Neocloud Contracts: How AI Compute Is Sold

In our previous article, we looked at how the AI boom created neoclouds, specialized cloud providers built around graphics processing units (GPUs). We closed by pointing out that contracts worth hundreds of billions of dollars have already been signed, even though no standard documentation exists for them yet, and almost every agreement is still written from scratch.

This article covers what neoclouds actually sell, why training a model and running inference (answering user requests and generating responses) look so different to a buyer, and what contracts the parties end up signing. To see how these contracts work in practice, we walk through six large, publicly disclosed deals between Nebius, CoreWeave, IREN, and WhiteFiber and their customers – Meta, Microsoft, OpenAI, and Nscale.

All six deals were put together by U.S. lawyers, and we take them on their own terms. How other legal systems would treat these structures is outside the scope of this article. Every agreement we discuss is public on EDGAR, the filing database of the U.S. Securities and Exchange Commission (SEC), and we link to each one.

Throughout this series, we use “provider” and “neocloud” interchangeably.

What Customers Buy from Neoclouds

Neoclouds sell time on their GPUs. The basic billing unit is the GPU-hour – one GPU working for one customer for one hour. Neocloud customers are AI companies, and they typically need compute for two things: (1) training new models and (2) running inference. Each places different demands on the infrastructure and on the contract, so it’s worth looking at them separately.

Training AI Models

Training is how an AI model learns from a massive dataset. The model processes text, images, or code over and over, adjusting billions of internal parameters until its predictions are accurate enough. Under the hood, most of this is matrix multiplication. GPUs can run huge numbers of identical operations in parallel, which makes them better suited to the job than general-purpose central processing units (CPUs).

Training a frontier model takes infrastructure on an enormous scale. These models train on clusters of thousands of GPUs (tens or even hundreds of thousands in the newest projects) that work in lockstep and constantly trade intermediate results. That traffic runs over high-speed interconnects. NVIDIA NVLink, for example, ties together GPUs within a server or rack, while InfiniBand and high-speed Ethernet link servers and racks to one another. The hardware is also packed very densely, which places heavy demands on power and cooling. A single NVIDIA DGX GB200 NVL72 rack holds 72 GPUs, draws about 120 kW, and needs liquid cooling for its compute components. A standard server rack draws 10–15 kW.

Two economic features of training directly shape the contract:

1. The load is constant and indivisible. A training run keeps every GPU in the cluster running flat out for weeks or months, so there is no spare capacity to lend to another customer.
2. The load is finite. Once the run ends, the cluster sits idle, and the neocloud has to find something else for it to do.

That shapes what each side wants. The neocloud wants to sell capacity years in advance, ideally before it is even built. The customer wants a guarantee that a scarce cluster will be ready when it needs it and, for training, that it can run that cluster at full capacity.

Inference

Inference is what a trained model does once it is in use – handling user requests and generating responses. It is measured in tokens (not to be confused with crypto tokens). A token is a chunk of text about four characters long, or roughly three-quarters of a word in English. Providers usually price inference per million input tokens (the text sent to the model, including the prompt and context) and per million output tokens (the text the model sends back).

Inference has the opposite load profile. First, demand is broadly steady and doesn’t stop when training ends. As long as people use the product, the model answers requests around the clock, and the load grows as users multiply and agentic workflows spread. Second, inference doesn’t need thousands of GPUs working in sync. Requests from different users are spread across servers and batched together, so how hard each GPU works depends more on the provider’s software than on the physics of the cluster.

The table below sums up the key differences between training and inference.

By 2030, McKinsey estimates, inference will be the single largest data center workload, at 93 GW versus 62 GW for training. Some neoclouds already see the shift. On CoreWeave’s first-quarter earnings call in May 2026, the CEO said that well over half of the company’s capacity was going to inference.

