"Some investors are betting that the artificial-intelligence revolution of the 21st century will give rise to a huge, actively traded market for computing power, the same way that global energy demand led to the emergence of the crude-oil market last century.
Compute Exchange, a startup that announced its launch on Wednesday, plans to conduct auctions where data-center operators and cloud providers can sell their computing capacity -- or "compute" in tech-industry jargon. Buyers such as AI companies can use the auctions to buy chunks of processing time for arrays of chips, such as Nvidia's H100 and H200 systems.
The trading platform's co-founder and best-known investor is Donald Wilson Jr., founder of high-frequency trading firm DRW Holdings and a veteran Chicago trader.
"The total dollars spent on compute will, over the next 10 years, exceed total dollars spent on oil," Wilson predicted. "Obviously, oil is the largest commodity right now, so I believe it will be displaced by compute."
Compute Exchange plans to hold its first public auction Feb. 25, after quietly holding two test auctions late last year. About a dozen sellers are expected to take part, according to the Palo Alto, Calif.-based startup.
The debut comes just after the tech industry was roiled by China's DeepSeek. The Chinese startup triggered a huge selloff in AI-linked stocks when it unveiled a sophisticated model that it trained using fewer, less-advanced chips than its Silicon Valley rivals and at a fraction of the cost.
There are obstacles to the emergence of a traded market for computing power. One challenge is that the largest sellers of compute -- the hyperscalers such as Amazon Web Services, Google Cloud and Microsoft Azure -- have so far stayed away from startup trading platforms such as Compute Exchange, preferring to sign big-ticket deals. Similarly, huge buyers such as OpenAI tend to secure massive amounts of computing capacity, rather than buying small portions on the spot market.
Skeptics also question whether computing power can truly be considered a commodity. Compute comes in a diverse array of underlying chips and configurations, without a commonly agreed upon standard. Typically, AI companies can't just switch from one provider to another to get a lower price, unless they are willing to put up with technical headaches.
"Compute isn't necessarily a fungible resource," said Meltem Demirors, founder of early-stage investment fund Crucible Capital. "A compute-hour in one facility with one type of chip is not equal to a compute-hour in another facility with another type of chip."
Besides Compute Exchange, Wilson has invested in Silicon Data, a startup that researches the market for computing power. Silicon Data aims to publish indexes that could become the basis for futures contracts tied to the price of compute. In theory, data-center operators could use such futures to hedge against the risk of a drop in prices, the same way energy companies trade futures to protect against a potential crash in the oil market.
Others also have sought to develop organized markets for computing power. Perhaps the best known is the San Francisco Compute Company, or SF Compute for short. Founded in 2023, the company says it has sold more than $50 million of compute contracts over the past year. Its investors include the venture-capital firm of Jack Altman, brother of OpenAI CEO Sam Altman.
The founders of SF Compute were initially seeking to train a model for an AI audio startup when they discovered it was difficult to get cloud providers to sell them computing capacity in the quantities they needed.
"They couldn't sell you an hour at a time, or even a week at a time," said Evan Conrad, a co-founder of SF Compute. "You had to buy for a year or longer."
SF Compute's founders pivoted to creating a marketplace for computing power. Unlike Compute Exchange, SF Compute doesn't run intermittent auctions. Instead, it runs a continuous market where AI companies can buy computing power as they need it, for times as short as one hour." [1]
1. Startup Envisions Market for 'Compute' Power. Osipovich, Alexander. Wall Street Journal, Eastern edition; New York, N.Y.. 30 Jan 2025: B4.
Komentarų nėra:
Rašyti komentarą