The computing power market lacks pricing, hedging, and financing infrastructure despite rapid growth.

Author: Multicoin Capital (Shayon Sengupta)

Compiled by Deep潮 TechFlow

DeepChain Summary: Multicoin views computing power as a novel productive asset, not merely hardware that depreciates in data centers. They believe that starting with physical computing power trading, then layering on futures, options, and lending, represents the most natural pathway for crypto to integrate with the real economy. For professionals focused on the intersection of DeFi and infrastructure, this piece clearly explains “why now.”

Computing power is the fastest-growing asset class in the world. Over the past few years, there has been extensive discussion regarding data center construction, the capital expenditures required to fund computing power facilities of all sizes, and the various innovative offtake agreements signed by AI labs and hyperscale cloud providers to secure capacity.

The first conclusion we draw from these phenomena is that demand for computing power will continue to grow, while the supply of chips, memory, electricity, and data center infrastructure will remain constrained.

A more significant implication is that a new universe of productive financial assets has emerged, meaning that mature markets will inevitably arise to price, finance, and hedge exposure to these underlying assets. Hundreds of billions of dollars in GPUs sit in data centers, generating cash flows under contracts of varying durations to serve workloads with differing performance requirements.

Today, most of these transactions are completed through customized bilateral agreements between large producers and consumers, and in some cases, via highly fragmented broker networks or matching markets. We believe there is a significant opportunity to build the first-generation correct exchange infrastructure: a market embedded with financial primitives that enables producers and consumers of compute to precisely express their risk preferences along price and availability curves—and across major hardware categories. We expect these new capital markets, which further financialize compute, are naturally suited to run on crypto rails.

In our view, the most compelling immediate objective is to build a platform that enables mid- and long-tail producers and consumers to manage their compute exposure over time. Today, this relies on a fragmented network of over-the-counter brokers and trading desks—but this is the first step toward a full-fledged exchange. Crucially, because compute is largely non-fungible, we believe all meaningful progress must begin by mediating transactions through physical delivery, rather than starting with synthetic or cash-settled instruments. Trading actual compute units will naturally lead to clustering or standardization over time, forming the foundation upon which financial instruments can be layered: futures or options contracts, new capital structures to fund new projects, and ultimately, the deployment of more efficient systems across their full lifecycle by using the underlying assets as collateral in an open and transparent risk engine.

The importance of the crypto rail here lies in verification, transparency, and composability. A trusted, neutral ledger can turn verified capacity rights into collateral; stablecoin lenders can extend loans against this collateral; and prime brokers can consolidate physical inventory and financial positions for margin management. Hash power has the potential to extend core financial primitives honed on the DeFi rail into one of the most exciting areas of the productive economy.

We believe this is a particularly unique moment to build compute finance primitives. The first significant change is that recent increases in inference demand and improvements in inference services have triggered a sharp reassessment of the residual and resale values of previous-generation chips.

Over the past period, NVIDIA’s product cadence led many in the data center capital structure to believe that GPUs would become economically obsolete within a few years. The argument was roughly this: each new architecture represents a step-function improvement in performance, causing older chips to fall out of favor among large-scale compute consumers, while asset owners depreciating their hardware over five to six years have effectively overestimated both returns and asset value.

We have recently seen compelling evidence contradicting this view. CoreWeave disclosed that it has signed customer contracts for A100 capacity extending through 2029. Since NVIDIA launched the A100 in 2020, this indicates that customers are willing to commit to using this chip nearly a decade after its release. Additionally, CoreWeave secured a $2.6 billion, five-year credit facility backed by a portfolio of customer contracts with an average term of three years. Lenders are explicitly underwriting the residual value of the GPUs and CoreWeave’s ability to resell this capacity after the initial contracts expire.

The A100 may no longer be the chip of choice for training cutting-edge models, but it remains well-suited for inference, fine-tuning, batch processing, academic research, and thousands of enterprise workloads where absolute performance is less critical than price or availability. Given this, you can now see much greater variance in spot prices for the A100/H100 chip category than ever before.

