Multicoin Capital says the most valuable target in compute markets right now is not an abstract index product. It is a platform that lets mid-tail and long-tail producers and consumers manage compute exposure across time.
In the essay, written by Shayon Sengupta and translated by TechFlow, Multicoin frames compute as a new productive financial asset rather than hardware sitting in data centers and losing value. Its argument is that once these underlying assets exist at scale, markets for pricing, financing and hedging them will follow, and crypto rails are a natural place for that market structure to develop.
Compute is being treated as a fast-growing asset class
The article says the past few years have produced extensive discussion around data center construction, the capital expenditure needed to fund compute factories of different sizes, and the many forms of offtake agreements signed by AI labs and hyperscalers to lock in capacity.
From that, the authors draw a first conclusion: demand for compute will keep rising, while supply of chips, memory, power and data center infrastructure remains constrained.
The more important implication, in their view, is that a new universe of productive financial assets is taking shape. Data centers hold hundreds of billions of dollars in GPUs, those GPUs generate cash flow under contracts with different durations, and they serve workloads with very different performance requirements. Once that is true, the authors argue, mature markets for pricing, financing and hedging those assets are likely to emerge.
Today, most of those transactions still happen through customized bilateral agreements between large producers and large consumers, with some activity routed through fragmented broker networks or matching venues. Multicoin says that leaves room to build the first exchange-grade infrastructure for compute: a market with embedded financial primitives that lets producers and consumers express risk preferences across price and availability curves, and across major hardware categories.
Why the starting point should be physical compute
One of the essay’s core claims is that compute is non-fungible in many important ways. Because of that, meaningful progress should begin with intermediating trades through physical delivery rather than starting with synthetic or cash-settled instruments.
The logic is straightforward. If the market first trades real units of compute, some clustering and standardization should emerge over time. That can become the base layer for additional instruments: futures and options, new capital structures for funding new projects, and eventually systems that treat underlying assets as collateral inside open and transparent risk engines.
Multicoin says crypto matters here because of verification, transparency and composability. A credibly neutral ledger can turn verified rights to capacity into collateral. Stablecoin lenders can lend against that collateral. Prime brokers can manage margin across physical inventory and financial positions in one place. In that framing, compute becomes a path for extending financial primitives refined in DeFi into a productive part of the real economy.
Three shifts are opening the window for compute finance
Older GPUs are being repriced
The first shift, according to the article, is a sharp repricing of residual value and resale value for older chips, driven by inference demand and improvements in inference services.
For some time, Nvidia’s product cadence led many participants in data center capital structures to believe GPUs would become economically obsolete within a few years. The argument was that each new architecture delivered a step change in performance, older chips would fall out of favor with large compute buyers, and owners depreciating assets over five to six years were overstating both earnings and asset value.
Multicoin says recent evidence cuts against that view. CoreWeave disclosed that it signed a customer contract for A100 capacity that runs through 2029. Nvidia launched the A100 in 2020, which means at least one customer is willing to commit to using that chip close to a decade after release. CoreWeave also raised a $2.6 billion credit facility with a five-year term, backed by customer contracts with an average duration of three years. In the authors’ reading, lenders are explicitly underwriting GPU residual value and CoreWeave’s ability to resell that capacity after the initial contracts expire.
The A100 may no longer be the chip for training frontier models, but the article says it still works for inference, fine-tuning, batch processing, academic research and a wide range of enterprise workloads where absolute performance matters less than price or availability. On that basis, the authors say spot prices for chip classes such as A100 and H100 now show much wider dispersion than before.
The broader conclusion is that GPUs may stop behaving like consumer electronics, where one generation quickly replaces the last, and start behaving more like stepwise productive assets that serve different demand segments over time. As chips age, they move down the cost curve and find new buyers, workloads and geographies.
Demand is becoming more fragmented
The second shift is fragmentation in compute demand. The first phase of AI infrastructure buildout was led by a small number of frontier labs and hyperscalers. Those firms bought large clusters to train increasingly large proprietary models. Because competing at the frontier required securing supply, they were largely price takers and expected to monetize later through inference.
The rise of open-weight models changed that market structure, the article says. Today, many enterprises with some scale have teams deploying capable models on their own infrastructure, and inference providers can serve the same model across many different GPU configurations. End customers choose among dozens of models based on cost, latency, geography and performance.
Sovereign AI programs add another layer. Governments and regulated enterprises increasingly want models deployed within specific jurisdictions and under specific security or data residency requirements. The result, in Multicoin’s view, is a much larger buyer base and demand that is far more heterogeneous than before.
Routing layers are exposing price discovery in real time
The third shift is the emergence of model routing. The article points to OpenRouter as an early example of what an auction for inference order flow could look like. Applications submit requests, and the router decides which model and provider should serve them based on price and latency.