Buying AI Compute: Commercial Models and Contract Structures

Buyers of AI compute care about two things: (1) the commercial terms on which they get capacity and (2) what the provider actually commits to. The first comes down to price, term, and availability. The second covers what exactly the contract is for, who maintains the equipment, and who is on the hook for outages.

Four Commercial Models

The labels below are not an industry-wide classification. Terms differ from provider to provider, and reservations and long-term commitments are often combined in a single deal. Still, for comparison, it helps to sort compute purchases into four models:

1. On-demand. The customer spins up whatever capacity is available when it needs it and pays for the time it holds that capacity, usually with no long-term commitment. This suits experiments and short jobs, but there is no guarantee that enough GPUs will be free on a given date. A billed hour also isn’t necessarily an hour of full GPU use, because capacity can sit idle after it is allocated.
2. Spot or preemptible. The customer usually pays less, but the provider can reclaim the capacity on agreed terms. The discount comes with the risk that a job gets cut off, so this model works for tasks that can wait or pick up again from a saved checkpoint. “Spot” here refers to interruptible cloud capacity, not a trade on an exchange.
3. Reserved. The customer locks in an agreed amount of capacity for a fixed term. It usually commits to pay for that capacity whether or not it uses it, and in return gets predictable access at a set price. The contract spells out the specific availability guarantees and payment rules.
4. Committed contract. This is a custom multiyear contract with a firm commitment to pay. The parties agree on a large block of capacity, a delivery schedule, and payments for several years out. The provider may buy hardware or build a new cluster just for this customer. These contracts are often take-or-pay, and once the capacity is delivered as agreed, the customer pays for the committed volume even if it does not use it. Late delivery and unavailable capacity are handled separately. Economically, a committed contract is often just a large custom reservation rather than a fundamentally different product.

What the Contract Covers

The same commercial terms can be papered in different ways. A single project, or even a single contract, can mix different levels of access to the infrastructure, so the lines are not always sharp. Even so, the industry relies on three main structures:

1. Compute services (GPU-as-a-Service). The neocloud gives the customer remote access to compute in an agreed configuration and runs the underlying infrastructure as the contract requires. That compute can come from a shared pool, from dedicated physical servers, or from a separate cluster. The provider maintains the hardware, while the service scope determines who sets up the operating system and software stack and who runs the jobs. What the provider ultimately promises is working compute that meets the agreed specs and service levels. The hardware remains under the provider’s control. The customer never touches it, and if a server goes down, the provider swaps in an equivalent one.
2. Equipment lease. The customer gets the right to use specific servers or GPUs for an agreed term, but doesn’t own them. What the customer is paying for is the hardware itself, not what it produces. In the basic setup, the customer finds a place to host the equipment and runs it, handling maintenance and configuration in-house or bringing in a contractor. The contract spells out what equipment is delivered and in what condition, how it can be used, who pays for repairs, and how it is returned at the end of the term. The customer usually has physical access to the servers, subject to the rules of the facility where they sit.
3. Colocation. The data center operator provides space, power, cooling, and agreed ancillary services. The customer brings its own or leased equipment and runs it. For an extra fee, the customer can also use “smart hands” services, where data center staff rack servers, run cables, or swap out components. The operator is responsible for the facility’s infrastructure, not for whether the customer’s servers and software work.

These structures can be combined. A neocloud may, for example, lease its hardware, house it in a third-party data center, and sell compute services to its own customers. What really separates them is what the provider promises and how the parties split responsibility. Accounting and depreciation are a separate matter.

Who Neoclouds Sell To

Contract terms depend heavily on who is on the other side of the table. Hyperscalers (Microsoft, Meta, Google, and Amazon) turn to neoclouds for the capacity they cannot build fast enough themselves, but they do it for different reasons. Microsoft resells the capacity through Azure, so its contracts extend usage rights to its own customers. Meta runs its own models on the capacity and contracts as an end user. Model developers such as OpenAI and Anthropic have enormous demand but weaker credit, which is why their deals come with equity in the provider and special solvency covenants. Enterprises and startups buy on providers’ standard terms, governments buy under confidential agreements, and industry players buy from each other. The neocloud Nscale, for example, leases data halls from the data center operator WhiteFiber. In September 2025, NVIDIA committed to buy all of CoreWeave’s unsold capacity through April 2032, a commitment that starts at $6.3 billion. In effect, NVIDIA guaranteed demand for one of its own largest chip buyers.

