C:\MARKETS> type MAPPIN~1.MD
Mapping the compute economy
Compute is already being sold forward. Can those commitments become a market?
Questions or disagreements? Highlight any passage and press ASK ABOUT THIS to send it to me with your question.
Introduction
This is the first piece in what I hope will become an ongoing series exploring compute markets. I'm approaching the subject as someone learning about the industry, not as an expert. I'll undoubtedly get some things wrong along the way, but I'm sharing the research openly in case others find it useful. Corrections, feedback and conversations are very welcome: highlight any passage and press ASK ABOUT THIS to email me a question or comment about it. If you'd like to follow along, you can sign up at the bottom of the page to receive future pieces by email.
One of the things I find most interesting about finance is its ability to move value across time. Future income can be brought into the present to finance investment, while today's savings can be deployed to generate value in the future.
This was part of what attracted me to crypto more than a decade ago: witnessing a new asset class emerge and the markets developing around it.
Today, I'm interested in how similar mechanisms are taking shape around compute.
AI companies are committing billions of dollars to computing capacity they will consume years from now. Infrastructure providers use those commitments to finance the data centres, GPUs and networks needed to deliver it. In effect, compute is already being sold forward, through long-term commercial contracts rather than standardized financial instruments.
These arrangements work, but they also bundle several risks together: future prices, capacity availability, delivery obligations and customer credit.
The question isn't whether a market for future compute can exist. One already does. It's whether parts of that market can become more transparent, transferable and standardized, and whether doing so would create real value.
Before exploring that possibility, I want to understand how the industry works, who holds the risks and how those relationships are evolving.
1. What exactly is being sold?
Demand for accelerated computing, particularly for AI training and inference, is driving investment in specialized hardware and infrastructure.
Training is the process of developing or refining an AI model using data and computational resources. Inference is the process of running a trained model to produce outputs, such as generating text or analyzing images. Both consume compute, but their infrastructure requirements can differ significantly.
GPUs are central to these workloads, alongside alternatives such as Google's TPUs and other accelerators. But owning a GPU and buying compute are different things.
Four terms help distinguish the products:
- Colocation (colo): Renting space, power, cooling and connectivity in a third-party data centre, typically to operate your own hardware. The colocation operator provides the physical infrastructure, while its customer generally owns or controls the servers.
- GPU rental: Paying to use GPUs without owning them, from hourly access to long-term capacity commitments. Rental pricing is commonly expressed in dollars per GPU-hour.
- Bare-metal compute: Renting a physical server rather than a virtual machine. The customer gets direct access to the server's hardware instead of sharing its resources through a virtualization layer.
- Managed compute: Purchasing a higher-level service, such as model training or inference, without directly operating the underlying infrastructure.
From individual GPUs to clusters
A GPU-hour represents access to one GPU for one hour. It's a convenient pricing unit, but not a standardized measure of computational output.
Different GPU generations have different capabilities. An NVIDIA H100 and B200, for example, vary substantially in performance, memory and the types of workloads they can efficiently execute.
Configuration matters too. A GPU cluster is a group of accelerators connected to operate together. For large-scale training, the speed and architecture of those connections can be as important as the performance of individual chips.
NVIDIA's HGX H100 platform commonly packages eight H100 GPUs into a single server node, connected internally through NVLink and NVSwitch. Larger training clusters connect many of these nodes through additional high-speed networking. This is a common configuration, not a universal minimum: other systems and cloud products provide smaller allocations.12
The difference becomes clearer with numbers:
| Configuration | H100 GPUs | GPU-hours over 730 hours |
|---|---|---|
| Individual GPU | 1 | 730 |
| HGX H100 server node | 8 | 5,840 |
| 32-node training cluster | 256 | 186,880 |
CoreWeave, for example, lists an eight-H100 HGX node at $49.24 per hour on demand, equivalent to $6.16 per GPU-hour. Renting that node continuously for a standardized 730-hour month would cost approximately $35,945 at the published rate.13
This illustrates the difference between a pricing unit and a physical product. Eight GPUs in one tightly interconnected server are not necessarily interchangeable with eight individual GPUs rented independently, even if their aggregate GPU-hours are identical.
A research laboratory training a frontier model may need thousands of tightly interconnected GPUs for months. A startup fine-tuning a model might rent eight GPUs for a few days. An enterprise purchasing inference services might care only about the cost, latency and quality of the output.
These customers consume compute, but they aren't buying interchangeable products.
For the purposes of this research, I'll distinguish physical infrastructure, contractual access to computing capacity, and the useful computational output delivered to customers.
