Opening Stakes
If your company is about to commit to a multi-year cloud and AI compute agreement, the number worth modeling is not the monthly bill. It's what you'll pay to leave.
That cost is set right now, in clauses most founders skim: how you get your data and your trained models out, whether your price is actually locked, what happens to your committed spend if the business pivots, and what you're owed when the provider can't deliver. None of those are discovered at exit. They're decided at signing.
Major cloud and AI partnership agreements show how contractual commitments can increase switching costs. In January 2025, Federal Trade Commission staff reported that these arrangements included billions of dollars in cloud-spending commitments, along with restrictions that could increase switching costs. Growth-stage companies should check their own provider agreements for less visible versions of the same mechanisms.
So this is a contract-drafting problem before it's a risk-register line. The good news is that each of these risks can be addressed before signing, through the contract, the product selection, the deployment architecture, or the amount of commitment you accept. Sometimes the answer is choosing a different service or refusing to place a nonportable workload there.
Why Compute Lock-In Has Become a Contract Issue
Three developments matter.
The first is FTC staff identifying contractual mechanisms that could increase switching costs. Cloud-spending commitments and related restrictions show how lock-in can arise from provisions negotiated at signing, rather than only from technical decisions made during deployment.
The second is competition authorities mapping the structure of the market. A November 2025 OECD Secretariat background note identified concentration, entry barriers, and switching barriers as important features of AI infrastructure markets. On 17 December 2025, France's competition authority published a study finding that electricity runs an estimated 30 to 50 percent of a data center's operating costs. That makes energy pass-through a live pricing variable, not a footnote.
The third is regulators targeting specific contractual and technical barriers to switching. In the EU, the Data Act generally limits the notice period for initiating a switch to two months and phases out switching charges entirely beginning 12 January 2027 (Articles 23 through 30 of Regulation (EU) 2023/2854). In the UK, the Competition and Markets Authority closed its cloud market investigation on 31 July 2025, finding that egress fees suppress switching. It also found committed-spend agreements widespread but, in their current form, not competition-harming. On 31 March 2026, after engagement with the CMA, Microsoft and Amazon announced steps to lower egress fees and improve interoperability. The CMA continued assessing whether those actions would sufficiently benefit customers. These actions show that egress and interoperability mechanics are not operationally immutable, although they do not establish what any particular buyer can obtain in a private negotiation.
The Four Clauses to Redline
Four clause families often drive much of the exit cost.
1. Data and trained-model portability on exit
Many agreements are particularly thin here, and it's where AI buyers have the most to lose. A standard clause says your data comes back "in a usable format." That sentence does almost no work. Usable to whom, in what schema, and including which layers?
For customer data, define the export scope, format, timing, assistance, and cost. Name the format; the EU's statutory phrasing, "structured, commonly used and machine-readable," is a clean benchmark. Name the scope: raw records, derived and enriched data, logs, and metadata. Then set a timeline in days, an assistance obligation, and who pays.
For customer-created or customer-specific model artifacts, the questions change. If you fine-tuned a model on the provider's infrastructure, ask what the service can technically export: which checkpoints, adapter weights, embeddings, and vector indexes, and in what format. Then ask what ownership and reuse rights the agreement actually gives you in each.
Keep four things separate: a contractual export right, the technical ability to export, ownership and reuse rights, and the difference between receiving files and receiving materials you can actually run. None of this reaches the provider's proprietary base-model weights; the subject is the artifacts you created or paid to create. The mistake is assuming your model is portable because your data is. Put it in the agreement, not in a support ticket you file the week you're trying to leave. This is considerably easier to address before deployment than during migration.
2. Pricing and pass-through
A locked unit price for the initial term is not the same as a locked cost. Two mechanisms move your real number: renewal repricing and pass-through. The French authority's finding that energy can run 30 to 50 percent of a data center's operating cost makes pass-through a priced risk, not a rounding error. The decision point is whether your agreement caps renewal increases and either excludes or bounds energy and other pass-through charges. If the agreement does neither, pass-through provisions can move your costs during the term, while an uncapped renewal provision can reprice the deal later, potentially when switching is most difficult. Finance should monitor both as separate exposure categories.
3. Commitment flexibility
Committed-use discounts trade price for flexibility, and the value of the flexibility you give up can exceed the discount if demand, architecture, or ownership changes during the term. The FinOps Foundation, a standards body, documents the mechanics: with spend-based commitments, if your consumption falls short, you pay the committed amount anyway. Some reservation-based commitments can be modified, exchanged, transferred, or sold through a provider-supported marketplace. Spend-based commitments generally cannot. An over-commitment can therefore become unused spend you still owe through the end of the term.
If an acquisition is plausible during the term, read the commitment together with the assignment and change-of-control provisions. Confirm whether the buyer or surviving entity can assume the agreement and unused commitment without provider consent, repricing, or termination.