It is becoming increasingly clear that GPUs will no longer behave like consumer electronics, where new generations rapidly replace old ones, but rather like stepped productive assets that serve different market segments over time. After aging, chips will move down the cost curve, finding new buyers, workloads, and geographic regions.

The second important change is that computing power demand is becoming more fragmented.

The first phase of AI infrastructure development is dominated by a small number of cutting-edge labs and hyperscale cloud providers. These companies purchase large-scale clusters to train increasingly large proprietary models. Due to the need to secure supply to compete effectively at the frontier, they are largely price takers, while aiming to monetize through inference in the future.

The rise of open-weight models has significantly transformed this market structure. Today, every moderately sized company has teams deploying capability models on their own infrastructure, while inference providers can offer the same model across many different GPU configurations. End users choose from dozens of models based on cost, latency, geography, and performance. Sovereign AI initiatives add another dimension, as governments and regulated enterprises increasingly desire model deployments within specific jurisdictions and compliance with particular security or data residency requirements.

As a result, the number of buyers has increased significantly, and demand is now far more heterogeneous than before.

The third change is the emergence of model routing. OpenRouter is an early example of what a bidding system for inference order flow might look like. Applications submit requests, and the router determines which model and provider will fulfill them based on price and latency. Underlying provider pricing can vary significantly. For example, OpenRouter recently showed that the input price for Llama 3.3 70B ranged from $0.10 to over $1.00 per million tokens across different providers.

This price dispersion reflects a key fact: inference service providers are less like wholesale providers of compute power and more like market makers. They generate profits through batching, caching, quantization, model placement, and most importantly, maintaining high utilization across multiple clients. Two providers with identical hardware can produce vastly different numbers of tokens per dollar—much like two refineries deriving entirely different economic returns from the same barrel of crude oil. These aggregators are vital sources of price discovery, as they continuously reveal the marginal value of different models, providers, and hardware configurations in real time.

These developments have made the computing market more complex, but have also significantly increased the financial value surrounding computing. With more buyers, more sellers, and greater allocation, price disparities persist, and the economic lifecycle for underlying assets that can be contracted and financed has become longer.

There has been much discussion recently about whether computing qualifies as a commodity and whether traditional commodity financial instruments can be directly applied to computing. We believe that, at the current stage of the market, such analogies are not particularly useful.

For example, with oil of a particular grade, units of oil within that grade are largely interchangeable. Computing does not have this property, because the value of a GPU depends on its deployment configuration and environment, as well as the specific state it is processing.

Look at the variables contained in what appears to be a simple H100 contract:

PCIe or SXM

Number of GPUs in the node

Available memory

Interconnection, network topology

Storage, CPU, geographic location

Security, uptime, support, and scheduled duration and continuity

Renting an H100 for just a few hours is a different product from renting one continuously for three months. Clusters in Virginia and clusters in Iceland are also distinct, even if the GPUs themselves are identical. The same hardware may sell for $2 per hour per GPU in one market and $15 in another. Part of this difference stems from market fragmentation, but a larger portion reflects inherent differences in the products themselves. Moreover, reservations are notoriously difficult to cancel because computing is deeply embedded in operations: workloads carry state, data, dependencies, and performance requirements, making migration between providers a complex undertaking—not simply transferring SSH keys to another buyer.

Today, most buyers manage their compute exposure through a combination of spot capacity and long-term reservations. Spot and on-demand contracts offer flexibility, but buyers must accept uncertainty in pricing and availability. Long-term reservations guarantee access, but buyers risk overestimating demand and leaving expensive capacity idle.

The issue at hand is less a supply crisis and more a maturity mismatch, which is most apparent during periods of fluctuating demand. A reasoning service provider may operate smoothly for months, then experience a surge in demand when a client launches a product. A consumer AI application might suddenly go viral, requiring several times its usual capacity. A company may need a large cluster to complete a fine-tuning or evaluation task lasting only a few weeks. In each case, buyers face a dilemma: either overcommit to long-term capacity or rely on the spot market—which is precisely the market most likely to fail when capacity is most valuable.