The Contract Stack

The market is young, but its contracts already follow a consistent three-tier structure. Each tier is defined by its role in the deal, not by the title on the document.

The first tier sets the general terms: liability, confidentiality, data, termination, and governing law. This may be the provider’s master services agreement, as with CoreWeave, or the customer’s master agreement for vendors, as in Meta’s contracts with Nebius. It may also be the customer’s standard purchase terms, which nobody signs but which are incorporated into the deal by reference, as with Microsoft’s Purchase Order Terms. Whatever the form, the job is the same.

The second tier covers the specific deal: what is being bought, how much, where, when, and at what price. These documents go by different names, such as Order Form, Service Order, or Statement of Work (SOW). The real difference is length, which depends on how much the document has to spell out and how generic the general terms are. Microsoft’s purchase terms, for example, are written broadly enough to fit any supplier, so its compute SOWs run longer than other customers’ deal documents.

The third tier is the exhibits: the service level agreement (SLA), acceptance criteria, security requirements, acceptable use policy, and data processing agreement. These are modules, not standalone contracts, and they get attached wherever is most convenient. CoreWeave’s SLA is an exhibit to its master agreement, while in Nebius’s deal with Microsoft the SLA is attached to the SOW.

How do we know what these contracts say? Companies listed in the United States must disclose their material contracts by filing them as exhibits to their reports. U.S. issuers do this under Item 601(b)(10) of Regulation S-K, and foreign issuers such as Nebius under the Form 20-F rules. Since 2019, companies have been allowed to redact confidential information that is not material to investors. In practice, they black out prices, volumes, and sites, but the legal architecture (risk allocation, liability, and termination rights) stays visible.

Now that we have covered how training differs from inference, how compute is bought, and how those purchases are documented, we can turn to specific contracts that listed companies have made public.

A Closer Look at Six Neocloud Deals

1. Nebius and Meta: A Master Agreement Drafted by the Customer

Nebius and Meta have signed two agreements: (1) a Cloud Infrastructure Services Agreement dated November 1, 2025 (EDGAR), and (2) an Infrastructure Services Agreement dated March 13, 2026 (EDGAR). The first was announced as a five-year deal worth about $3 billion. The second covers $12 billion of dedicated capacity on next-generation Vera Rubin chips starting in 2027, plus Meta’s commitment to buy up to $15 billion of additional capacity.

Both agreements are built on Meta’s standard vendor template. Nebius is named as the “Vendor,” Meta’s policies are attached as exhibits, and capacity is ordered in batches and delivered in tranches. If something goes wrong with one tranche, only that tranche can be terminated, and the rest of the deal stays in place.

When capacity goes down, Meta receives service credits rather than damages. Credits are capped at a set percentage of the monthly fee, and anything above the cap rolls over to the following months. This keeps the provider’s cash flow steady, which matters to its lenders as much as to the provider itself.

The agreement expressly treats the arrangement as a service, not a lease, and leaves the depreciation deductions with Nebius. Meta’s rights are structured as an intellectual property license so that Section 365(n) of the U.S. Bankruptcy Code applies. That provision lets a licensee keep its license if the licensor goes bankrupt.

In the second agreement, the financing section overrides every other provision. The parties agree to run the deal through a bankruptcy-remote special purpose vehicle, because that is “a minimum pre-requisite of lenders.” From day one, the customer contract is written to serve as collateral for a future loan.

