Compute Offtake Agreements

Compute offtake agreements are how neocloud providers sell GPU usage ("compute") to customers. With offtake agreements in-hand, operators have the revenue visibility to raise GPU-backed debt, negotiate colocation terms, and order hardware. This article draws on compute offtake master services agreements (MSAs) filed publicly with the SEC by CoreWeave, IREN, SharonAI, Nebius, Bit Digital, and others, along with American Compute's review of private offtake agreements. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

An offtake agreement is the combination of MSA + Order Forms

Every compute offtake deal is structured as two layers.

1. The MSA sets the rules: liability caps, termination rights, SLA framework, etc.
2. The Order Form details the money: exact GPU type and count, pricing, term length, payment schedule, upfront deposits. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Not every provider calls it an "MSA." IREN's $9.7 billion deal with Microsoft is structured as a "Partner Statement of Work," ^{[9]IREN Limited / Microsoft, Partner Statement of Work, SEC Form 8-K (2025)https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_8k.htm} and Nebius's $2.9 billion deal with Meta is filed as a "Commercial Agreement." ^{[11]Nebius / Meta Platforms, Commercial Agreement, SEC Form 6-K (2025)https://www.sec.gov/Archives/edgar/data/1513845/000110465925109803/tm2530882d2_6k.htm} The underlying structure is the same: a framework agreement governing all future orders.

The specific financial terms that matter the most and get negotiated the most, always live in the Order Forms.

The benefit of the split is that the MSA can be negotiated once, and the Order Forms can be negotiated many times. This is crucial in enterprise sales, where procurement teams and legal counsel need to review MSA terms heavily, but not as often as the Order Forms terms.

The MSA continues until all Order Forms have expired or been terminated. There is no fixed end date on the MSA itself. A single customer can sign multiple Order Forms over time, each with its own term. Microsoft had at least five separate Order Forms with CoreWeave, representing approximately $10 billion in committed spend between 2023 and 2030 and 62% of CoreWeave's 2024 revenue. ^{[2]CoreWeave, Inc., S-1 Registration Statement, SEC filing (2025)https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001769628&type=S-1&dateb=&owner=include&count=40}

Most MSAs allow customer affiliate entities (subsidiaries and related companies under the same corporate umbrella) to enter their own Order Forms, bound by the same MSA as if they were an original party. ^{[1]From American Compute's review of private offtake agreements and similar contracts} This matters for international expansion. Microsoft Corp. signs the MSA with CoreWeave in the US, and Microsoft Singapore Pte Ltd. can place its own Order Forms under the same agreement without renegotiating the framework.

How the deal process works

The sales process for a compute offtake deal follows standard enterprise sales: the neocloud's sales team qualifies the buyer, solutions engineers scope the deployment (GPU type, cluster size, interconnect, software stack), and procurement negotiates pricing, term length, and commitment model.

For new customers, the buyer's legal team then reviews the MSA, which is typically the bottleneck. Once terms are agreed, both sides sign the Order Form, the buyer wires a deposit, and the provider provisions hardware.

First-time customers can expect three to six months from first call to live deployment. Existing customers adding a new Order Form move in weeks, because the MSA is already signed.

How pricing and commitment work

Compute offtake agreements sell compute capacity in three models: Reserved Instances, Bulk Credits, and on-demand. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Commitment models ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Reserved Instances is when the customer gets exclusive access to a specified hardware configuration, a specific number of GPU nodes with a specific GPU type, interconnect, and memory. The customer pays the full Reserved Instance fee whether it uses the capacity or not. This is also commonly known as "take-or-pay". Reserved Instances require a deposit, typically 30-60 days before the contract takes effect. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Bulk Credits are a committed-spend model. The customer buys credits in advance, at a cheaper rate than on-demand, and use them as needed. Unused credits expire at term end and are not refundable. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