The largest buyers address this issue through scale. They acquire capacity far exceeding that needed for any single workload, maintain inventory across multiple hardware generations and data centers, and continuously reallocate workloads across their portfolio. Their size allows them to absorb underutilization and manage demand internally.

For mid-tier and long-tail companies in the market, access to computing may be a matter of survival. These companies cannot afford large amounts of idle capacity, yet they also cannot risk discovering unavailable capacity when business demands arise. Over the past few years, numerous consumer-grade AI applications have demonstrated that even with massive user demand, a business model can still fundamentally collapse if inference costs and capacity requirements are not properly managed.

Nevertheless, buyers have almost no ability to hedge against the calculated price or availability. Sellers have limited means to monetize future capacity, transfer existing obligations, or finance GPUs without long-term customer contracts. The market has created physical assets and commercial demand, but has not yet created standardized claims against these assets.

The most direct attempt to financialize computation—or any asset—is to construct a price index and list cash-settled futures based on it. This provides participants with a reference price to express views on the average cost fluctuations of a particular GPU. We have already seen early proposals and implementations by exchanges such as Lighter and Architect.

However, the basis between the reference contract and the physical product is likely to be very large.

A buyer requiring continuous use of a B200 cluster in Europe for 90 days receives very limited protection from a cash-settled futures contract based on the average hourly price of U.S.-based H100 GPUs. Even within a single GPU generation, differences in networking, memory, location, contract duration, and service levels can cause the actual price paid by the buyer to diverge from the index movement.

Cash settlement also introduces capital efficiency concerns. New cloud service providers have taken on significant debt to finance GPUs and data center infrastructure, while inference providers need working capital to cover computation costs before receiving revenue from customers. Requiring these companies to post additional cash margin for futures positions increases the hedging costs for natural participants.

Finally, a cash-settled market without strong physical market backing will primarily attract speculators. Speculators are a useful source of liquidity, but they cannot reliably force the price of financial contracts to converge with the value of actual deliverable capacity. This requires market makers capable of procuring, delivering, replacing, and converting physical capacity.

Therefore, we are confident that any meaningful financial primitive serving producers and consumers must be built on physical settlement.

The current compute market is dominated by hyperscale cloud providers, direct sales teams, brokers, and an increasing number of spot markets. Brokers facilitate customized transactions and rely heavily on relationships. Market aggregators consolidate inventory and disclose more pricing information, but typically do not standardize contracts or act as counterparties in trades.

We believe the path to a computed capital market begins with a proprietary trading desk positioned between buyers and sellers, responsible for structuring contracts, managing physical delivery, and gradually standardizing the most common capacity forms. By intermediating real transactions, this desk accumulates proprietary insights into configuration-level pricing, counterparty quality, utilization rates, delivery failures, and actual basis differentials between different types of compute.

This information should enable the creation of transferable claims against physical capacity. For example, a seller can commit to providing an eight-GPU H100 SXM node in the Eastern U.S. for 30 days and receive a standardized credential representing that capacity. This credential can be transferred, used as collateral, or delivered into a forward contract prior to the booking period. Upon expiration, the holder gains access to the underlying cluster.

Our assessment is that computation will never be fully homogenized, but homogenization is not a binary concept. Two clusters may still be sufficiently interchangeable for a certain class of workloads, even if they cannot replace all buyers.

For example, many inference providers serving open-source models accept eight-GPU H200 SXM nodes from multiple vendors, provided they feature NVLink, can deliver continuous operation for 30 days within an acceptable region, and meet a minimum uptime threshold. We expect repeat demand to concentrate on a limited number of configurations and booking windows, providing a natural starting point for market standardization. Prime brokerage desks can discover these groupings through actual trades and verify whether capacity from different sellers meets the same delivery standards. If enough buyers treat certain clusters as practically interchangeable commodities, future capacity could be traded around those clusters using a common reference contract, creating a liquidity Schelling point.

Once trading venues are embedded within physical supply chains, the design space for financial products expands significantly.

Transferable forward: Buyers can lock in future capacity and later sell their position if demand changes. Sellers gain greater revenue visibility without permanently tying up capacity with a single customer.

