5 reasons compute trading stays opaque for mid-market buyers

Compute trading happens in private group chats because the product is not fungible: GPU capacity trades as bespoke bilateral contracts with cluster-specific specs, prepayments and delivery terms. Affaan Mustafa and Alejandro Soumah Fernandez, who lead compute trading at Jump Trading, described that structure at AI Engineer and explained why no exchange has taken over.

What compute trading actually is, and why it stays in chat rooms

Compute trading is the private, bilateral buying and selling of GPU capacity, usually as long-term contracts with hardware-specific terms, and it has not consolidated into a public exchange. Affaan Mustafa, who leads compute trading at the proprietary trading firm Jump Trading, described at AI Engineer how deals move through Slack, Telegram and broker channels.

The unit of trade is a commercial agreement, not a token. A typical deal mixes a master services agreement, a service-level agreement and an order form, with cluster-specific and hardware-specific language in each order. The chip SKU (H100, H200 or B300 class) anchors the negotiation, but the contract covers power, rack space, delivery dates and access level.

Mustafa borrowed the term dark pool from equities to describe this. A dark pool in finance is an unlit trading venue where orders are not publicly displayed. Compute works the same way: public on-demand and spot listings exist, but tenured capacity with a fixed term and a defined total contract value trades out of public view.

The trading vocabulary is deliberate. These are commercial agreements, and treating them as trades lets participants reason about price exposure, hedging and counterparty risk the way commodity desks do. Brokers carry that role today. They see quotes for similar specs, know which buyers are inelastic, source the capacity, negotiate, and hold risk by operating the cluster while earning a spread.

The mismatch between falling token prices and firm compute prices

Token prices and compute prices have decoupled, which is the core puzzle behind compute trading today. Cheaper inference per token has not pushed GPU capacity prices down, and Mustafa pointed to Jevons paradox as the likely explanation.

The mechanism is straightforward. A cheaper task unlocks more tasks, so a team runs more tokens on the same budget, and demand for the underlying capacity rises even as serving gets more efficient per token. Mustafa described this as latent demand that has not been hit yet, and separated three quantities that are often conflated: the price of intelligence, the number of tasks people want to run, and the hardware required to serve them.

Capacity still has to be built and paid for. Mustafa listed the stack that has to line up before one usable GPU hour exists: chips, power, buildings, grid connections and capital. The International Energy Agency projects global data center electricity use rising from 485 TWh in 2025 to 915 TWh in 2030, a figure from the IEA's April 2025 update that Mustafa cited to show why power procurement now sits inside compute pricing.

Every upstream constraint flows down into the hourly cost. Land, interconnection queues, colocation space and lender appetite all shape what a cluster costs to stand up, and therefore what its owner must charge per hour to service the debt behind it.

Why mid-market buyers cannot sign the contracts that are on offer

Mid-market companies cannot sign the multi-year compute contracts that lenders want to underwrite, because the prepayment alone can consume an entire funding round. Mustafa described this as the central structural gap in compute trading.

The arithmetic is blunt. CoreWeave's public filings show large-scale infrastructure financing supported by take-or-pay customer contracts, and Mustafa cited a 15% to 25% median prepayment on a $100 million total contract value cluster. That is $15 million to $25 million due up front, structured over roughly 36 months.

For a startup that has just raised a $40 million round, the math leaves little room. Mustafa walked through an illustrative case: a $100 million three-year contract with a 30% prepayment leaves $10 million to run the business. Contracted capacity then moves up in steps tied to equity rounds, while demand arrives continuously, so the company caps out and turns customers away between raises.

On-demand capacity can bridge the shortfall, but only if the economics work. Mustafa said buyers must compare financing the prepayment against buying spot capacity at the moment of need, and the answer depends on the yield on the financed prepayment versus the spot price when the capacity is actually available.

Hyperscalers, frontier labs and larger neo clouds do not face the same constraint. They can access large equity, take on more debt and commit to ten-year deals, which is why the contract sizes on offer do not fit companies in the middle.

Ownership versus rental: the equilibrium compute buyers weigh

Renting capacity through an off-take contract usually approaches the cost of building and financing your own cluster, so the decision turns on utilization and how much control your workload needs. Mustafa presented this as an equilibrium where rental should sit below ownership cost.

The operating models differ in what you actually get:

• Build and operate the full stack. You own hardware and site, and your team handles power, permitting, assembly and management. Highest control, highest capital requirement.

• Own servers in a colocation facility. A partner handles most assembly and operations, and you keep the hardware asset.

• Off-take contract. The provider owns and operates the asset. Access to bare metal may be partial, and multi-tenant configurations can affect both pricing and what you can reconfigure.

Mustafa noted that some businesses value physical access, confidential compute, topology changes or the ability to reconfigure the system, and those firms may rationally pay more to own rather than rent. Companies with spiky or highly variable utilization can also find ownership cheaper in aggregate, despite the capital outlay.

The balance sheet effect is the part investors now watch. Procurement terms and payment schedules change how much runway a growing company has, which gives venture investors a direct interest in how their portfolio companies source capacity.