On-demand still requires the customer to sign the MSA and set up an account with the provider. Enterprise buyers can't just go to a website and start spinning up GPUs; their legal and procurement teams need to review the vendor's terms first. The difference from reserved or bulk credits is that the customer makes no minimum spend or term commitment. On-demand is billed monthly at the provider's published rates, which can change at any time without notice.  ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Depending on the contract, the hourly rate can be fixed at the beginning of the session or variable. For example, if the rate in the first hour is $3.00 per GPU-hour and moves to $3.50 per GPU-hour in the next hour, some providers may bill you $6.00 in total, while others might bill you $6.50 in total. ^{[1]From American Compute's review of private offtake agreements and similar contracts} Bit Digital's SEC filings show contracted GPU cloud rates of $2.15 to $3.40 per card-hour for H200 GPUs. ^{[12]Bit Digital, Inc., Annual Report on Form 10-K, SEC filing (2025)https://www.sec.gov/Archives/edgar/data/1710350/000101376225000307/ea0233723-10k_bitdigit.htm}

Reserved Instances are the focus of large neocloud providers. HPE required Soluna to pay $10,293,350 before the 36-month, $34.3 million contract became effective. ^{[3]HPE / Soluna AL CloudCo, HPE GreenLake GPU Infrastructure-as-a-Service + HPC & AI Cloud Services Agreement, SEC Exhibit 10.105 (2024)https://www.sec.gov/Archives/edgar/data/64463/000149315224032123/ex10-105.htm} That is 30% of total contract value paid before a single GPU-hour was delivered. IREN's Microsoft deal requires 20% of each tranche's value paid before the delivery date, credited against fees due after the 24th month. ^{[9]IREN Limited / Microsoft, Partner Statement of Work, SEC Form 8-K (2025)https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_8k.htm} The upfront payment helps the provider and provider's lenders de-risk the customer, and goes towards funding hardware procurement.

The scale of these commitments is large. OpenAI committed approximately $11.9 billion over five years to CoreWeave, with CoreWeave agreeing to issue OpenAI $350 million in stock at its IPO price, a compute-for-equity arrangement embedded in the services contract. ^{[4]CoreWeave / OpenAI, Master Services Agreement, Bare Metal Environment, SEC Exhibit 10.24 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525052207/d899798dex1024.htm}

These commitment levels are not unique to CoreWeave. IREN committed $9.7 billion in GPU capacity to Microsoft through 2031, ^{[9]IREN Limited / Microsoft, Partner Statement of Work, SEC Form 8-K (2025)https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_8k.htm} Nebius signed a $2.9 billion deal with Meta over five years, ^{[11]Nebius / Meta Platforms, Commercial Agreement, SEC Form 6-K (2025)https://www.sec.gov/Archives/edgar/data/1513845/000110465925109803/tm2530882d2_6k.htm} and SharonAI signed a $1.25 billion MSA with ESDS Software Solutions to deploy 8,200 NVIDIA B300 GPUs in Australia. ^{[10]SharonAI Holdings / ESDS Software Solutions, Master Services Agreement, SEC Form 8-K (2026)https://www.sec.gov/Archives/edgar/data/2068385/000182912626001479/sharonai_8k.htm}

Committed contract value ^{[4]CoreWeave / OpenAI, Master Services Agreement, Bare Metal Environment, SEC Exhibit 10.24 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525052207/d899798dex1024.htm [2]CoreWeave, Inc., S-1 Registration Statement, SEC filing (2025)https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001769628&type=S-1&dateb=&owner=include&count=40 [9]IREN Limited / Microsoft, Partner Statement of Work, SEC Form 8-K (2025)https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_8k.htm [11]Nebius / Meta Platforms, Commercial Agreement, SEC Form 6-K (2025)https://www.sec.gov/Archives/edgar/data/1513845/000110465925109803/tm2530882d2_6k.htm [10]SharonAI Holdings / ESDS Software Solutions, Master Services Agreement, SEC Form 8-K (2026)https://www.sec.gov/Archives/edgar/data/2068385/000182912626001479/sharonai_8k.htm}

Payment terms protect the provider's cash flow

Payment terms in compute offtake agreements protect the provider's cash flow, and are important to get right for both the provider and the provider's lenders.

