On 29 September 2026, Reuters reported that Anthropic expected to spend at least $518 billion on AI infrastructure over a decade. The scale is exceptional. The management issue is familiar: a long-term commitment should be assessed against the conditions under which the business might need to change course.

What happened

Citing a confidential IPO prospectus it had reviewed, Reuters said about 80% of the planned spending was either non-cancellable or payable regardless of usage. Anthropic’s stated rationale was the need to secure computing capacity for future demand. The report concerned commitments and expectations, not a claim that the full sum had already been spent.

Securing scarce capacity can have strategic value. So can retaining flexibility. The useful question is how a company weighs those benefits against obligations that continue when its assumptions change. The answer depends on its demand, alternatives and ability to absorb a different outcome.

Why it matters

Most businesses will never negotiate infrastructure at this scale, but they encounter the same structure in smaller AI decisions. A discounted annual commitment can look attractive beside a monthly price. A broad licence bundle can appear efficient compared with individual purchases. The comparison is incomplete if it ignores adoption uncertainty and the cost of leaving.

Imagine a company buying agent capacity before a pilot has established how much of the process can run reliably. If the workload later proves smaller, the unused commitment may remain. If the preferred model changes, moving could require new integration work alongside the original bill. These are planning scenarios, not predictions about Anthropic’s agreements.

Procurement should therefore compare usable outcomes rather than headline unit prices. Estimate a low, central and high usage case. Include implementation, supervision, migration and support. Ask which obligations survive a decision to stop the project and which assets the organisation can retain or transfer.

The bigger shift

AI strategy has a timing problem. Waiting too long can delay learning; committing too early can bind the business to assumptions it has not tested. A staged approach helps distinguish the cost of learning from the cost of operating at scale.

An early pilot should create evidence about demand, quality and the surrounding workload. A later commitment can then reflect what that evidence supports. Larger commitments may still be justified, but the decision should state why capacity certainty is worth the reduced flexibility and which downside the business is prepared to carry.

This is part of choosing an AI strategy before accumulating tools. The strategy should explain the dependencies it creates, including data access, contractual terms, specialist skills and integration. An exit plan is useful even when nobody intends to leave, because it makes the real cost of dependency visible.

My take

Before approving a substantial AI commitment, I would ask for one page describing a change-of-plan scenario. Suppose adoption is slower, a workflow is withdrawn or a better option becomes available. What would the organisation still owe, what could it move, and how long would a transition take?

The answer should come from the people responsible for the agreement and the operating process, with specialist advice where required. A generic assurance that data can be exported says little about whether the service can actually be replaced.

Then record the trade-off honestly. A business may choose to pay for certainty, accept a dependency or reserve capacity ahead of demand. Those can be deliberate strategic choices. They become harder to defend when the commitment was approved as a simple efficiency saving and nobody examined the conditions under which the saving disappears.