2. Nebius and Microsoft: A Statement of Work Under Microsoft’s Purchase Terms

The deal rests on a Statement of Work dated September 7, 2025 (EDGAR), and Amendment No. 1 dated January 21, 2026 (EDGAR). According to Nebius, the contract is worth $17.4 billion through 2031 and could grow to $19.4 billion. The capacity comes from a data center in Vineland, New Jersey.

The SOW runs half the length of the Meta agreements, yet it covers more ground. Nearly half of it rewrites 25 provisions of Microsoft’s standard purchase terms. The rest sets out nine GPU tranches, the price per GPU-hour, a payment schedule, and liquidated damages. The fee is calculated by multiplying the number of GPUs by the price per GPU-hour and by the number of hours in the term. Microsoft pays an advance on each tranche and, once a tranche is accepted, owes the fee “irrespective of the actual usage.” This is a classic take-or-pay clause, which some readers will recognize from natural gas supply contracts.

Acceptance works much like it does in construction. The provider validates the capacity and sends a readiness notice, Microsoft runs its own tests, and silence counts as acceptance. If delivery is late, the provider owes liquidated damages equal to a portion of the daily fee for each day of delay. Once a grace period runs out, Microsoft may reject the tranche and get its upfront payment back.

Three provisions stand out. First, the agreement takes effect only once Nebius raises financing or confirms that it will pay for the build-out itself. Second, the contract expressly lets Nebius assign it and its payment rights to lenders as security, and Microsoft agrees to sign an acknowledgment with them. Third, Microsoft may terminate if Amazon or Google takes control of Nebius. Liability is capped at one-tenth of the total fees for all services.

3. CoreWeave and Microsoft: A Master Agreement Drafted by the Provider

The parties work under a Master Services Agreement dated February 22, 2023 (EDGAR), governed by New York law. The agreement itself sets no total value. That depends on the orders Microsoft places. Microsoft matters enormously to CoreWeave. It accounted for about 62% of CoreWeave’s revenue in 2024, according to the IPO prospectus, and 67% in 2025, according to the annual report.

What sets this deal apart is that the master agreement is CoreWeave’s own template, not Microsoft’s. In 2023, CoreWeave was one of very few suppliers that could offer GPU capacity at the scale Microsoft needed, and that leverage let it sign the world’s largest corporate buyer on its own terms. The contract reflects this. CoreWeave owns all the infrastructure, keeps the rights to anonymized usage data, and sells capacity as Reserved Instances or prepaid Bulk Credits.

Liability is capped at six months of fees (12 months for data and confidentiality breaches), and there are no liquidated damages for delay. If a GPU server fails (a “Hardware Failure” in the contract’s terms) or a force majeure event hits, and the outage lasts more than two business days, the customer simply stops paying for the affected capacity until it comes back. A regional power shortage counts as force majeure and excuses the provider. In Nebius’s contracts with Meta, by contrast, the provider remains on the hook for the power it has contracted for.

If the customer pays late, CoreWeave may repurpose or resell the servers reserved for it and apply the proceeds to the debt. What the customer owes is tied to the gap between the price it paid and the on-demand price. That is the first time a market price shows up in these documents.

4. CoreWeave and OpenAI: A Model Developer as the Customer

The deal is built on a Master Services Agreement – Bare Metal Environment dated March 7, 2025 (EDGAR), a related $350 million stock issuance agreement (EDGAR), and a second, similar agreement dated May 8, 2025 (EDGAR).

The initial deal was announced at $11.9 billion, and later expansions took it to about $22.4 billion. It uses the same CoreWeave master agreement, adapted for a customer without a hyperscaler’s balance sheet. CoreWeave provides the capacity as bare metal, which means OpenAI gets entire physical servers with no virtual machine layer in between and installs its own software environment on them.

