Portability of AI Compute Infrastructure in AI Acquisitions

In acquisitions of AI companies, the strategically valuable assets are not solely the models themselves but the AI compute infrastructure arrangements that power them. Contracts for AI infrastructure now routinely represent billions of dollars in committed capacity. For any acquirer of an AI-centric business, whether these agreements survive a change of control has become a threshold question, rather than a secondary diligence item, that determines whether the deal delivers its intended value.

This Legal Update examines the key legal and practical considerations around the portability of AI compute infrastructure contracts in M&A transactions, drawing on our experience advising on technology transactions, outsourcing arrangements, data center and cloud services, and AI-focused acquisitions. For convenience, this Legal Update refers to these arrangements as “AI compute contracts,” but the compute capacity can (depending on the business) come from agreements with a hyperscaler, a specialized GPU cloud provider, owned or leased hardware through a co-location provider, or a combination of these models. 

Why AI Compute Contracts Matter

AI companies depend on substantial, and often growing, quantities of computing power to train and run their models and serve customers. Unlike ordinary enterprise software and IT services agreements, AI compute contracts have several features that make them especially significant in the M&A context:

• Scale and commitment: AI compute contracts may involve multi-year commitments, minimum spend obligations, take-or-pay capacity, prepayments, or hardware financing exposure. For larger AI businesses and infrastructure providers, those obligations can be central to the company’s operating model and transaction value.
• Pricing advantages: Some AI companies have secured favorable pricing, reserved capacity, or allocation rights for scarce GPU or other accelerator resources. These contractual advantages may not be replicable at current market rates, making the contracts themselves a source of competitive value.
• Technical and physical dependencies: AI workloads may be tied to a provider’s architecture, including proprietary chip designs and custom networking and storage configurations. Switching providers may require significant engineering work, migration planning, and interim operating arrangements.
• Capacity scarcity: GPU and accelerator capacity remains constrained. Even as new providers enter the market and capacity expands, losing access to committed capacity for a production AI workload may not be remediable on comparable terms within commercially reasonable timeframes.

For these reasons, an AI compute contract that terminates, becomes terminable, loses reserved capacity, triggers repricing, or requires a costly migration upon a change of control can erode the value of the acquired business. In the extreme case, it can leave the acquirer with a target stripped of the infrastructure that justified the price.

Assignment and Change-of-Control Provisions

The starting point for any analysis is the assignment, change-of-control, and transfer language in the relevant AI compute contract. In our experience, the following patterns are common in material cloud, GPU, co-location, and related infrastructure contracts:

• Anti-assignment clauses: AI compute contracts often prohibit assignment without the provider’s prior written consent and typically deem a non-conforming assignment void from the outset. The decisive question in M&A is whether a deal structure that practitioners would ordinarily regard as falling outside such language (for example, a stock acquisition or a reverse triangular merger) nonetheless constitutes an “assignment” within the meaning of the clause.
• Deemed assignment provisions: Many contracts go further and specifically provide that a change of control of the customer, whether by merger, stock purchase or otherwise by operation of law, is deemed to be an assignment requiring consent. Where such language exists, the contract may trigger a consent requirement even if the acquisition structure would not otherwise constitute an assignment under default rules.
• Consent not to be unreasonably withheld: The presence or absence of a “not to be unreasonably withheld” qualifier is critical, and it frequently dictates how the consent process unfolds. Where consent may not be unreasonably withheld, the customer has a basis to challenge a refusal or undue delay; where consent rests in the provider’s “sole discretion,” the customer may have limited leverage and the provider may seek concessions as the price of cooperation.