One honest counterpoint. If your workload is stable and your forecast is reliable, committed spend is not a trap; the discount can be real money and the flexibility you trade is flexibility you would not use. The point is to price what you're giving up and keep an exit that works.
4. Performance failure, capacity shortage, and termination relief
Where accelerated-compute capacity is constrained, this clause family can be especially important. Many provider forms say relatively little about what happens when reserved capacity is unavailable. Four drafting objectives close the gap.
First, define unavailable capacity. The agreement should say what counts as unavailable or undelivered capacity, because a provider and a buyer will read a shortage differently.
Second, reduce what you owe when the provider cannot deliver. Your commitment's shortfall term runs one direction: you pay when your own consumption falls short. Form contracts tend to stay silent on the mirror case, where the provider is the one who misses. Say expressly whether committed spend is reduced during a provider-caused shortage, and whether prepaid or unused commitment is refunded rather than forfeited.
Third, preserve remedies beyond service credits. Many provider forms make service credits the exclusive remedy, capping the provider's exposure at a fraction of a monthly bill no matter what a shortage costs you. Decide whether credits are your only recourse or a floor above which other remedies survive.
Fourth, allow termination and migration assistance after chronic failure. A pattern of underdelivery should create a penalty-free exit, so leaving a provider that broke the deal doesn't cost you a termination charge. Termination assistance, the provider's obligation to help move your data and workloads out, should survive termination itself, so the duty doesn't expire the moment you need it.
One related claim to read skeptically: several providers announced in 2024 that they "eliminated egress fees." As the CMA's cloud work documented, those waivers generally apply only when you leave the platform entirely, often require account termination within a fixed window, and don't touch day-to-day transfer. Do not treat a benefit announced on a provider's blog as a negotiated contractual right unless it is incorporated into the agreement or otherwise made binding.
The Diligence Consequence
The four clauses also shape how outsiders read you. A sole-provider, multi-year commitment with no meaningful portability terms can look like concentrated, unpriced dependency. A buyer may reflect that dependency in valuation, integration planning, or the amount of flexibility it believes it is acquiring.
The specifics are the ones you have already met: unused committed spend, assignment and consent restrictions, migration feasibility, and how much stack flexibility survives the deal. The switching-cost review a board runs after a term sheet lands is the same review you could run before you sign, while you can still change the terms. For an investor or board, that makes it a diligence item: ask to see a portfolio company's portability, pricing, and commitment terms, and read them as concentration risk. Not every compute commitment reduces valuation or impairs a transaction, but an unpriced one invites the question.
A Four-Question Pre-Signing Screen
Four questions leadership and counsel can answer in a room, before the business terms harden. Each maps to one clause family, and the drafting action sits under it.
1. Can we leave with usable data and artifacts?
Draft for: a named format, a defined scope, a day-count timeline, and an assistance obligation in place of "usable format," plus confirmation of what model artifacts the service can export and what rights you hold in them.
If unanswered: you discover at exit that your data returns in a form you cannot use and your models do not come out at all.
Owner: engineering, with product.
2. Can our price move during the term or at renewal?
Draft for: bounds or objective formulas for in-term pass-through charges, and a cap or defined methodology for renewal pricing.
If unanswered: pass-through can move your costs during the term, and renewal can reprice the deal later, when your leverage is lowest.
Owner: finance, with legal.
3. What reduces what we owe?
Draft for: the shortfall and shortage terms read together, and a commitment made transferable through the assignment and change-of-control provisions.
If unanswered: an over-commitment or a pivot leaves you paying for capacity you no longer use.
Owner: finance and operations, with legal.
4. Can we exit if the provider cannot deliver?
Draft for: defined undelivered capacity, committed-spend reduction during a provider-caused shortage, remedies beyond service credits, and penalty-free termination with migration help after chronic failure.
If unanswered: a provider that misses still collects, and you cannot leave.
Owner: operations and legal, with finance.
Closing Perspective
Lock-in is a contract term before it's a risk-register line. That reframing is the whole game: it moves the fix from a project you run under pressure later to a redline you make from strength now. Your bargaining power with the provider is front-loaded. It lives in the window between the term sheet and the signature, and it drains from the moment your data and models sit on one provider's infrastructure.
What I keep coming back to is how one-sided the timing is. The EU and UK developments provide concrete benchmarks for switching notice, egress charges, and interoperability. They do not govern a typical US agreement or establish what a particular customer can negotiate, but they make the relevant exit mechanics easier to identify and evaluate. Waiting has a price: the renewal you can't walk away from and the model you can't extract. Price the cost to leave before you sign.
This article is for informational purposes only and does not constitute legal advice. Every company's situation and technology stack are different, and qualified counsel should review the applicable agreement before the company makes material contracting or migration decisions.
