Every other vertical in this atlas runs on an information asymmetry. The buyer does not know what an expert costs, what a designer is paid, what a teleoperator earns. The middleman rents that ignorance and keeps 40–80% of the money that passes through.
Compute does not have that asymmetry. Street prices for GPU rental are published as an index (GPUsmith), Silicon Data runs GPU rental price indices as a product (Silicon Data), and CME Group and Silicon Data are launching compute futures on 5 October 2026 (CME Group).
The moment a supply becomes fungible and indexed, the spread collapses. The intermediary's revenue stops being margin per unit and becomes volume × tiny fee, plus adjacent services — financing, hedging, scheduling, reliability SLAs. No labour vertical in this atlas has a price index. There is no published rate for an hour of elite red-teaming, a bimanual demonstration, a Yoruba recording session or an expert call, and there are no futures on any of them. That absence is not a market inefficiency waiting to be corrected; it is the product the labour middleman sells. Every 70% take rate in What a rake can actually be is a bet that no index will ever exist for that unit.
Whose budget
The deepest in the atlas, without qualification. budget: 5.
Compute is new budget in the purest sense: it is often a startup's single largest cost line, it is approved at board level, and there is no "we would hire someone instead" alternative. Contrast Outbound and GTM-as-a-service, where the price ceiling is one SDR's salary.
The magnitudes: OpenAI's 2026 R&D compute budget is roughly $19B, about double 2025 (Epoch AI); its gross margin is 33%, constrained by inference, with inference costs going from $8.4B in 2025 to a projected $14.1B in 2026 (Sacra). Anthropic has a SpaceX compute agreement at $1.25B per month through May 2029, covering roughly 325,000 GPUs, alongside AWS, Azure and Fluidstack commitments (Sacra).
Below the labs, the pattern generalises. ICONIQ puts internal AI infrastructure spend at 11% of revenue in 2025, 16% projected for 2026 and 19% for 2027, and notes explicitly that "talent's share of cost falls while inference rises" (ICONIQ). The buyer is actively converting a people budget into a capacity budget — which is the tailwind for this vertical and, read the other way, the headwind for every labour vertical in the atlas.
Can you get the supply
It was hard in 2023 and it is getting easier every quarter, which is the wrong direction.
SF Compute runs a spot-and-forward market for GPU clusters and manages over $100M of hardware it does not own, listing H100 and H200 with B300 coming, having raised $40M at a $300M post-money valuation from DCVC and Wing VC with Electric Capital and Alt Capital (Data Center Dynamics). Assembling that inventory — idle clusters across neoclouds, over-provisioned enterprises and crypto miners — required real hustle in 2023.
Standardisation is dismantling that advantage. Its most telling feature is that the market lets buyers resell their own unused capacity (Data Center Dynamics): every customer is also a potential supplier, so the supply side bootstraps itself once liquidity exists. That is an elegant answer to Which side you build first and a terrible answer to "why can't a competitor do this."
supply: 2. Hard is good, and a fungible unit with a public index is the opposite of hard. This is a liquidity race, not a sourcing one.
speed: 3. A liquidity race sets a slower clock than a labour one. You cannot invoice on one side of this market: a broker needs listed inventory and a matched buyer in the same week, and the buyer is signing a board-level cost line rather than approving a retainer. Against that, the unit needs no training, no licence and no rig, and the buyer-resale mechanism means every customer you win is also supply you did not have to source — which is why this is not the 2 that Data-centre labour and site brokerage carries. Months to a real invoice, gated on the first matched trade rather than on assembling anything.
What the spread looks like
Low single digits, and this is an inference rather than a measurement.
SF Compute does not publish a fee schedule; the docs page for it returns a 404. Nobody in this vertical has disclosed a take rate. "Low single-digit percentage" is derived from what brokered markets in indexed, fungible commodities converge on, not from any operator's accounts. It is [UNVERIFIED] and should be read as an argument about market structure. If someone publishes a real number and it is 15%, this page is wrong.
Put the estimate against the rest of the atlas and the point becomes obvious: Expert networks keep 70–80% — a $200/hour expert bills the client $800–$900 (Inex One forum). Design, video and content production appears to keep 70–88%. Engineer marketplaces keep 33–36%, the one disclosed figure in the sweep, because engineer rates are semi-public. Compute, fully public, keeps almost nothing.
