(Bloomberg) -- A central challenge of the current artificial intelligence boom is finding enough computing power. Getting access to the semiconductors that handle the massive computational loads demanded by AI can make or break a young startup. And often, only the biggest, richest and best-connected companies are able to secure the processing they need.
To help get small players a piece of those resources, a new startup called San Francisco Compute Co. wants to make it easier to buy a chunk of access at a time. The company’s idea is to create a kind of trading platform for coveted compute power.
On Tuesday, the company plans to announce it’s raised $12 million in an early funding round to get its platform off the ground. Already, the company has secured the equivalent of 8,000 H100s, Nvidia’s highly prized AI chip, to jump-start the endeavor.
Tech entrepreneur Jack Altman led the funding round, which valued SF Compute at $70 million. Altman’s firm, Alt Capital, invested alongside a handful of individual AI backers and founders. The goal, Altman said in an interview, is to “allow regular startups to use humongous amounts of compute for a short period of time.”
SF Compute’s founders, Evan Conrad and Alex Gajewski, happened upon the idea for the company last year when they were trying to build an AI music startup. They quickly found that compute power was out of reach.
“If you aren’t one of the holy few, you are effectively priced out of the market,” said Conrad. “There is no option for you without major funding.”
Conrad, who previously worked at OpenAI and at startup accelerator AI Grant, met Gajewski, who had co-founded an AI search startup, while they were housemates in San Francisco. They became friends and embarked on the AI music startup idea. To build something to show potential investors, they wanted to tap clusters of graphics processing units, or GPUs, the chips that help develop AI models. They needed a month of access, but suppliers were only offering terms of a year or longer.
Their solution was to reach out to other founders that also needed computing power, reasoning that if a dozen other startups also signed up, they could all share a yearlong contract.
Within a few weeks, 170 had signed on. Seeing the need, the pair abandoned their AI music dreams and decided to become a GPU cloud provider specializing in short-term access — also called “hero” runs in the industry. The company’s early customers now include research labs like Harvard University and Princeton University and startups like PlayHT Inc., Phind and Liquid AI Inc.
SF Compute is not the only startup providing fractional access to computing power or the first to pivot to providing infrastructure. Lambda Inc., which started as a facial recognition company, shifted its focus to provide on-demand GPU clusters for customers including Intel Corp. and the Defense Department. Other companies in the space include upstarts like Vast.ai Inc. and RunPod Inc. as well as industry heavyweight CoreWeave Inc., which is said to be nearing a valuation north of $19 billion.
Altman, who is the brother of OpenAI co-founder Sam Altman, said he expects a range of use cases for the SF Compute, and that venture capital firms and others who secured long-term GPU deals could use the trading platform to buy or sell access. “A bunch of VC firms could become customers,” Altman said. “SF Compute has had those conversations, and I think that’s a very plausible direction for this to play out.”
The startup will use the recent funding to build out the trading platform and has plans to double its mostly engineering staff to 30 people over the next year.
©2024 Bloomberg L.P.