One notable provision is the Alternate Operator Option in Section 10(c). If CoreWeave goes bankrupt, OpenAI can require that all rights under the order, the master agreement, and the data center contract be transferred to a pre-selected third-party operator within two business days. The protection runs both ways. OpenAI must tell CoreWeave if it becomes unable to pay or is at risk of becoming unable to pay within the next four months.

The agreement has another unusual provision. Under Section 11(b), if CoreWeave went public, it had to issue $350 million of its stock to OpenAI at the IPO price. It did so under a separate stock issuance agreement right after its IPO in March 2025. In effect, the provider gave its customer a discount in the form of equity. Customers taking stakes in their suppliers is a hallmark of this market.

5. IREN and Microsoft: Same Template, Different Provider

The deal consists of a Partner Statement of Work dated November 2, 2025, with Microsoft’s standard purchase terms attached (EDGAR), a server purchase agreement with Dell (EDGAR), and an agreement making that order non-cancellable (EDGAR). Unlike the other contracts, this one leaves the figures unredacted. The total value is $9.67 billion, upfront payments come to $1.93 billion (exactly 20% of each tranche), and the capacity arrives in four tranches of NVIDIA GB300 chips with terms running through 2031.

IREN, a former bitcoin miner, signed the same kind of contract Nebius had signed two months earlier. Putting the two side by side shows how the hyperscaler’s template is evolving.

Instead of liquidated damages, for example, the contract provides for Delay Credits and makes them Microsoft’s sole remedy. The delivery deadline moves if NVIDIA ships the chips late, as long as the provider can show it ordered them on time. That shifts the risk of a late shipment from the chipmaker onto the buyer.

Microsoft also got an option to swap the GB300s for next-generation chips, with the price adjusted by a formula. Under a step-in right, IREN must get letters from its lenders and data center operators that allow Microsoft, if IREN goes bankrupt, to make the payments on IREN’s loans and colocation agreements itself.

6. WhiteFiber and Nscale: One Layer Down

A WhiteFiber subsidiary and the British neocloud Nscale signed a Master Services Agreement dated November 22, 2025 (EDGAR), governed by North Carolina law. According to WhiteFiber’s announcement, the agreement runs for an initial term of 10 years and covers 40 MW of critical IT capacity for about $865 million, with a 3% annual escalator. That figure doesn’t include power, which is passed through to the customer.

This deal is not about compute at all but about the space to house it. Nscale gets the exclusive right to place its own GPUs in data halls at a North Carolina data center, while WhiteFiber supplies power, cooling, and connectivity.

Legally, Nscale does not lease the space but receives a license to use it, and under U.S. real property law that difference matters a great deal. A lessee holds a property interest in the premises, which survives a sale of the building and is protected against the owner’s creditors. A licensee has only a personal, contractual right to be on the premises and keep its equipment there. The agreement makes the choice explicit, stating that the customer “is a licensee and not a tenant or lessee” and acquires no interest in real property (Section 1(a)). That leaves the operator free to finance or sell its property without the customer’s rights getting in the way.

Fees are pegged to electrical capacity, and Nscale pays a minimum power charge no matter how much power it actually draws. As in the other deals, Nscale may assign the agreement to its lenders, either outright or as security, and WhiteFiber must sign equipment access agreements with those lenders.

What Comes Next

In this article, we looked at what neoclouds sell, the commercial models and contract structures they use, and how six large contracts filed on EDGAR are put together.

If the six contracts share one takeaway, it is that the GPU-hour price in each of them comes out of private, bilateral negotiations. None of them points to an external market benchmark, and the parties did not simply overlook it. The first GPU rental price indexes appeared in 2025, but there is still no generally accepted market price for AI compute. Contracts worth tens of billions of dollars get signed without reference to the market because, in the exchange sense, there is no market yet.

That is about to change. Compute increasingly looks like an exchange-traded commodity, and the first futures on GPU rental price indexes are on the way.

In the next articles in this series, we will explain how these indexes are built, how GPU rental futures will work, and how they could change the deals that finance neocloud infrastructure.