The San Francisco Compute Company

Here is a thing that happens in the AI industry, and it is a slightly absurd thing. You are two founders, Evan Conrad and Alex Gajewski, and in 2023 you want to train an AI audio model. To train a model you need GPUs, and GPUs are scarce, and the people who have them do not want to rent you a few for a weekend. They want to sell you a 12-month contract for far more compute than you need. So you sign it. And now you are the owner of a large, depreciating pile of graphics cards that you are mostly not using, which is a financial position roughly nobody chooses on purpose.

The reasonable response to this is despair. The Conrad-and-Gajewski response was to notice that other people were in exactly the same bind - over-committed, under-utilized - and that if you could resell the hours you weren't using to someone who needed them, everyone came out ahead. That side activity turned out to be a much better business than the audio model. It became The San Francisco Compute Company, which everyone calls SF Compute, and which is now trying to do to GPUs what commodities exchanges long ago did to wheat and oil: make them trade.

The core insight is that a GPU-hour is a commodity. An hour on an Nvidia H100 in one data center is, give or take some plumbing, an hour on an H100 in another. Commodities want to be traded on markets, priced in real time, bought and sold and resold. But the AI compute market mostly doesn't work that way. It works like commercial real estate: long leases, opaque pricing, take-it-or-leave-it terms. SF Compute's bet is that the real-estate model is a temporary accident of a supply shortage, and that underneath it there's a proper market trying to get out.

So SF Compute built the market. On its platform you can buy thousands of H100s for a single hour, set a price limit and let the system buy when compute drops below it, or list capacity you're not using and let someone else take it off your hands. Prices move. During 2024's volatility, H100s traded anywhere from roughly $1 to $8 an hour depending on the day; SF Compute's average has hovered around $2 an hour. The point isn't the specific number. The point is that there is a number, visible to everyone, that moves - which is what a market looks like and what most of AI compute conspicuously lacks.

How It Works

Own the risk nobody else wants

Now, the clever part, and the part that makes SF Compute more interesting than "GPU Airbnb," is where it sits in the risk stack. A lot of GPU companies own hardware and then wrap it in software services, hoping the software margins cover the brutal depreciation on the hardware. Conrad's view, stated with the confidence of a man who has read the balance sheets, is that this quietly loses money. Owning compute is a capital-intensive, high-risk business, and bolting a nice API onto it does not change the underlying physics.

SF Compute's structure is designed around that observation. It secures long-term GPU capacity - the boring, cheap-per-hour kind - and then sells it in flexible, short-term, resellable slices at market prices. The arbitrage between a cheap long-term contract and expensive short-term demand is the business. When it works, utilization approaches 100%, because anything the primary buyer doesn't use gets sold to someone who will. The company takes a spread and, crucially, tries to structure things so it is not the one holding the depreciation bag.

To make the commodity fungible, SF Compute runs its clusters, in its own phrase, "from UEFI on up." It controls the boot layer, standardizes the hardware, burns each cluster in with LINPACK testing, and monitors it actively so that an H100-hour you buy behaves like the H100-hour the next person buys. It also does the customer-friendly things traditional clouds resist: it never charges you to move your data out, and it passes SLAs and automated refunds straight through to buyers. Egress fees are how clouds trap you; removing them is a quiet statement of confidence that they'd rather keep you by being good.

The far horizon is more ambitious than a spot market. What SF Compute really wants to build is futures - cash-settled contracts that let a data center lock in a price for compute it will deliver next winter, the way a farmer locks in the price of corn. That's what Conrad means when he says futures will "chill out the entire industry." A lot of the frantic energy in AI infrastructure comes from everyone guessing where GPU prices go. If you can hedge that risk instead of guessing, the whole system gets calmer, and valuations stop being a bet on a single unknowable number. It is a genuinely financial vision of a business that most people think of as cables and cooling.

The tell that a market is real is that the smallest and largest participants use the same one. On SF Compute, a grad student with a two-week deadline and a burst-capacity need buys through the same spot market as large AI labs and enterprises. Named users span Nvidia, Roboflow, Datology, Liquid AI, and MIT, alongside AI labs such as Standard Intelligence and Phind. Academic researchers - including Schmidt Futures grantees - use it precisely because they need a lot of compute for a short, defined window and then nothing, which is exactly the shape traditional long-term contracts handle worst.

In November 2025, SF Compute closed a $40 million Series A that valued it at $300 million, led by DCVC and Wing Venture Capital, with participation from existing backers Electric Capital and Alt Capital. Total funding sits around $52 million. At the raise it managed over $100 million in hardware with a team of roughly 30, since grown toward the mid-50s. The pitch to investors was notably un-flashy: not "another GPU cloud," but a marketplace - the exchange where all that compute gets priced. DCVC's Ali Tamaseb framed it as "a true marketplace for compute," letting companies buy on flexible short-term contracts while data centers sell excess capacity.

The team is built for the unglamorous half of the job. CTO Eric Park previously ran Voltage Park; other executives came from Sun Microsystems and Lambda. It is, deliberately, an infrastructure team wrapped around a financial idea.

There's a nice irony in SF Compute's aesthetic. Its website is calm and nature-forward, almost pointedly under-hyped in an industry that runs on superlatives, with a hover Easter egg tucked onto the buy page for people paying attention. The company sets low expectations and then tries to beat them with the product, which is the opposite of how most AI infrastructure markets itself. The bet underneath the calm is that the most contrarian thing you can build in AI right now is not a bigger model but a working market - the mundane financial plumbing that turns a speculative frenzy into something you can price, resell, and hedge. If that bet is right, SF Compute won't be the loudest company in AI. It'll just be the one quietly running the exchange.