As one example, the piece says OpenRouter recently showed input pricing for Llama 3.3 70B ranging from $0.10 per million tokens to more than $1.00 per million tokens across providers. That kind of dispersion, the authors argue, shows that inference providers look less like wholesale compute resellers and more like market makers. They generate profit through batching, caching, quantization, model placement and, above all, maintaining high utilization across many customers.
Two providers with the same hardware can produce very different numbers of tokens per dollar. The article compares that to two refineries extracting very different economic value from the same barrel of crude. In that sense, aggregators become an important source of price discovery because they reveal the marginal value of different models, providers and hardware configurations in real time.
Those developments make compute markets more complex, but they also make financial markets around compute more valuable. There are more buyers, more sellers, more configurations, persistent price differences and a longer economic life for assets that can be contracted and financed.
Why a mature compute finance market does not exist yet
The article notes that many recent discussions compare compute with commodities and ask whether traditional commodity finance tools can be applied directly. Multicoin says that analogy is not very useful at the current stage.
Oil of a given grade is broadly interchangeable within that grade. Compute is not. The value of a GPU depends on the configuration and environment in which it is deployed, and on the specific state it is handling.
The article lists variables embedded even in what looks like a simple H100 contract:
- PCIe or SXM
- the number of GPUs in a node
- available memory
- interconnect and network topology
- storage, CPU and geography
- security, uptime, support, and the duration and continuity of the reservation
An H100 rented for a few hours and an H100 rented continuously for three months are different products. A cluster in Virginia and a cluster in Iceland are different products too, even if the GPU itself is identical. The same hardware may sell for $2 per GPU hour in one market and $15 in another. Part of that gap reflects market fragmentation, but the article says a larger part reflects real product differences.
Reservations are also hard to cancel because compute is deeply embedded in operations. Workloads carry state, data, dependencies and performance requirements, so moving them between providers takes real migration work rather than a simple handoff of access credentials.
That leaves most buyers managing exposure through some mix of spot capacity and long-term reservations. Spot and on-demand contracts preserve flexibility, but buyers accept uncertainty around price and availability. Long-term reservations secure access, but buyers take the risk of overestimating demand and leaving expensive capacity idle.
Multicoin argues that the central problem is less a supply crisis than a maturity mismatch. The issue becomes most visible when demand is volatile. An inference provider may run steadily for months and then face a surge when a customer launches a product. A consumer AI application may suddenly go viral and need several times its normal capacity. An enterprise may need a large cluster for only a few weeks to complete fine-tuning or evaluation. In each case, the buyer faces the same choice: overcommit to long-duration capacity or rely on spot markets that may fail exactly when capacity is most valuable.
The largest buyers solve this through scale. They secure far more capacity than any single workload needs, maintain inventory across hardware generations and data centers, and continuously reallocate workloads across a portfolio. Their size lets them absorb underutilization and manage demand internally.
Mid-sized and long-tail companies do not have that option. For them, access to compute can be existential. They cannot afford large pools of idle capacity, but they also cannot afford to discover that capacity is unavailable when the business needs it. The article says the past few years have already shown that even strong user demand does not save a consumer AI application if inference costs and capacity needs are not managed properly.
Even so, buyers have almost no way to hedge compute price or availability. Sellers have limited tools to monetize future capacity, transfer existing obligations or finance GPUs without long-term customer contracts. In Multicoin’s words, the market has created physical assets and commercial demand, but not standardized claims on those assets.
Why cash-settled futures are not enough
The most direct way to financialize compute, the article says, would be to build a price index and list cash-settled futures against it. That would give participants a reference price and a way to express a view on the average cost of a given GPU class. The piece notes that exchanges such as Lighter and Architect have already proposed and implemented early versions.
But Multicoin argues that basis risk between a reference contract and the physical product is likely to be very large. A buyer that needs a B200 cluster in Europe for 90 continuous days gets limited protection from a cash-settled future based on average hourly H100 pricing in the United States. Even within a single GPU generation, differences in networking, memory, location, contract term and service level can break the link between the index and the price a buyer actually pays.
Cash settlement also creates capital-efficiency problems. New cloud providers are already borrowing heavily to finance GPUs and data center infrastructure, while inference providers need working capital to pay for compute before collecting revenue from customers. Requiring those firms to post additional cash margin for futures positions raises hedging costs for the natural users of the market.
The article adds that a cash-settled market without deep physical-market support will mainly attract speculators. Speculators can provide liquidity, but they cannot reliably force financial prices to converge with the value of deliverable capacity. That requires market makers that can source, deliver, substitute and transform physical compute.
For that reason, Multicoin says any meaningful financial primitive for producers and consumers has to be built on physical delivery.
