Your AI capacity contract is a capital decision
B2B blog post, about 1,200 words. C-suite audience, AP style, sourced from public filings and releases only. Written as a portfolio sample for a cloud and SaaS content agency.
Most finance teams meet AI infrastructure through a line item. A vendor quote arrives, someone in engineering explains that the model needs more compute, and the conversation becomes a negotiation over price per hour.
That framing undersells the decision. The contract in front of you is not a procurement agreement with a monthly meter attached. It is a multi-year capacity commitment with a counterparty whose ability to deliver depends on construction schedules, utility interconnection, and a supply chain it does not control. Those commitments now run long enough, and large enough, to show up in your credit profile rather than your operating budget.
The companies on the other side of the table have started describing the same contracts that way. Oracle reported $664 billion of remaining performance obligations at Aug. 31, 2026, up $209 billion year over year, and said it put 850 MW of additional data-centre capacity into service in the quarter (Oracle Q1 FY27 release, Sept. 10, 2026). On the landlord side, TeraWulf reports about $27 billion of contracted revenue across 839 MW of contracted critical IT capacity, on leases it describes as 10 to 20 years and credit-backed (TeraWulf Q3 2026 investor release, Nov. 6, 2026). Individual commitments are now large enough to matter to a balance sheet rather than a monthly budget.
You do not need to be a lender to borrow their diligence. You need to ask the questions a lender asks before it writes the check.
1. The commitment outlives your depreciation schedule
Your finance team depreciates servers over five or six years. The building they sit in lasts decades, and the lease that pays for the building is increasingly written for 15 to 20 years, often with renewal options tacked on.
That mismatch is not a technicality. It means the contract you sign this quarter constrains capital allocation well past the planning horizon of the strategy that justified it. Before you commit, put the full term on one page: total obligation, annual obligation, the year each escalation lands, and the exit price in each year. If the exit price in year three is larger than the benefit you expect in year three, you have bought a decade of optionality you cannot afford to use.
2. Risk starts when the hall is delivered, not when the contract is signed
Read the commencement clause before you read the price. In TeraWulf's 20-year lease with Anthropic at its Hawesville, Kentucky campus, covering about 401 MW of critical IT load, the tenant's obligation to pay rent begins on delivery of each individual data hall rather than at signing or at financial close. Google's credit support, which stands behind the tenant, only becomes effective at the commencement date of the final data hall under each of the three leases (TeraWulf Form 8-K, Nov. 4, 2026).
Two sentences there matter to a customer as much as to a lender. First, a phased delivery schedule means your capacity arrives in tranches, so your model needs to know which workloads run on tranche one and what happens to the business case if tranche two slips two quarters. Second, the strength of the credit behind your contract can change over the life of the deal. Ask what happens to your pricing if the counterparty's support mechanism steps down.
3. Take-or-pay is a floor for them and a ceiling for you
A take-or-pay clause protects the provider's revenue. It does not protect your ability to use what you have paid for.
The question to ask is not whether you can reduce volume. It is what you can do with capacity you have paid for and cannot fill: resell it, sublease it, burst it to another region, or hand it to an affiliate. In a market where inference pricing moves faster than contract terms, secondary-use rights are worth real money, and they are usually negotiable early and immovable later.
4. Counterparty risk has moved to the contractors
When a provider tells you the campus will be ready in a given quarter, the answer depends on an engineering, procurement and construction contractor, a utility and a supply chain. At Hawesville, Fluor is the EPC contractor, and the contract is described as a guaranteed maximum price basis, with the contractor responsible for cost overruns only up to a percentage of its profit. Past that line, the risk returns to the owner.
You are not underwriting the contractor. You are pricing the probability that a slip becomes your problem. Ask what remedies exist if delivery is late, whether service credits begin on the scheduled date or the actual one, and how much notice you get before a slip is disclosed. A provider that answers those questions with a schedule and a contract clause is a provider worth signing with.
5. The renewal option is a pricing decision you have already made
Renewal options look like flexibility. They are usually a hedge written by the party that expects to need it. Two 5-year options at the tenant's election, in a market where power costs and land costs have both moved, is a valuable right that you should either price or trade away rather than accept as boilerplate.
Model the option at a stress case, not a base case. If power prices move 30%, if the local utility raises capacity charges, if a competitor's capacity comes online next door, what does renewal cost you and who bears it?
What to ask for before you sign
- The delivery schedule by hall, in writing, with the service-credit trigger attached to the scheduled date.
- The full payment schedule including escalators, pass-throughs and any utility-related charges the provider can pass on.
- Secondary-use and resale rights, stated as permissions rather than prohibitions.
- Termination and exit pricing, year by year, for the full term.
- Disclosure of who carries construction risk past the contractor's cap, and what happens to your pricing if it crystallises.
None of these requests is unusual. All of them are cheaper to negotiate before signature than after.
The point is not that AI capacity contracts are dangerous. It is that they have become large enough, long enough and structured enough to deserve the same treatment as any other capital commitment: a term sheet you can defend to your board, a counterparty you have actually diligenced, and an exit you can afford. The companies selling capacity already think this way. The companies buying it should too.
Sources: Oracle Corp., Q1 fiscal 2027 results release (Exhibit 99.1 to Form 8-K), Sept. 10, 2026; TeraWulf Inc., Form 8-K, Nov. 4, 2026, and Q3 2026 investor release (Exhibit 99.1), Nov. 6, 2026; Fluor Corp., contract award release, Nov. 4, 2026. Figures are quoted as published. No estimate in this piece is presented as a reported number.