Due Diligence Considerations

Buyers should conduct targeted diligence on AI compute contracts early in the deal process. Key questions include:

• What does the target’s compute stack actually include? Map the target’s infrastructure by provider, workload, GPU or accelerator type, region or facility, committed capacity, renewal date, and migration path. The objective is to determine whether the target depends on a single hyperscaler agreement, multiple GPU cloud providers, owned or leased hardware in co-location, or informal capacity reservations.
• What are the assignment and change-of-control provisions? Obtain and review the full text of each material AI compute contract, including all amendments, order forms, service-specific terms, co-location exhibits, and side letters. Do not rely on management summaries.
• Does the transaction structure trigger a consent requirement? Analyze whether the proposed deal (stock purchase, merger, asset sale) constitutes an “assignment,” “change of control,” or other consent-triggering event under the specific contractual language. The analysis can also turn on whether the contract includes a license of intellectual property from the provider, or on jurisdiction-specific authority regarding whether a reverse triangular merger effects an assignment by operation of law. The full contents of the compute contracts, including the governing law provision, should be reviewed alongside the assignment language.
• What happens if consent is not obtained? Identify the consequences: termination rights, pricing resets, loss of committed capacity or reserved instances, loss of credits, acceleration of minimum commitments, or renegotiation triggers. Pay particular attention to minimum commitment acceleration provisions. If the compute agreement provides that the provider may accelerate the full remaining committed spend upon a non-consented change of control, the buyer faces an immediate and potentially significant cash liability, on top of the loss of future capacity.
• Are there competitor restrictions? Some AI compute contracts restrict assignment to, or use by, competitors of the provider (or affiliates of those competitors). If the buyer or its affiliates competes with the AI infrastructure provider, consent may be refused on legitimate commercial grounds.
• Is the pricing and capacity package transferable? Even where the agreement itself may remain in place, pricing, credits, reserved capacity, hardware allocation, support level, or committed-use discounts may be subject to adjustment upon a change of control; for example, by moving from negotiated enterprise rates to standard published pricing.
• What is the remaining term, committed spend, and full scope of compute-related financial obligations? A compute contract with 18 months remaining and $50 million in committed spend presents a different risk profile than one with five years remaining and $2 billion in commitments. How do the specifics impact your transaction model? Importantly, buyers should look beyond the four corners of executed AI compute contracts. Targets may have secured compute capacity through binding letters of intent, capacity reservations, pre-payment arrangements, or off-balance-sheet financing structures that do not appear as straightforward contracts on a schedule of material agreements.
• Are there technical and physical lock-in considerations? Assess the degree to which the target’s AI workloads are architecturally or physically dependent on a specific provider’s infrastructure. This affects both the cost of migration and the practical alternatives if the contract is lost or reprised.
• Are there export-control-based legal restrictions? If the compute is deployed outside the United States, US government authorization may be required before assignment of remote or physical access may occur.
• What is the hardware generation, remaining useful life, and obsolescence risk of the allocated infrastructure? Diligence not only the commercial terms of the AI compute contract, but also the specific hardware generation and remaining useful life of the GPU or accelerator infrastructure allocated to the target. A favorable price on aging infrastructure may be less valuable than it appears if the buyer will need to migrate to newer hardware shortly after closing.
• Are the data, security, and use restrictions compatible with the target’s business? Review restrictions on regulated data, customer data, training data, model weights, inference outputs, logging, benchmarking, security testing, prohibited AI uses, and provider access to telemetry or usage data. The buyer should confirm that the target’s use of the infrastructure is consistent with its commitments to customers and applicable law.
• Are the service levels and remedies adequate? Review provisioning obligations, delivery milestones, uptime commitments, maintenance windows, service credits, suspension rights, support commitments, disaster recovery, data durability, and limitations of liability. Ordinary service credits may be inadequate if failure to deliver reserved capacity impairs the deal thesis.