Net terms typically range from net-10 to net-60, with discounts for faster payment. This means that the customer has 10-60 days to pay the invoice after issued. Many contracts bill in advance, with monthly payments duebefore the upcoming month's service begins. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

A no-offset clause is critical. Amounts owed by the customer cannot be withheld or offset for any reason. If the customer has a billing dispute, a service complaint, or a pending claim, they must still pay the invoice on time. They can dispute separately, and get a refund in the future, but they need to always pay their invoices to get the service.

Some contracts require additional credit support beyond deposits. SharonAI's $1.25 billion MSA with ESDS requires $140 million in letters of credit or bank guarantees as customer security, on top of monthly advance payments. ^{[10]SharonAI Holdings / ESDS Software Solutions, Master Services Agreement, SEC Form 8-K (2026)https://www.sec.gov/Archives/edgar/data/2068385/000182912626001479/sharonai_8k.htm}

Suspension is the provider's primary enforcement tool. For unpaid fees, the seller can deactivate access and suspend running instances after providing written notice and giving the customer a cure period (a defined window, typically 30 days, to fix the problem before consequences kick in). ^{[1]From American Compute's review of private offtake agreements and similar contracts} The customer remains on the hook for all fees during suspension, and their bill will continue to grow.

Repeated non-payment triggers harsher consequences. Some contracts allow immediate termination of the entire agreement, if the customer is late on payment a specified number of times within a specified period (both number and period are negotiated in the Order Form). ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Payment escalation: invoice to termination ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Compute buyers do have some protections. For example, the Stale-invoice clause protects the customer. This clause waives any obligation to pay invoices received more than 120 days after the date they should have been sent. ^{[5]CoreWeave / Microsoft, Master Services Agreement, SEC Exhibit 10.23, CoreWeave S-1 filing (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798dex1023.htm} If you are a provider, you need to invoice on time.

What the SLA commits the provider to

The service level agreements (SLAs) define what the provider owes when the infrastructure goes down. Uptime targets, credit formulas, claim deadlines, and escalation triggers vary across contracts, but the structures are consistent. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Service credits, not cash refunds, are the standard way to compensate the customer for downtime. The provider typically offers credits equal to the duration of validated outages. They usually apply only toward future purchases, which means it cannot be used to discount previous invoices.

Customers must submit claims within a defined window, typically 30 days to two months after the incident. ^{[1]From American Compute's review of private offtake agreements and similar contracts} This is clearly worse for the customer, because the interruption in AI workloads likely cost more in business delays than the cost of the GPU-hours. That's why customers should always pick more reputable providers, over less reliable providers with more generous SLAs.

Hardware-specific SLAs are included in some contracts. Some specify that if a GPU node continues to be broken for multiple consecutive business days, the customer isn't charged for that hardware until it is restored. ^{[5]CoreWeave / Microsoft, Master Services Agreement, SEC Exhibit 10.23, CoreWeave S-1 filing (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798dex1023.htm}

The distinction matters. Under a standard credit SLA, the customer still pays for broken hardware upfront and gets refunded in service credits for future purchases, not cash. Under a hardware-specific SLA, the customer simply owes nothing for that node until it is back online. The credit SLA approach is worse for the customer: they are out the cash, and the credits may not be worth much if they are already planning to wind down usage.

Sustained SLA failure can trigger full termination. Some contracts include a form of Material Adverse Change (MAC): if the provider breaches the SLA for a sustained period, the customer can terminate and receive a refund. The HPE/Soluna contract sets this threshold at six consecutive months.  ^{[3]HPE / Soluna AL CloudCo, HPE GreenLake GPU Infrastructure-as-a-Service + HPC & AI Cloud Services Agreement, SEC Exhibit 10.105 (2024)https://www.sec.gov/Archives/edgar/data/64463/000149315224032123/ex10-105.htm} The exact duration is negotiated per deal.

