Most long-term GPU capacity is traded through private quotes, brokers, and bilateral contracts in Slack, X, and Telegram rather than through a widely used exchange.
2
Falling token prices have increased demand for compute because cheaper inference encourages teams to use more tokens, while the underlying hardware, power, buildings, and capital still have to be paid for.
3
Mid-size companies struggle to finance capacity because suppliers want long commitments and large prepayments, while buyers need smaller, flexible blocks as their demand develops.
Summary
Affaan Mustafa and Alejandro Soumah Fernandez describe compute as an unlit market built around private quotes, brokers, and bespoke contracts. Long-term GPU capacity is negotiated in group chats and documented through master service agreements, service-level agreements, and order forms. The speakers explain why cheaper tokens have not produced cheaper GPU hours: lower inference costs encourage more usage, while chips, power, buildings, grid access, and financing remain expensive. Large hyperscalers and frontier labs can sign long contracts, but mid-size companies may have to put 15 to 25 percent down on a three-year agreement. That prepayment can consume an entire funding round. The talk compares building a cluster, owning servers in colocation, and renting capacity. It then examines brokers, marketplaces, futures, flexible contracts, and credit providers. The speakers argue that the market needs standard specifications and contracts, plus well-capitalized intermediaries that can carry risk between suppliers and growing buyers.
Compute capacity is traded through private, bespoke channels
Mustafa uses "dark pool" to describe an unlit market rather than a formal exchange. Physical compute usually trades in blocks through private quotes, brokers, and long bilateral contracts. The paperwork combines a master service agreement, service-level agreement, and order form, with language specific to the cluster and hardware. The main public references often focus on the GPU type, such as H100, H200, or B300, but the actual negotiation happens outside public view. Long-term capacity deals can be arranged through Slack messages, X messages, and Telegram group chats. Public marketplaces and on-demand prices exist, but they do not provide a widely adopted venue for trading fixed capacity over a defined term at a set contract value.
Cheaper inference can increase demand for GPU hours
The speakers separate the price of intelligence from the number of tasks people want to run and the hardware required to serve those tasks. Cheaper tokens let teams do more work within the same budget, and applications that were previously too expensive become possible. Mustafa connects this to the Jevons paradox. More efficient inference can encourage people to consume more tokens, so compute demand rises even as the cost per million input, cached, and output tokens falls. Inference clouds may serve more applications and tokens, while companies building on token infrastructure can serve more customers. This helps explain why token prices have declined while compute prices have remained strong.
A GPU hour depends on an entire physical and financial system
A usable GPU hour requires more than a chip. The speakers list chips, power, buildings, grid connections, and capital as inputs that have to come together before capacity can operate. They cite an IEA April update projecting that global data center electricity use will rise from 485 terawatt hours in 2025 to 915 terawatt hours in 2030. That growth makes power availability, energy sources, land, and data center build-outs part of compute pricing. Chip and system vendors supply GPU servers and networking. Site operators provide power, space, assembly, and management. Hyperscalers, neoclouds, and other capacity providers resell or operate bare-metal capacity for startups, inference clouds, laboratories, enterprises, and application companies.
Large buyers can commit for years, while mid-size companies cannot
Hyperscalers, frontier labs, and larger neoclouds can make substantial commitments because they can raise equity, take on debt, and sign long-horizon contracts. A Series B startup usually cannot commit to a $200 million-plus, three-year compute contract without consuming capital needed for the rest of the business. Lenders may finance infrastructure against contracted revenue, but a private, pre-profit or pre-revenue company gives them limited evidence beyond that contract. The speakers use CoreWeave's public filings as an example of infrastructure financing supported by take-or-pay customer contracts. They say CoreWeave has quoted median prepayments of 15 to 25 percent, which would mean $15 million to $25 million upfront on a $100 million total contract value cluster.
Prepayments turn compute procurement into a runway problem
The illustrative example has a company raising $40 million and signing a $100 million, three-year contract with a 30 percent prepayment. After putting down $30 million, only $10 million remains to run the business. A company whose demand is below its contracted capacity may need to wait for another equity round before adding a cluster, or use on-demand capacity to cover the gap. It could also borrow to finance the prepayment, but that only works if the financing cost makes sense compared with buying the needed capacity on demand. This connects compute procurement to venture financing. Investors have a direct interest in payment terms and access because a poor capacity deal can reduce the operating runway of a portfolio company.
Ownership, colocation, and rental fit different operating needs
The speakers describe three broad models. A company can build and operate the full stack, including hardware, the site, power, permits, assembly, and management. It can own servers in a colocation facility while a partner handles more of the assembly and operation. Or it can sign an off-take contract in which the provider owns and operates the asset. Rental can offer less control over bare-metal access, topology changes, confidential compute, and reconfiguration. Businesses that need those capabilities may accept the cost of owning and operating more of the stack. At high and consistent utilization, ownership may make more economic sense. With variable usage, an off-take contract can be preferable.
Standard contracts and risk-bearing intermediaries could connect the market
Every deal is currently bespoke. Specs, formats, bills of materials, delivery dates, and the meaning of terms such as memory can vary between negotiations. A financial hedge can address a reference price, but it cannot deliver a cluster or guarantee utilization. Flexible contracts could let a buyer upgrade to a new chip, roll capacity over, or sell it back. Brokers, marketplaces, futures, options, equipment lenders, and specialist credit providers each address different parts of the problem. Mustafa and Fernandez argue that the market needs common specifications, definitions, and a standard capacity contract. Well-capitalized participants could bridge suppliers that need long commitments and growing buyers that want to expand in smaller blocks.
"Chips, power, buildings, money, they all have to come together before you can even get to an hour of actual usable GPU per hour compute."Affaan Mustafa05:16
Who should watch
You are buying or financing GPU capacity and need to understand why public spot prices do not fully describe a long-term contract.
You run a growing AI company whose compute commitments compete with product spending and operating runway.
You are building a marketplace, financing product, futures contract, or flexible capacity agreement for data center infrastructure.