Purchase Agreement Considerations

Where AI compute contracts are material to the value of the target, the transaction documentation should address the following:

• Representations and warranties: Ensure that material AI compute agreements are included in “Material Contracts,” which should cover notices of defaults, termination or intention not to renew. The target should also confirm that: (i) the consummation of the transaction will not require the consent of any AI infrastructure provider, give rise to any termination right, or result in any loss of capacity, acceleration of committed spend or change to pricing or other material terms under any compute agreement, in each case except as set forth in the disclosure schedules; and (ii) it is in compliance with all minimum commitment obligations under the AI compute contracts.
• Pre-closing covenants: Between signing and closing, and in addition to using commercially reasonable efforts to obtain any consents required under the material AI compute contracts and not amending, terminating, waiving or accelerating any material right or obligation under any such agreement, the target should be required to: (i) promptly notify the buyer of, and provide copies of, any communication from an AI infrastructure provider concerning the transaction, and, subject to compliance with applicable law, afford the buyer a reasonable opportunity to participate in substantive discussions with the provider; (ii) continue to satisfy minimum commitment and consumption obligations and maintain reserved capacity in the ordinary course of business consistent with past practice; and (iii) not agree to any condition or undertaking proposed by a provider as the price of its consent that would impose new economic or operational obligations on the business following closing without the buyer’s approval.
• Closing conditions: In transactions where a specific AI compute contract is fundamental to the deal thesis, consider making receipt of the provider’s consent a condition to closing. This is aggressive, and of course sellers will resist it as introducing third-party deal risk, but it may be appropriate where the relevant relationship is essential to the acquired business.
• Price reduction or indemnification: Where consent is not a condition to closing, or where the risk of termination, non-renewal, or repricing is accepted, buyers should negotiate specific provisions allocating losses arising from those events. Depending on materiality, the parties may address the risk through a purchase price adjustment, dedicated escrow, special indemnity, covenant to obtain consent, or post-closing migration plan.

Practical Takeaways

AI compute contracts can be a central value driver in AI acquisitions. A few practical points for M&A practitioners:

• Treat material AI compute contracts as key operational contracts, alongside significant customer agreements, intellectual property licenses, data use agreements, and critical supply arrangements. They deserve dedicated diligence attention from the outset and tailored representations, warranties, covenants, and closing deliverables.
• Do not assume that a stock acquisition or reverse-triangular merger avoids the consent requirement. Standard anti-assignment language, express change-of-control language, deemed-assignment language, competitor restrictions, governing law, and any embedded IP license all need to be reviewed separately.
• Engage with the relevant AI infrastructure provider early where possible. Hyperscalers, neoclouds, co-location providers, and equipment lessors may have different consent processes, timelines, commercial leverage, and operational requirements. Early engagement can reduce timeline uncertainty.
• Consider whether the AI compute contract is a source of value or merely a necessary input. If favorable pricing, committed capacity, hardware allocation, or co-location rights are a primary driver of deal value, the diligence workstream and contractual protections in the purchase agreement must reflect that.
• Assess the hardware generation underlying the target’s compute allocation, and whether the contract’s pricing advantage reflects genuinely durable value or merely a legacy rate on aging infrastructure. A contract that guarantees access to hardware approaching obsolescence may be a depreciating asset rather than a competitive moat.
• Diligence the target’s full compute-related financial exposure, including off-balance-sheet commitments. Capacity reservations, pre-payment arrangements, residual value guarantees, and lease obligations that have not yet commenced can represent material future cash commitments that are invisible in standard financial statements. These obligations survive a change of control and may accelerate upon one.
• Remember that compute scarcity is not permanent. Market dynamics are shifting as hyperscalers expand capacity, neoclouds grow, co-location providers build AI-ready facilities, and GPU pricing evolves. But in the near term, the inability to replace lost capacity on comparable terms within a commercially reasonable timeframe remains a real deal risk.

As AI companies continue to scale and infrastructure commitments grow, these issues will only become more central to the M&A process. We expect assignment, change-of-control, portability, and infrastructure-continuity issues in AI compute contracts to receive the same level of scrutiny that has traditionally been reserved for key intellectual property licenses, data use agreements, key customer contracts, and other core operating assets—and M&A transaction documentation should evolve accordingly.