Four frontier AI labs — OpenAI, Anthropic, Google and xAI — each hold Defense Department agreements worth up to $200m to prototype decision-making tools, according to contracts obtained through freedom-of-information litigation. The reporting describes two provisions worth isolating: bidirectional data exchange including model benchmarks, and engineers embedded with the military.
Most of the discussion of these deals will be about whether frontier models should be used for military decision support, which is a serious question and not this one. This is about the structure of the arrangement, which is unusual and has consequences regardless of where anybody lands on the first question.
What a normal procurement looks like
An agency buys software. Requirements go out, vendors bid, one wins, the product is delivered and operated under a support agreement. Data flows from the vendor to the customer in the form of the product. If the agency wants to change supplier at renewal it faces migration cost — real, often large, and bounded by the data it holds and the interfaces it built.
The described structure differs in both directions at once. Data moves toward the supplier. People move toward the customer.
Each of those separately creates a tie. Together they create something closer to a joint operation than a purchase.
Why four contracts is not the same as four suppliers
Awarding in parallel to four labs looks like textbook competitive procurement, and at the award stage it is. Nobody is locked in, prices are disciplined, and the department gets to compare.
The question is what the position looks like at renewal, after two years of integration. Whichever vendor's models have been tuned against the department's benchmarks, whose engineers know the workflows, and whose formats the data has accumulated in, is not competing on equal terms with the other three any more. It is competing against its own switching cost, which is the strongest position a supplier can occupy.
This paper has written about the same dynamic elsewhere: enterprises spent a year making sure they could switch between model providers and discovered the switch ran through a small number of clouds and chips underneath, and buyers who cannot audit software have started auditing the vendor instead. Procurement keeps solving for the visible layer while the dependency forms below it.
The benchmark provision is the interesting one
Bidirectional exchange of frontier-model benchmarks is not a routine clause.
A benchmark is how a lab knows what its model can do relative to others. Benchmarks derived from a customer's actual operational problems are considerably more valuable than public ones, because public benchmarks are saturated and increasingly contaminated by training data. A supplier receiving evaluation signal from real military decision problems is receiving something it could not otherwise buy.
That is a benefit flowing to the vendor, and it should be priced. Whether it is priced — whether the $200m reflects the value of what goes back — is not something anybody outside can assess, and it is the sort of question that would ordinarily be asked at award rather than after litigation.
The embedded-staff question is older than this technology
Placing supplier personnel inside a customer is standard in defence, in large systems integration, and in enterprise software of a certain size. It is efficient and it is how difficult deployments actually succeed.
It also blurs a line that matters when things go wrong. Somebody has to sign off on what a model said, and an organisation whose expertise on a system sits with contractors from the company that supplied it has a harder time answering that question than one with the knowledge in-house.
What to watch
Not the next contract, which will be reported as a dollar figure.
Watch whether the renewal in 2027 is competed or extended. A competed renewal means the department preserved its optionality through two years of integration, which would be a genuine achievement and worth studying. A sole-source extension on the grounds of continuity means the structure did what structures like this do — and that it was foreseeable at signing, in a clause that was public the whole time and took a lawsuit to read.
The description of July 2025 Department of Defense agreements with OpenAI, Anthropic, Google and xAI, each worth up to $200m, to prototype military decision-making tools, and of provisions for bidirectional data exchange including frontier-model benchmarks and engineers embedded with the military, are as reported by The Intercept following freedom-of-information litigation and summarised in industry coverage during the week of 9 September 2026. This publication has not independently reviewed the contracts, and the specific scope, security controls and limitations in them are not described here. The observations about switching costs and procurement structure are general and are our own.
Topics technologyprocurementdefenceartificial intelligencecontracts