Capacity options: Buyers can purchase the right to use a specific cluster over a future period, which is especially valuable for product launches, model deployments, fine-tuning, evaluations, and other variable workloads.

Portfolio margin: Prime brokers can identify hedging exposures between spot inventory, forwards, options, hardware generations, positions, and tenors. A supplier long H100 capacity and short B200 capacity should post margin based on the spread risk, rather than the combined value of both positions.

Request for Quote and Order Book: Standard contracts trade transparently, while large and customized demands are matched through competitive RFQs. These two structures coexist because buyers with annual procurement needs of $100 million in computing power have fundamentally different requirements from those with $1 million in procurement needs.

Verification: The marketplace can verify that the capacity genuinely exists, meets the claimed performance characteristics, and has not been sold multiple times to different buyers. For inference markets, the same verification layer can measure token count, latency, throughput, and uptime to prevent providers from falsifying performance.

Financing: Smaller new clouds can borrow against certified inventory, transferable forwards, and hedged future capacity, rather than relying solely on locked-in customer revenue. Observable forward curves also provide lenders with a better framework for evaluating residual value and renewal risk.

These products are broadly similar to a hash rate prime broker or commercial bank.

AWS and CoreWeave have already taken on many of these functions internally. They first commit to data centers and chips, then sell capacity across various terms. Their portfolio and deep demand-side liquidity enable them to manage the resulting mismatches.

A smaller new cloud essentially functions as an undercapitalized merchant. It borrows money to purchase GPUs, sells contracts with varying maturities, assumes the risk of underutilization, and bears residual value risk. Each company manages these exposures independently, even though these exposures could be netted across the entire market.

Independent prime brokers can aggregate market-wide risk by taking a position and hedging it with an opposite position. A diversified pool of positions can theoretically support new forms of credit. Such market participants aim to transform the existing "mismatch" in computing power to reduce idle capacity and lower the capital cost of new financing.

Over time, reference configurations may actually support credible price indices and useful cash-settled derivatives. Their effectiveness as hedging tools depends on how well the index tracks a repeatable workload from a sufficient number of buyers. Ultimately, even with perfect price hedging, buyers remain responsible for securing available capacity when needed. Therefore, standardization makes both contract forms more useful: physical delivery addresses availability risk, while cash settlement manages price risk.

Its core primitives can be viewed as trusted digital claims on physical assets or future services.

For example, this may include:

Security interest in GPU or server

The right to use a cluster during a future time period

Receivables under the underwriting agreement are aggregated into standardized claims.

Here, tokenization aims to make the underlying assets and associated cash flows easier to transfer, pledge, and re-collateralize. In each case, it is essential that these claims are supported by strong legal rights. This itself represents a design space full of opportunities, involving regulatory and verification infrastructure, and serves as a prerequisite for building robust capital markets around this asset class.

It’s easy to imagine that, after scaling, a new cloud provider could issue vouchers against its proven future capacity and sell a portion of them forward. The remaining inventory could support a margin loan denominated in a stablecoin. Lenders could underwrite the loan using the voucher’s forward price and contracted cash flows, with utilization history influencing the discount rate. In the event of borrower default, the collateral could be transferred to a market maker who already knows how to deliver or resell that computing power.

Buyers can purchase call options on products and resell them if plans change, with settlement conducted via delivery versus payment upon exercise. Reasoning providers can finance reservations backed by customer revenue and hold capacity options within the same margin account. Prime brokers can accept capacity certificates and extend credit against them, enabling these certificates to support broader secured financing arrangements.

In the short term, its core role is that of a "compute power prime broker," which today is essentially building its own collateral ledger. It records assets and corresponding obligations, maps the position of each claim within the capital structure, and then handles margin and settlement. Once scaled, we believe DeFi provides a natural architecture for this market, which features numerous asset originators and diverse sources of capital.

There are several reasons:

Collateral transparency: Lenders can see whether a capacity certificate has been pledged and which debt has priority. Discount rates and terms are visible on the same ledger.

Collateral interoperability: Capacity certificates held in the lending pool can also serve as margin for forward contracts designed to protect their value. Both platforms rely on the same custody layer.