SLA exclusions protect the provider from liability for events outside of their control: force majeure (events beyond either party's control, such as natural disasters, war, or government actions), customer-caused incidents, scheduled maintenance, internet connectivity issues beyond the provider's infrastructure, and third-party software. Customers with past-due accounts are also typically excluded from credits. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Liability caps and indemnification

Liability caps set the maximum a provider can owe their customer, no matter how badly things go. The standard structure is three-tiers: ^{[1]From American Compute's review of private offtake agreements and similar contracts}

1. a general cap (maximum the provider has to pay the customer) at the fees the customer paid and owed in the six months before the event
2. an elevated cap for more serious security and data breaches at twelve months of fees, and
3. uncapped for extreme liabilities like willful misconduct, gross negligence, IP infringement indemnification, and confidentiality breaches.

Liability cap tiers on a $10M/month contract ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Consequential damages are excluded. These include lost revenue, lost goodwill, business interruption. If the customer loses business because the provider's infrastructure was misconfigured, that is not something the provider is responsible for. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Some liabilities are explicitly treated as direct damages inside the caps even though they look consequential: government fines, breach notification costs, credit monitoring for individuals whose data was exposed (up to one year), and legally required mitigation. These are not excluded by the consequential damages waiver. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Indemnification answers a simple question: if a third party sues, who pays? These clauses are a standard artifact of enterprise software and technology contracts, carried over into compute agreements.  ^{[1]From American Compute's review of private offtake agreements and similar contracts}

The provider covers two categories. First, IP infringement: if the provider's platform uses patented technology without a license (say, a proprietary networking protocol or orchestration software), and the patent holder sues, the provider pays, not the customer. Second, security incidents: if a breach in the provider's infrastructure exposes customer data, the provider covers the resulting claims.  ^{[1]From American Compute's review of private offtake agreements and similar contracts}

If an IP infringement claim comes in, the provider has three options: modify the service to remove the infringement, obtain a license for continued use, or terminate the Order Form on 30 days' notice and refund prepaid fees for the remainder of the term. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

The customer indemnifies the provider for their own misuse. If the customer uses the provider's infrastructure for illegal activity or violates the acceptable use policy (AUP), the customer covers the resulting damages.  ^{[1]From American Compute's review of private offtake agreements and similar contracts}

Termination and early exit

One of the biggest concerns for providers and their lenders is what happens to the remaining contract balance if the customer tries to terminate the contract early.

If the customer breaches the contract and fails to fix it within the cure period, the provider can collect the full remaining balance. The customer pays all remaining fees for the entire remaining term. ^{[1]From American Compute's review of private offtake agreements and similar contracts} The provider can also transfer or sell the freed-up compute to other customers, but the proceeds of those would offset what the customer owes.

If the provider breaches first and then the customer terminates, the provider will have to refund only the prepaid fees for the remainder of all affected Order Forms. The asymmetry is intentional: the customer has to pay the rest of the contract, the provider only has to refund the prepaid portion.

Some contracts explicitly restrict termination for convenience. SharonAI's ESDS contract prohibits either party from terminating for convenience during the first 36 months. ^{[10]SharonAI Holdings / ESDS Software Solutions, Master Services Agreement, SEC Form 8-K (2026)https://www.sec.gov/Archives/edgar/data/2068385/000182912626001479/sharonai_8k.htm}

Cure periods give both sides time to fix problems before termination takes effect. Typical cure periods are 30 to 45 days for material breach and non-payment. In practice, the most common trigger is late payment, but cure periods also apply to other material breaches such as a provider failing to meet deployment timelines specified in an Order Form or a customer violating usage restrictions.

