave Inc. is worth more than $50 billion even as its losses are projected to increase tenfold this year, to $650 million. Oracle Corp., meanwhile, recently added some $250 billion to its market capitalization overnight after revealing the size of its upcoming order book. What’s unclear is whether those valuations are based on a scarcity thesis — that demand is likely to outstrip supply for years to come — or an abundance thesis, foreseeing a multitrillion-dollar market where these companies will be among the many winners.
In a world where the promise of AI is fulfilled and today’s valuations for companies like OpenAI are justified, compute providers will be more akin to Walmart, with its high-volume, low-margin stores, than a high-margin, medium-volume business like LVMH. That world aligns roughly with McKinsey & Co.’s bull-case scenario of $7.9 trillion in data center investment over the next five years, and would also be good news for the industry: Walmart Inc. after all, is a lot more profitable than LVMH Moët Hennessy Louis Vuitton SE.
Even McKinsey’s bear case of $3.7 trillion, however, implies a magnitude of capital flows with little precedent in human history. And if that’s going to happen, one necessary precondition is that compute needs to be priced a lot more transparently.
In the clunky current system, the chief operating officer of an AI company calls up Amazon Web Services Inc. or CoreWeave and asks for a price. That’s a major bottleneck, since it’s difficult for consumers of compute to shop around, especially if there are many different configurations they’d be happy with at the right price. And while vendors enjoy the extra pricing power the status quo gives them, they’d also love to be able to compete effortlessly on every contract being negotiated anywhere in the world. That means sooner or later we’re going to have to have a public market for compute, much as we do for all other valuable commodities.
The good news is that transitioning to a market-based system has happened many times before, generally with great success. The oil and gas industry, for instance, struggled with significant capital constraints when it came to exploration up until around 1983, when the introduction of crude oil futures brought a lot more liquidity into the market. Something similar happened in the 1990s, when chunks of the electromagnetic spectrum previously allocated by governments started being auctioned, priced and invested in.
The godfather of spectrum auctions is Paul Milgrom, the Stanford economist who won the Nobel Prize in 2020 for his work on auction design. Now, Milgrom and his company, Auctionomics Inc., have teamed up with OneChronos Markets, a company that builds smart exchanges, to do for compute what Milgrom did for spectrum.
With AI technology, that’s now possible. As a rule, the simpler and more fungible the commodity, the easier it is to trade. Grain futures date back almost 4,000 years, to the Mesopotamian Code of Hammurabi; oil futures, by contrast, required the creation of standards like Brent Crude and West Texas Intermediate. Compute markets — both spot and futures — involve so many variables that they’ve become a possibility only with the advent of AI agents that can accept natural-language inputs. Those agents then turn high-level English-language goals and instructions into detailed structured “expressive bidding” that can be executed on-exchange.
The trick is to use what’s known as combinatorial auctions. Bidders might care very much about having an uninterrupted stretch of time to do their computations, for instance, or they might be willing to break it up if someone else wanted the data center urgently and briefly. Similarly, a bidder might ideally want to use B200 chips, but would be fine with paying slightly less money to use more H100 chips for more hours. Right now the market in compute is almost entirely bespoke: A client explains what they’re looking for, and a vendor gives them a price. But with combinatorial auctions, a large number of buyers and sellers can collectively arrive at prices for the optimal allocation of every chip during every hour of every day.
That in turn maximizes value, not least because it significantly reduces the risk associated with buying compute. “One of the things that can really hurt the value of an asset is liquidity risk, or the inability to resell it,” says OneChronos co-founder and Chief Executive Officer Kelly Littlepage. Once compute spot markets are in place, every buyer will become a potential reseller of the commodity, creating vastly more potential vendors from whom chip time can be bought....
....MUCH MORE