The spread tracks the opacity of the unit almost perfectly. spread: 1.
The implication for valuation is uncomfortable but clear: SF Compute's $300M mark on a brokerage of $100M+ of hardware only makes sense as a bet on volume and on becoming the price-formation venue — an exchange, not an agency. Exchanges are extraordinary businesses and there is usually one of them per instrument. See What the public market pays for labour and Marketplace, staffing firm, BPO or agency.
Can you hold it
Structurally poorly, with one narrow escape.
A fungible unit at a published price gives the buyer no reason to be loyal and the seller no reason to be exclusive. Both sides multi-home by default — the same behaviour visible on the model side, where builders now use 3.1–3.3 model providers on average, up from 2.8 six months earlier (ICONIQ). Getting cut out is not a risk here; it is the assumed steady state.
The escape is liquidity. If your venue is where the price forms, both sides come to you because that is where the counterparties are, and that is self-reinforcing in a way no relationship-based marketplace can match. But it is winner-take-most: venue number three in a liquid instrument has no business. And the incumbents arriving are not startups. CME listing compute futures means the price-formation venue for the forward curve may end up being CME, with the brokers reduced to physical delivery.
hold: 2. The buyer cannot be locked in, and the one mechanism that would lock them in is available to at most one company.
What AI does to it
ai: 5 — the only score shared with Adversarial evals and red-team crowds, for a completely different reason.
Models do not eat this work; models are the demand. Every capability increase raises training and inference requirements, and the buyer's own cost structure is shifting toward it by design — 11% to 16% to 19% of revenue in ICONIQ's projection, with talent's share falling as inference rises (ICONIQ).
The nuance is that a growing budget and a growing spread are different things. Compute demand can triple while the broker's percentage halves, and the futures market makes that more likely rather than less: hedging instruments exist to remove pricing risk, and pricing risk is where an unhedged broker's margin hides. See What better models do to each layer.
What would kill it
The futures market institutionalises the price and the broker becomes a plumber. From 5 October 2026, compute has a listed forward curve (CME Group). Once buyers can hedge on an exchange, the value of a broker who "finds you capacity at a good price" collapses to execution and delivery — genuinely useful, structurally low-margin, and exactly the role a bank's commodities desk plays.
Two other endings. The hyperscalers close the arbitrage — the spare capacity being brokered exists because provisioning is lumpy, and better internal scheduling at the largest providers removes the inventory. The counterparties consolidate — OpenAI and Anthropic together took 43% of all global H1 2026 venture funding (Crunchbase), and both have signed direct multi-year compute agreements at hundreds of billions of dollars. Buyers of that size do not need a broker; see One customer is a binary event.
Who is already there
| Who | Position | Money | What is public |
|---|---|---|---|
| SF Compute | Spot-and-forward GPU market; buyers can resell unused capacity | $40M Series A at $300M post; DCVC, Wing VC, Electric Capital, Alt Capital | Funding and >$100M of managed hardware; fee schedule 404s |
| Together, Lambda, CoreWeave-adjacent resellers | Neocloud capacity and resale | Not established in this record | Listed capacity only |
| Spheron | Decentralised compute brokerage | Not established in this record | None found |
| Silicon Data | GPU rental price indices | Not established in this record | Index product published |
| CME Group | Compute futures, listing 5 October 2026 | Public company | Launch announced |
Note what the last two rows do to the first. The infrastructure that makes a market efficient is being built by index providers and an exchange, not by the brokers — and efficiency is precisely what compresses a broker's take.
Where the record is thin
SF Compute's fee schedule page 404s and no competitor publishes one either. There is no revenue figure, no gross margin, no customer count and no utilisation rate anywhere in this vertical. The "low single digits" that this page's frontmatter carries is an inference from market structure, not a number anyone has seen.
Also unmeasured: how much of the brokered volume is genuinely spare capacity versus primary sales dressed as resale; whether the buyer-resale mechanism actually produces meaningful supply or is a marketing feature; and what happens to spot volumes once the futures contract has a year of history. That last one is answerable by simply waiting until late 2027.