The most immediately promising unit of standardization is access to a specified amount of capacity over a specified period. Whether that is sufficiently fungible to support liquid markets remains to be seen.
2. The economic relationships
Rather than mapping compute as a conventional supply chain, I find it more useful to follow capital, capacity and risk.
The principal participants are hardware suppliers, electricity providers, data-centre operators, GPU cloud providers, hyperscalers, customers and financiers.
Figure 1 — How capital and compute move
flowchart TB
HW["Hardware suppliers<br/>NVIDIA · AMD · OEMs"]
PW["Utilities & grid"]
DC["Data-centre operators<br/>Colocation"]
GPU["GPU cloud providers<br/>CoreWeave · Lambda"]
HS["Hyperscalers<br/>AWS · Azure · Google"]
CU["AI labs & enterprises"]
FIN["Lenders & investors"]
PW -->|"Electricity"| DC
DC -->|"Powered space"| GPU
GPU -->|"Lease payments"| DC
HW -->|"GPU systems"| GPU
GPU -->|"Equipment payments"| HW
FIN -->|"Project finance"| DC
FIN -->|"Hardware & corporate finance"| GPU
GPU -->|"Debt service & returns"| FIN
DC -->|"Debt service & returns"| FIN
GPU -->|"Compute capacity"| CU
CU -->|"Contract payments"| GPU
GPU -->|"Wholesale capacity"| HS
HS -->|"Capacity payments"| GPU
HS -->|"Cloud services"| CU
class HW supply
class PW,DC infra
class GPU,HS provider
class CU customer
class FIN capitalFigure 1. Simplified economic flows. Large companies operate across multiple roles; hyperscalers can be infrastructure owners, compute providers and customers simultaneously. The arrows represent selected commercial relationships, not every transaction or ownership structure.
The owner of a data centre may not own the GPUs inside it. A GPU provider can lease powered capacity from one business, purchase equipment from another and sell computing services to a third.
Hyperscalers don't always supply compute exclusively from their own facilities. They can also purchase capacity from specialized providers. Microsoft's relationship with CoreWeave is an example.
Capital enters at several levels, often supported by different assets and contracts. The resulting risks are distributed unevenly.
| Participant | Principal risks | Common arrangements |
|---|---|---|
| Colocation operator | Construction, power availability, tenant credit | Long-term leases and capacity commitments |
| GPU provider | Hardware depreciation, utilization, deployment | Reserved capacity and take-or-pay contracts |
| Lender | Customer credit, asset value, delivery and borrower performance | Contract-backed and asset-backed financing |
| Compute customer | Future requirements, availability, rental prices | Reservations, long-term commitments and on-demand purchasing |
The distinction between these roles becomes clearer through a concrete example.
3. CoreWeave: bringing future revenue into the present
CoreWeave provides a useful case study because its public filings reveal how customer commitments support substantial infrastructure investment.
The company secures data-centre capacity, deploys GPU systems and sells computing services to customers, including Microsoft and OpenAI.
Infrastructure commitments
One of CoreWeave's colocation partners is Core Scientific, which has been transitioning from Bitcoin mining into AI and high-performance computing infrastructure.
At September 30, 2026, Core Scientific reported approximately 590 MW of leased customer power capacity contracted to CoreWeave across five sites. Of that, 485 MW had commenced billing, while the remaining 105 MW was under construction or commissioning.1
Data-centre capacity is commonly measured in megawatts (MW), reflecting the electrical power available to support computing infrastructure.
A distinction matters here: IT capacity refers to power available for computing equipment, whereas total facility power also includes cooling and other supporting systems. Published capacity figures don't always use the same definition.
Nor does electrical capacity tell us exactly how much useful compute is available.
Three measures are worth separating:
- Contracted capacity: Power or compute that a provider has agreed to supply, potentially before the infrastructure is ready.
- Billable capacity: Capacity that has reached the contractual stage at which the provider can begin charging its customer.
- Utilized capacity: The portion actually being consumed by operating workloads or equipment.
A facility can be billable without its computing equipment operating at full utilization. It can also have available electrical power before GPUs have been installed.
Core Scientific's figures distinguish contracted from billable capacity, but don't reveal the actual utilization of CoreWeave's GPUs.
