SF Compute runs a spot-and-forward market for GPU clusters. It manages more than $100M of hardware it does not own, lists H100 and H200 with B300 coming, and raised $40M at a $300M post-money valuation from DCVC and Wing VC with Electric Capital and Alt Capital (Data Center Dynamics).
The structurally interesting feature is not the inventory. It is that the market lets buyers resell their own unused capacity (Data Center Dynamics). Every customer is a potential supplier, which is an unusually elegant answer to Which side you build first — the supply side bootstraps out of the demand side once there is enough liquidity to make idle hours worth listing.
It is also, read the other way, the admission that this is a liquidity business rather than a sourcing business.
The money
$40M Series A at $300M post. Revenue: not disclosed. Gross margin: not disclosed. Utilisation, take rate, customer count, average contract size: none disclosed, anywhere.
The $100M+ of managed hardware is not revenue and must not be read as such. It is inventory under management — third-party assets whose owners SF Compute matches to renters. A brokerage's revenue is the fee on the flow, not the value of the flow, and the gap between those two numbers here is roughly two orders of magnitude. This is the GMV is not revenue error the atlas exists to prevent, and this vertical invites it more than most, because the headline figure is enormous and the fee figure is invisible.
What it charges — and why nobody knows
SF Compute publishes no fee schedule; the documentation page for it returns a 404. No competitor publishes one either. The atlas carries "low single-digit percentage take" for compute brokerage, and that number is a structural inference, not a sourced figure — derived from what brokered markets in indexed, fungible commodities converge on, not from any operator's accounts or any press report. It is [UNVERIFIED]. If SF Compute publishes a real number tomorrow and it is 15%, both this page and Compute and capacity brokerage are wrong.
The reasoning behind the inference is worth stating plainly, because it is the only thing supporting the number. GPU rental prices are published as an index (GPUsmith); Silicon Data sells GPU rental price indices as a product (Silicon Data); and CME Group and Silicon Data are listing compute futures on 5 October 2026 (CME Group). A buyer who can look up the spot price and hedge the forward curve does not pay a broker an expert-network margin.
Compare the disclosed and estimated take rates elsewhere in the atlas: Expert networks at 70–80%, Design, video and content production at an apparent 70–88%, engineer marketplaces at a disclosed 33–36%. The variable that best explains the ranking is how visible the underlying unit's price is. Compute's is fully visible. See What a rake can actually be.
What the valuation implies
$300M on a brokerage with undisclosed revenue is not a bet on margin. It is a bet on volume and on becoming the venue where the price forms.
That is a coherent thesis — exchanges are among the best businesses in existence — but it has two hard properties. It is winner-take-most, so the second and third venues in a liquid instrument are worth very little. And the competition is not other startups: CME is arriving with a listed contract, and if the forward curve forms there, the brokers are left with physical delivery and execution. See What the public market pays for labour and Marketplace, staffing firm, BPO or agency.
The demand side is not the risk. Compute is the deepest budget in this atlas — OpenAI's 2026 R&D compute budget alone is roughly $19B (Epoch AI), and ICONIQ has internal AI infrastructure spend going from 11% of revenue in 2025 to 16% in 2026 to 19% in 2027 (ICONIQ). The risk is that a budget can triple while the intermediary's percentage of it halves.
What to watch
Whether a fee schedule ever appears. A broker that publishes its take is confident the take is defensible. Continued silence is itself weak evidence about the level.
Spot volumes after October 2026. Once compute futures have a few quarters of history, the split between exchange-cleared hedging and bilateral brokerage becomes observable, and that split decides whether SF Compute is a venue or a service.
Whether the largest buyers ever transact here. OpenAI and Anthropic have signed direct multi-year compute agreements — Anthropic's SpaceX arrangement runs at $1.25B a month through May 2029 for roughly 325,000 GPUs (Sacra). Counterparties of that size negotiate directly. A brokerage whose addressable market excludes the two buyers who took 43% of all global H1 2026 venture funding (Crunchbase) is serving the mid-market of a barbell-shaped How much money is actually in the buyer pool.
The caveat
Almost everything above is structure rather than evidence. What is actually sourced about SF Compute is one funding round, one valuation, one inventory figure, a product description and a hardware list — from a single trade-press article. Founding date, headquarters, headcount, revenue, take rate and customer concentration are all absent from the record. Confidence on this page is low for that reason, and the number readers are most likely to quote from it — the take rate — is the one number nobody has ever published.