Force majeure clauses vary significantly across contracts. Some contracts allow the customer to terminate an affected Order Form if force majeure continues for 30 or more consecutive days, with a full refund of prepaid fees not yet applied to delivered services. Other providers insist that the customer's payment obligations "may be delayed but are not excused." even in force majeure events.

Force majeure: how payment obligations vary by contract ^{[4]CoreWeave / OpenAI, Master Services Agreement, Bare Metal Environment, SEC Exhibit 10.24 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525052207/d899798dex1024.htm [6]CoreWeave / Meta Platforms, Master Services Agreement, SEC Exhibit 10.5 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000176962825000062/exhibit105.htm [7]CoreWeave / NVIDIA, Master Services Agreement, SEC Exhibit 10.31 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm}

Competitor acquisition triggers are common. If either party is acquired by a direct competitor of the other, the non-acquired party can terminate on written notice. For example, if CoreWeave is acquired by a direct competitor of Meta, Meta as a customer can terminate the contract on written notice. ^{[1]From American Compute's review of private offtake agreements and similar contracts} Some buyers go further: Microsoft reserved a unilateral right to terminate in the event of any CoreWeave acquisition, regardless of whether the acquirer is a competitor. ^{[5]CoreWeave / Microsoft, Master Services Agreement, SEC Exhibit 10.23, CoreWeave S-1 filing (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798dex1023.htm}

Customer restrictions providers will accept

Large offtake customers can impose restrictions that reach deep into provider's operations. These terms are standard across the contracts we reviewed. ^{[1]From American Compute's review of private offtake agreements and similar contracts}

• No subcontracting of compute. Customers prohibit or require prior written approval for subcontracting the performance of compute services. ^{[1]From American Compute's review of private offtake agreements and similar contracts} The customer wants assurance that their workloads run on the provider's hardware, not someone else's. OpenAI's MSA with CoreWeave explicitly prohibits subcontracting the bare metal environment without prior written consent.  ^{[4]CoreWeave / OpenAI, Master Services Agreement, Bare Metal Environment, SEC Exhibit 10.24 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525052207/d899798dex1024.htm}
• Publicity treated as material breach. The provider cannot disclose the customer's identity, the nature of the relationship, or use their name or logo in any advertising without consent. Any breach is treated as a material breach of the agreement, triggering termination rights. Meta's CoreWeave MSA treats any unauthorized public disclosure as a material breach.  ^{[6]CoreWeave / Meta Platforms, Master Services Agreement, SEC Exhibit 10.5 (2025)https://www.sec.gov/Archives/edgar/data/1769628/000176962825000062/exhibit105.htm}
• Customer's Data Protection Agreement governs. The customer's data handling framework, not the providers, controls how their data is processed. Microsoft required CoreWeave to comply with Microsoft's own Data Protection Agreement, not CoreWeave's standard terms.  ^{[5]CoreWeave / Microsoft, Master Services Agreement, SEC Exhibit 10.23, CoreWeave S-1 filing (2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798dex1023.htm}
• Feature removal notice. Providers must provide advance notice (often 12 months) before removing features or functionality that could materially affect the customer's use of the services. ^{[1]From American Compute's review of private offtake agreements and similar contracts}
• Capacity concentration cap. Some contracts cap the percentage of available capacity a provider can dedicate to a single customer. This protects the customer because they want to know the operator has resilient, multi-tenant infrastructure and isn't too dependent on their contract with them.

A couple of restrictions work in the provider's favor.

• Assignment to an affiliate "cannot be unreasonably withheld." Neoclouds typically sign the customer MSA under a parent entity but hold the hardware and raise debt in a separate entity. This clause lets the provider assign the Order Form to that entity so lenders can take a security interest (ability to seize the hardware).
• Benchmarking restrictions. Customers may be prohibited from disclosing performance or benchmarking results to third parties without the provider's prior written consent.