CoreWeave also attempted to acquire Core Scientific in 2025. The transaction was terminated after Core Scientific shareholders rejected it, illustrating how the boundary between infrastructure ownership and compute provision can itself become a strategic question.2
Financing future capacity
At September 30, 2026, CoreWeave reported $17.6 billion outstanding under delayed-draw term loan facilities, part of approximately $47.5 billion in total indebtedness.3
The delayed-draw facilities are secured by assets associated with customer contracts and pledged contractual cash flows, generally from investment-grade counterparties. The financing arrangements support infrastructure deployment, with repayments tied to the facilities' contractual terms and revenue expectations.
A customer's future payments help support borrowing today. The borrowing funds infrastructure needed to deliver the service, and revenue from that service helps repay the debt.
The lender is underwriting more than the physical GPUs. The quality and enforceability of customer contracts are central, alongside collateral value, deployment and the borrower's ability to perform.
NVIDIA is another participant with overlapping interests. It supplies the technology underpinning CoreWeave's business and invested $2 billion in its equity in January 2026. A critical supplier is therefore also an investor in a major customer.4
Future revenue commitments
Perhaps the most revealing figure is the size of CoreWeave's contracted future business.
At September 30, 2026, the company reported $132.3 billion in remaining performance obligations (RPO). Of that amount, 36% was expected to be recognized as revenue within the following 24 months.3
RPO represents the portion of contracted transaction value allocated to services not yet delivered. It is not cash already collected or unconditional future revenue.
Nevertheless, the figure illustrates the scale of future commitments supporting today's investment.
CoreWeave's committed customer contracts represented 98% of revenue in Q3 2026.
Its FY2025 10-K describes these contracts as generally structured on a take-or-pay basis: customers commit to paying for specified compute capacity regardless of whether they actually use it, provided CoreWeave meets its contractual obligations.
The company also reports receiving upfront payments under some committed contracts.35
Take-or-pay arrangements are important to the economics of GPU infrastructure. They separate the provider's contracted revenues from the customer's actual utilization, giving the provider greater revenue visibility while transferring some demand risk to the customer.
However, they also create dependence on a relatively small number of large counterparties.
In Q3 2026, CoreWeave's two largest disclosed customer concentrations were 30% and 24% of revenue, or 54% combined. The filing anonymizes these customers, so these percentages should not be assigned to specific named companies.3
If contractual revenues support financing, the creditworthiness and performance of those customers become part of the financing risk.
Different economic clocks
There is also a mismatch worth examining between the lives of the underlying assets and the obligations supporting them.
CoreWeave's filings provide several reference points:
| Economic commitment | Disclosed timeframe |
|---|---|
| Computing equipment | Six-year estimated accounting useful life |
| Initial committed customer contracts | Generally one to six years |
| Operating leases | Twelve-year weighted-average remaining term at September 30, 2026 |
| Delayed-draw financing | Facility-specific maturities and repayment schedules |
Sources: CoreWeave FY2025 10-K5 for equipment useful life and initial customer contract terms; Q3 2026 10-Q3 for lease terms and financing. These are company-wide measures, not matched terms for an individual GPU deployment.
Accounting life isn't necessarily economic life. Nor do all customer contracts, leases and loans finance the same assets.
But the different time horizons explain why the ability to renew contracts, redeploy equipment and maintain revenues after initial commitments expire is important.
Long-term contracts reduce uncertainty. They don't eliminate it; they determine who carries which parts of it.
4. Compute is already being sold forward
A long-term capacity agreement has much in common with a physically delivered forward sale.
A customer agrees today to purchase access to future compute. The provider undertakes to deliver that capacity, often at predetermined commercial terms.
The commitment gives the customer access and price certainty. For the provider, it supports capital investment and may secure financing.
But these agreements aren't equivalent to standardized exchange-traded forwards.
A capacity contract can bundle:
- A specific hardware and cluster configuration.
- Delivery dates, duration and location.
- Availability and performance guarantees.
- Customer-specific service terms.
- Credit exposure to the contracting parties.
Its contractual rights also may not be freely assignable or resalable. Some arrangements permit capacity resale or reassignment under defined conditions, but that is different from trading a standardized instrument.
Consider a customer that commits to 1,000 GPUs for three years and discovers after twelve months that it needs only half of them.
The customer may remain obligated to pay for the full commitment. Meanwhile, another buyer might be willing to purchase that unused capacity.
Connecting them is possible in principle, but the contract may not permit resale, the hardware may not suit the second buyer, or the provider may be unwilling to accept the resulting delivery and credit arrangements.
The opportunity is not to invent future compute commitments. It is to make existing commitments easier to reallocate and their risks easier to manage.
Transferability is only one potential improvement. Better price discovery, standardized specifications and the ability to hedge rental-price exposure independently of physical capacity could also be valuable.