From trading desk to exchange: a compute prime broker
The current compute market is dominated by hyperscalers, direct sales teams, brokers and a growing number of spot venues. Brokers arrange customized deals and rely heavily on relationships. Marketplaces aggregate inventory and reveal more pricing, but they usually do not standardize contracts or step in as principal.
Multicoin says the path to compute capital markets starts with a proprietary trading desk sitting between buyers and sellers. That desk would structure contracts, manage physical delivery and gradually standardize the most common forms of capacity. By intermediating real transactions, it would build proprietary information on configuration-level pricing, counterparty quality, utilization, delivery failures and the actual basis between different types of compute.
That information, the article argues, should make it possible to create transferable claims on physical capacity. One example is a seller committing to provide an eight-GPU H100 SXM node in the U.S. East region for 30 days and receiving a standardized certificate representing that capacity. That certificate could be transferred before the reservation starts, pledged as collateral or delivered into a forward contract. At maturity, the holder gets access to the underlying cluster.
Multicoin’s view is that compute will never become fully fungible, but fungibility is not binary. Two clusters may be interchangeable enough for a certain class of workload even if they are not substitutes for every buyer.
The article gives one example: many inference providers serving open models would accept eight-GPU H200 SXM nodes from multiple suppliers, with NVLink, as long as the node can be delivered continuously for 30 days in an acceptable region and meets a minimum uptime threshold. The authors expect repeat demand to cluster around a small number of configurations and booking windows, creating a natural starting point for standardization. A prime-broker desk could discover those groupings through actual trades and verify whether capacity from different sellers meets the same delivery standard. If enough buyers treat a set of clusters as practical substitutes, future capacity can trade around a shared reference contract and form a Schelling point for liquidity.
What financial products could sit on top of physical flows
Once a venue is embedded in physical supply flows, the design space opens up, the article says. It lists several products and functions:
- Transferable forwards: buyers can lock in future capacity and later sell the position if demand changes. Sellers gain better revenue visibility without permanently tying capacity to one customer.
- Capacity options: buyers can purchase the right to use a specific cluster during a future period, which is useful for product launches, model releases, fine-tuning, evaluation and other volatile workloads.
- Portfolio margining: a prime broker can recognize hedged exposures across spot inventory, forwards, options, hardware generations, locations and tenors. A supplier long H100 capacity and short B200 capacity should post margin on spread risk rather than on the gross value of both positions.
- RFQ and order books: standardized contracts can trade transparently, while large and customized demand can be matched through competitive requests for quote. The article says both structures can coexist because a buyer procuring $100 million of annualized compute and a buyer procuring $1 million do not have the same needs.
- Verification: the venue can verify that capacity exists, matches the claimed performance characteristics and has not been sold multiple times. In inference markets, the same verification layer can measure token counts, latency, throughput and uptime to prevent providers from faking performance.
- Financing: smaller new-cloud providers can borrow against certified inventory, transferable forwards and hedged future capacity instead of relying entirely on locked-in customer revenue. Observable forward curves also give lenders a better framework for assessing residual value and renewal risk.
Taken together, the article says, those functions resemble a compute prime broker or a commercial bank.
AWS and CoreWeave already perform many of these functions internally. They commit to data centers and chips first, then sell capacity across multiple durations. Their asset portfolios and deep demand-side liquidity let them manage the mismatch that follows.
A smaller new-cloud provider, by contrast, looks more like an undercapitalized merchant. It borrows to buy GPUs, sells contracts across different tenors, carries underutilization risk and bears residual value risk. Each company manages those exposures on its own even though many of them could be netted across the broader market.
Multicoin says an independent prime broker could aggregate risk across the market by holding one exposure and hedging it with the opposite position. A diversified pool of positions could, in theory, support new forms of credit. In that model, the intermediary is trying to transform the “form mismatch” that exists in compute today, reduce idle capacity and lower the cost of new financing.
Over time, the article says, reference configurations may support credible price indexes and useful cash-settled derivatives. Whether those tools hedge effectively will depend on how well the index tracks a repeatable set of workloads from a large enough buyer base. Even then, buyers still need actual capacity when they need it. Standardization therefore makes both contract types more useful: physical delivery addresses availability risk, while cash settlement addresses price risk.
Compute as collateral on crypto rails
The core primitive, in the article’s framing, is a trusted digital claim on a physical asset or a future service.
That could include:
- a security interest in GPUs or servers
- the right to use a cluster during a future time window
- receivables under offtake agreements, pooled into standardized claims
The purpose of tokenization here is to make the underlying assets and related cash flows easier to transfer, pledge and re-pledge as collateral. In every case, the article says, those claims need strong legal rights behind them. That is itself a large design space involving regulation and verification infrastructure, and it is a prerequisite for building durable capital markets around the asset class.