Settlement is programmable because computing power is delivered over time: funds can be released incrementally as capacity becomes available. Collateral is returned once service level targets are met; if targets are not met, predefined penalties are triggered. When physical facts are verified, funds and collateral can be automatically transferred.

Crypto rails can reach global capital and lend around the clock against digital collateral: stablecoins provide a unified settlement asset (for example, one vault can hold senior credit, another can finance inventory or options).

A trusted and neutral blockchain can coordinate claims and their settlement, while cryptographic hardware proofs can verify the existence of physical assets. These layers build upon necessary legal agreements to establish ownership and enforcement mechanisms, forming the foundation required to unlock capital efficiency in the computing power sector.

Computational power has already supported massive asset and credit volumes, but infrastructure for transferring risk remains minimal. Ultimately, we believe protocols do not capture value—DAOs manage risk. Protocols that build intermediaries between producers and consumers can transform these contracts into transparent collateral systems.

We believe this is an incredibly rich design space with rapid evolution. If you're researching these issues or have a different perspective on how this market should develop, please reach out to shayon@multicoin.capital.

Thank you to Kyle Morris for numerous discussions that helped shape the thinking behind this article, and to Tomasz Tunguz and Mason Nystrom for their feedback.

Disclosure: Unless otherwise noted, the views expressed in this article are solely those of the author and do not reflect the views of Multicoin Capital Management, LLC or its affiliates (collectively, “Multicoin”). Certain information contained herein may be sourced from third parties, including companies in which funds managed by Multicoin have invested. Multicoin believes the information provided is reliable, but makes no representations regarding its ongoing accuracy or applicability to specific circumstances. This article may contain links to third-party websites (“external websites”). The presence of any such link does not constitute an endorsement of such websites, their content, or their operators. These links are provided for convenience only and do not imply our approval of the content on these external websites. The content of such external websites is developed and provided by others, and Multicoin assumes no responsibility for any such content. Charts provided herein are for informational purposes only and should not be relied upon in making any investment decisions. Any forecasts, estimates, forward-looking statements, goals, outlooks, and/or opinions expressed in this blog are subject to change without notice and may differ from or contradict opinions expressed by others.

This content is for informational purposes only and should not be relied upon as a basis for investment decisions, nor is it intended to be comprehensive. The information herein should not be construed as legal, business, or tax advice. You should consult your own advisors regarding these matters. Any reference to securities or digital assets is for illustrative purposes only and does not constitute investment advice or an offer to provide investment advisory services. Any investments or portfolio companies mentioned do not represent all investments held by Multicoin’s managed vehicles, and no guarantee is made that such investments will be profitable or that future investments will have similar characteristics or outcomes. A list of investments made by Multicoin’s venture funds can be viewed at: https://multicoin.capital/portfolio/. Investments not yet publicly disclosed are not included in this list due to coordination with development teams or issuers regarding the timing and nature of public disclosure. Additionally, for strategic reasons, Multicoin Capital’s hedge fund does not disclose its holdings in publicly traded digital assets.

This blog does not constitute investment advice, nor does it constitute an offer to sell or a solicitation of an offer to buy any limited partnership interest in any investment vehicle managed by Multicoin. Any offer or solicitation to invest in any Multicoin investment vehicle will be made solely pursuant to the offering memorandum, limited partnership agreement, and subscription documents, and investors should rely exclusively on the information contained in such documents when making an investment decision.

Past performance does not guarantee future results. There is no assurance that any Multicoin investment vehicle will achieve its investment objectives, and investment outcomes may vary significantly from year to year or even month to month. As a result, investors may lose all or most of their investment. The investments or products mentioned in this blog may not be suitable for you or any other party. The valuations provided are based on detailed assumptions as of the date of publication and may no longer be applicable after that date. Our target prices or valuations, as well as any base or upside scenarios used to derive them, may not be realized.

Multicoin has established, maintained, and enforced written policies and procedures reasonably designed to identify and effectively manage conflicts of interest related to its investment activities. For additional important disclosures, please see the disclosures and applicable terms at https://multicoin.capital/disclosures and https://multicoin.capital/terms.