Whether the benefits justify the added complexity is an open commercial question.
5. From bilateral contracts to financial markets
Other industries have developed financial markets around physical services that are not perfectly interchangeable.
Freight offers an instructive precedent.
Different vessels, routes and contractual arrangements produce different shipping costs. Nevertheless, the freight industry developed benchmark assessments published by the Baltic Exchange and derivatives that settle against those indices.
Forward Freight Agreements allow participants to hedge changes in freight rates without necessarily chartering a vessel or physically transporting cargo under the financial contract.6
The analogy with compute isn't exact. Freight benefits from established physical trading conventions and considerable market history. Compute faces its own challenges involving hardware generations, performance and capacity availability.
But freight demonstrates that perfect physical fungibility isn't always necessary for financial risk transfer.
CME's proposed compute futures
This transition from physical contracts to financial price exposure is already underway.
In 2026, CME Group partnered with Silicon Data to develop cash-settled futures referencing the rental prices of NVIDIA H100 and B200 GPUs.10
The proposed contracts have straightforward specifications:
| H100 futures | B200 futures | |
|---|---|---|
| Contract code | GPU1 | GPU2 |
| Reference index | Silicon Data H100 Rental Index | Silicon Data B200 Rental Index |
| Contract size | 730 GPU-hours | 730 GPU-hours |
| Settlement | Financial, based on rental indices | Financial, based on rental indices |
| Listed maturities | 36 consecutive months | 36 consecutive months |
Both indices measure hourly on-demand rental prices from neocloud providers, excluding hyperscalers.
The contract size represents a standardized month of GPU capacity rather than the actual number of hours in each calendar month.
As of November 18, 2026, these contracts have not launched. CME's originally planned October 5 launch was postponed, with regulatory review extended through January 8, 2027.11
Questions surrounding benchmark integrity and potential manipulation are particularly relevant when financial products settle against prices from an emerging, relatively opaque underlying market.
The significance of CME's proposal isn't that it makes physical GPU capacity transferable. A futures position doesn't entitle its holder to receive GPUs.
Instead, it separates rental-price risk from the underlying obligation to deliver or consume compute.
A GPU provider could sell futures to protect against falling rental rates. A customer could buy futures to offset rising procurement costs. Neither would need to transfer or reserve physical capacity through the exchange.
The contract unit also highlights the difference between financial and physical capacity. Eight H100 futures contracts represent the same 5,840 GPU-hours as one continuously rented eight-GPU node over a standardized month. A 256-GPU cluster represents 256 such contracts by quantity.
That does not mean the economic exposures are equivalent. A dedicated training cluster has networking, availability and performance characteristics that a single-GPU rental benchmark may not capture.
Whether the futures become useful at scale depends on how closely the settlement indices track real commercial exposures, and whether sufficient trading liquidity develops.
Investment in market infrastructure
CME isn't the only participant attempting to develop this market. On November 17, 2026, Compute Desk announced a $45 million seed round led by Lightspeed Venture Partners to develop GPU pricing indices, trading infrastructure and technology intended to support physical settlement of compute contracts. The company is working with exchange partners pursuing regulatory approval for compute futures and options.14
The distinction is useful: CME's proposed contracts transfer price exposure through cash settlement, while Compute Desk is also pursuing infrastructure to connect financial positions with deliverable computing capacity. The funding demonstrates investor interest in building these markets, although neither funding nor proposed contracts establish that meaningful liquidity already exists.
What needs to be standardized?
The first requirement is a credible reference price.
Silicon Data already publishes GPU rental benchmarks, including an H100 rental index, with observations adjusted for differences such as rental terms, cluster scale and interconnect.7
These indices provide a reference point, not necessarily a price at which a particular customer can execute a transaction.
Consider a GPU operator selling dedicated H100 clusters under multi-year agreements.
The operator's revenues depend on more than the number of GPUs supplied. Customers may be paying for high-speed interconnects, guaranteed availability, dedicated networking and the ability to run large training workloads across the entire cluster.
An index based on short-term rentals of individual H100 GPUs may move differently from the price of those dedicated clusters.
The resulting basis could be substantial, even when both products use the same hardware. Similar problems arise across GPU generations, locations and contract durations.
The second requirement is genuine two-sided demand.
GPU owners have an economic interest in protecting future rental revenues against falling prices. Compute consumers may want protection against rising costs. But their desired contract sizes, tenors and underlying products may differ.
Finally, there is the dealer.