At scale, the article imagines a new-cloud provider issuing certificates against verified future capacity and selling some of those forwards. The remaining inventory could support a stablecoin-denominated margin loan. Lenders could underwrite the loan using the forward price of the certificate and the contract cash flows, while utilization history would affect the haircut. If the borrower defaults, the collateral could be transferred to a market maker that already knows how to deliver or resell the compute.
Buyers could purchase call options for product launches and resell them if plans change, with exercise settled delivery-versus-payment. Inference providers could finance reservations against customer revenue while holding capacity options in the same margin account. A prime broker could accept capacity certificates and extend credit against them, and those certificates could then support broader secured financing arrangements.
In the near term, Multicoin says the key role is a “compute prime broker” that is effectively building its own collateral ledger. It records assets and matching obligations, shows where each claim sits in the capital structure, and then handles margin and settlement. At larger scale, the authors believe DeFi offers a natural architecture for a market with many asset originators and many sources of capital.
The article gives several reasons:
- Collateral transparency: lenders can see whether a capacity certificate has already been pledged and which debt sits senior to which. Haircuts and maturities are visible on the same ledger.
- Collateral composability: a capacity certificate held in a lending pool can also serve as margin for a forward contract that protects its value. Both venues rely on the same custody layer.
- Programmable settlement: because compute is delivered over time, funds can be released as capacity becomes available. Collateral can be returned when service levels are met, or penalties can be triggered when they are not. If physical facts can be proven, funds and collateral can move automatically.
- Access to global capital: crypto rails can connect global capital to round-the-clock lending against digital collateral, while stablecoins provide a common settlement asset. The article gives the example of one vault holding senior credit and another financing inventory or options.
In the authors’ view, a credibly neutral blockchain can coordinate claims and settlement, while cryptographic hardware proofs can verify what actually exists. Those layers would sit on top of the legal agreements needed to establish ownership and enforcement, creating the foundation for better capital efficiency in compute.
Closing view and disclosures
The article concludes that compute already supports a large stock of assets and credit, but the infrastructure for transferring risk remains thin. Multicoin’s view is that protocols do not capture value and DAOs manage risk; protocols that intermediate between producers and consumers can turn these contracts into transparent collateral systems.
The firm describes the area as a rich design space that is evolving quickly. It says readers working on these questions, or those who disagree with how the market should develop, are welcome to reach out to the email address included in the original article.
The piece also thanks Kyle Morris for multiple discussions that helped shape the thinking behind it, and thanks Tomasz Tunguz and Mason Nystrom for feedback.
In its disclosure section, the article says that unless otherwise stated, the views expressed are the author’s own and do not represent the views of Multicoin Capital Management, LLC or its affiliates. Some information may come from third-party sources, including companies in funds managed by Multicoin. The firm says it believes the information is reliable but makes no representation about its continuing accuracy or suitability for any specific situation.
The article also says it may contain links to third-party websites, and that those links are provided for convenience rather than endorsement. Charts are for informational purposes only and should not be relied on for investment decisions. Any forecasts, estimates, forward-looking statements, targets, outlooks or opinions may change without notice and may differ from the views of others.
Multicoin adds that the content is for informational purposes only and should not be used as the basis for investment decisions. It should not be read as legal, business or tax advice. References to securities or digital assets are illustrative only and do not constitute investment advice or an offer to provide advisory services. Investments or portfolio companies mentioned do not represent all investments across Multicoin-managed vehicles, and there is no guarantee that any investment was or will be profitable or that future investments will have similar characteristics or results.
The article includes a link to the list of venture fund investments at https://multicoin.capital/portfolio/ and says unpublished investments are not included because of coordination with development teams or issuers around the timing and nature of public disclosure. It also says Multicoin Capital’s hedge funds do not disclose positions in publicly traded digital assets for strategic reasons.
The disclosure further states that the blog is not investment advice and does not constitute an offer to sell or a solicitation to buy limited partner interests in any Multicoin-managed investment vehicle. Any such offer or solicitation would be made only through offering memoranda, limited partnership agreements and subscription documents, and investment decisions should rely only on those materials.
Finally, the article says past performance does not guarantee future results. There is no assurance that the investment objectives of any Multicoin vehicle will be achieved, and results may vary materially from year to year or even month to month. Investors may lose all or a substantial portion of their investment. Investments or products mentioned may not be suitable for every person. Valuations are based on assumptions detailed at the time of publication, and those assumptions may no longer apply after publication. Price targets, valuations and the base or upside scenarios used to derive them may not be realized.
Multicoin says it has established, maintains and enforces written policies and procedures designed to identify and manage conflicts of interest related to its investment activities. It directs readers to https://multicoin.capital/disclosures and https://multicoin.capital/terms for additional disclosures and applicable terms.