The players forming around the GPU hour

A cluster of intermediaries has formed around the GPU hour, each addressing a different part of the gap between buyers and suppliers in compute trading. Mustafa grouped them into four roles.

• Price discovery. Index providers, data vendors and benchmark publishers whose reference rates could underlie futures contracts, subject to spec and pending approval.

• Physical matching. Auctions, aggregators, decentralized on-demand marketplaces, liquid compute pools with standardized contracts, bare-metal resellers and neo clouds selling larger blocks.

• Desks and brokers. Firms quoting bilateral prices, sourcing capacity, negotiating terms and carrying operational and price risk for a spread.

• Financial and credit participants. Exchanges building GPU-hour derivatives, prepayment facilities, equipment lenders, lessors and specialist credit providers.

Buy-side demand is forming too. Mustafa said buyers ask for flexible contracts that allow upgrades when a new chip generation ships, rollover rights, or the ability to sell capacity back. Those are hedging requests dressed as contract terms.

The common thread is risk transfer. A company selling fixed-price inference while its GPU bill floats carries price exposure; locking capacity for years removes some of that exposure but creates idle-hour risk instead. Mustafa was explicit that a financial hedge can address a reference price but cannot deliver a cluster or guarantee utilization.

Why every compute deal is bespoke, and what standardization would change

Every compute deal is negotiated from scratch because the product is only partly fungible, and that is the mechanical reason compute trading remains opaque. Mustafa described spec sheets, formats and bills of materials that vary between deals.

The ambiguities are ordinary and consequential. A bill of materials may not match what is delivered if a part is out of stock, and long lead times from OEM and ODM suppliers shift delivery dates. Even vocabulary drifts: when someone says memory during a fast negotiation, they may mean VRAM, system memory or storage, and the answer changes what workloads the cluster can run and what it should cost.

Public reference prices help as a starting point, but they get chopped. A listed H100 rate is a snapshot, not a clearing price, because the actual quote depends on the cluster specification and every other factor in the deal.

Mustafa's proposed fix is a standard capacity contract with common specs, common terms and shared definitions, so buyers and suppliers can compare offers without days of back-and-forth. That is a market-structure project, not a product launch, and it requires participants with balance sheets strong enough to bridge the timing mismatch.

Policy: export controls and open-weight models feed back into US compute economics

Alejandro Soumah Fernandez, co-lead of compute trading at Jump Trading, placed the whole market inside a policy frame shaped by export controls and open-weight AI strategy. He covered the policy section in the same AI Engineer talk.

Soumah Fernandez said the US Commerce Department moved qualifying H200-class license reviews for China and Macau to case-by-case review with conditions in January, which affects how chip supply reaches buyers outside the United States. He also noted that the US AI action plan encourages open source and open-weight AI and recommends improving the financial market for compute so startups and academics can access large-scale capacity.

Policy and market structure connect through the middle market. Open-weight models let smaller companies serve workloads without paying frontier-lab margins, but they still need the hardware, and hardware access is where the financing problem bites.

FAQ

• What is compute trading? Compute trading is the private buying and selling of GPU capacity, typically as multi-year contracts with hardware-specific terms and large prepayments. Most deals clear through brokers and chat channels rather than a public exchange.

• Why do GPU deals happen in Slack and Telegram? Because each contract is tied to a specific cluster, delivery date and access level, so quotes are not directly comparable. Brokers match buyers and sellers bilaterally and hold risk, which a public listing cannot do.

• How large is the prepayment on a compute contract? Affaan Mustafa of Jump Trading cited a 15% to 25% median prepayment reported by CoreWeave, which on a $100 million total contract value means $15 million to $25 million up front.

• Why does Jevons paradox matter for GPU pricing? Cheaper inference per token expands usage, so demand for capacity grows even as serving gets more efficient. That keeps compute prices firm while token prices fall.

• Can futures hedge compute price risk? A derivative can hedge a reference price but cannot deliver a cluster or guarantee utilization. Companies still need physical capacity contracts alongside any financial hedge.

• What would make compute markets more liquid? A standard capacity contract with common specs, terms and definitions, plus intermediaries with strong balance sheets that can bridge the gap between long supplier commitments and smaller buyer blocks.

• Do buyers get full access to bare metal? Not always. Off-take contracts may grant partial access, and multi-tenant configurations can limit reconfiguration, which affects both workload fit and price.

• What does the IEA project for data center power? The IEA's April 2025 update projects global data center electricity use rising from 485 TWh in 2025 to 915 TWh in 2030, which ties power procurement directly to compute costs.

• Who can sign ten-year compute deals? Hyperscalers, frontier labs and larger neo clouds, because they can raise large equity, take on debt and underwrite long-horizon commitments that mid-market companies cannot.

From a conference talk to a written explainer

The value in this talk sits in its detail: prepayment percentages, the ownership-versus-rental equilibrium, and the vocabulary drift that makes quotes hard to compare. That detail survives transcription, but only if someone turns it into prose a reader can follow without watching 21 minutes of video.

If you have similar material in a recorded talk, interview or walkthrough, you can paste the YouTube URL into Skalablog, transcribe it, and generate an article draft that keeps the figures and the causal chain intact.

Source video