A dealer offering a fixed-price hedge must be able to offset the exposure, find an opposing customer, secure suitable physical capacity, or retain the position on its balance sheet.
If the hedge is imperfect or liquidity disappears, the dealer remains exposed.
The commercial viability of a compute derivatives market may depend less on the ability to design a contract than on the ability to hedge it.
CME's initiative and Compute Desk's plans provide concrete tests of that proposition. Standardized contract specifications, reference indices and clearing infrastructure are necessary building blocks, but they don't by themselves establish a liquid market or eliminate the mismatch between short-term rental prices and long-term capacity agreements.
6. Geography: an introduction to basis risk
Compute can be consumed remotely, but its infrastructure is local.
Data centres depend on electricity, land, cooling and network access. These vary considerably by location.
Northern Virginia remained the largest of CBRE's eight primary North American data-centre markets by inventory in H1 2026, at approximately 4.5 GW. Across those markets, a record 7,481.1 MW was under construction, with Atlanta overtaking Northern Virginia in construction activity.8
These figures describe data-centre electrical capacity, not installed GPU capacity. A megawatt of available IT power cannot be converted into a fixed number of GPU-hours without knowing the hardware configuration, infrastructure efficiency and utilization.
The location of future capacity increasingly reflects the availability of power rather than simply proximity to traditional technology hubs.
There is an interesting comparison with Bitcoin mining, another industry whose economics are heavily influenced by electricity.
Bitcoin miners can often interrupt production in response to electricity-market conditions. Cambridge's 2025 mining-industry report recorded 888 GWh of curtailed electrical load among surveyed operators during 2023.9
Some AI workloads may offer similar flexibility, but the distinction matters: a mining operation generally produces a standardized output for a global network, while AI workloads have different networking, latency, continuity and data requirements.
For geographically flexible workloads, cheaper compute in one region can potentially substitute for more expensive capacity elsewhere. For tightly interconnected training clusters or latency-sensitive inference, relocation may be impractical.
The difference is especially pronounced for training workloads that require frequent communication between GPUs. A cluster's accelerators may need to exchange large quantities of data continuously, making high-speed interconnects and physical proximity essential to performance.
By contrast, a self-contained batch workload might move between regions with relatively little penalty, assuming the data can be transferred and suitable hardware is available.
Regional GPU prices may therefore reflect both differences in production costs and differences in the usefulness of the capacity available.
How easily workloads can move between locations will help determine how closely regional prices converge—and how much geographic basis risk remains.
That will be the subject of the next research piece.
7. Where the research goes next
The compute economy already contains substantial markets for future capacity, supported by commitments connecting consumers, infrastructure operators and lenders.
The next questions concern how efficiently those commitments allocate capacity and distribute risk.
I'll be exploring several connected themes:
Geography and portability. Where is compute being built, what determines its location, and when can workloads move economically between regions?
Ownership and financing. How are hardware and infrastructure financed, and what happens when customer contracts, debt and asset lives don't align?
Price discovery. How are rental prices determined, and what makes a benchmark representative of an actual commercial exposure?
Natural flows. Who has a genuine reason to buy or sell future compute, and which risks are already managed through bilateral contracts?
Dealer hedging. How can intermediaries manage outright price exposure, geographic and hardware basis, liquidity and delivery risk?
I'm interested in whether more transparent and transferable markets can improve on the contractual arrangements that already exist. Not every risk needs a derivative, and not every bilateral contract benefits from becoming tradable.
The goal of this research is to understand where financial markets genuinely improve the allocation of capital and risk, and where the existing structures are better suited to the problem.
If you're working in compute infrastructure, cloud services, procurement, financing or emerging compute markets, I'd welcome the opportunity to learn more about your industry.
Sources and further reading
- Core Scientific — Q3 2026 Form 10-Q
- CoreWeave — Termination of proposed Core Scientific acquisition, October 2025
- CoreWeave — Q3 2026 Form 10-Q
- NVIDIA — CoreWeave investment, January 2026
- CoreWeave — FY2025 Form 10-K
- Baltic Exchange — Freight futures settlement
- Silicon Data — H100 Rental Price Index
- CBRE — North America Data Center Trends, H1 2026
- Cambridge — Digital Mining Industry Report, 2025
- CME Group — Compute Futures, Contract Specifications
- CFTC — Designated Contract Market Product Filings and Approval Status
- NVIDIA — HGX H100 Architecture and Eight-GPU Server Configuration
- CoreWeave — Cloud Pricing
- Lightspeed Venture Partners — Why We Invested in Compute Desk, November 17, 2026