Summary: TechCrunch reported on April 28 that Google agreed to let the US Department of Defense run its AI on classified networks under effectively all-lawful-use terms, becoming the third vendor after OpenAI and xAI. The contract reportedly includes statements against domestic mass surveillance and autonomous weapons, though enforceability remains unclear. About 950 Google employees signed an open letter urging comparable guardrails.
TechCrunch reported on April 28 — leaning on The Wall Street Journal and other outlets — that Google has agreed to let the U.S. Department of Defense run its AI on classified networks, under wording that in effect permits “all lawful uses.” That makes Google the third vendor to sign a defense AI contract since Anthropic turned down the Pentagon’s terms, with OpenAI and xAI ahead of it in line. When outlets sought comment, Google gave none.
Nothing here turns on technology; it turns on market structure. The largest buyer (the Pentagon) refuses to bend on unrestricted use, the most principled seller (Anthropic) already carries the “supply-chain risk” label, and so every AI company still in the game has to answer one question — the market, or the guardrails? Google answered by putting pen to paper.
Inside the Agreement: Classified Networks and “All Lawful Uses”
According to the WSJ, the deal clears the DoD to deploy Google’s AI tools in classified settings, which in practice means all lawful uses. The same report points out that the contract carries language saying Google has no intention of seeing its AI used for domestic mass surveillance or autonomous weapons — phrasing that mirrors OpenAI’s earlier defense contract. What the WSJ is equally upfront about is what that language is worth: whether such provisions carry legal force or could be enforced at all remains unclear.
Put plainly, this is a contract that says “in principle, anything is fair game — just keep away from these two things.” The Pentagon walks off with the unrestricted access it has pressed for throughout; Google walks off with a carve-out statement it can hold up as public-relations and legal padding.
Anthropic Got There First: What Refusal Cost
To read this deal properly, rewind to the sharpest standoff between the AI industry and government of the past two months. The Pentagon pressed every AI company to accept unrestricted-use terms; Anthropic held out for guardrails on domestic mass surveillance and autonomous weapons, and the talks broke down. In early March, the DoD designated Anthropic a “supply-chain risk” — a label ordinarily reserved for foreign adversaries. Anthropic sued at once, and in late March a judge ordered the designation frozen for now, with the litigation still moving.
Over the same stretch, the vendors willing to play picked up their tickets in order: OpenAI put its name to its own DoD agreement in early March (paying with internal controversy and an executive departure), and xAI’s Grok had worked its way into the Pentagon’s classified systems still earlier. Google’s move effectively casts the deciding vote in this tug-of-war — among the major frontier-AI suppliers, only Anthropic, mid-litigation, remains outside.
For vendors observing from the sidelines, the sequence doubles as a case study in procurement leverage. The buyer never talked Anthropic’s guardrails down; it simply re-labeled the seller. Once that instrument sat on the table, every later negotiation opened with whoever cared most about usage limits already a step behind.
Dissent at Home: An Open Letter with 950 Signatures
Google’s hallways are not quiet, either. An open letter at notdivided.org carries roughly 950 employee signatures, pressing the company to copy Anthropic’s playbook: no sale of AI to the DoD unless comparable guardrails come with it. Set against Google’s overall headcount, that is a small slice — but it keeps alive the long-running divide inside big tech over defense work, with Project Maven then and this standoff now, the script rarely changing. The letter’s framing openly reaches back to 2018, when employee protests led Google to let its drone-imagery contract lapse.
Reading the Procurement Market After This Deal
Three takeaways. One, guardrails are switching jobs — from “product feature” to “bargaining chip”: sit across the table from the single largest government customer and negotiating terms is negotiating market structure. Two, carve-outs that cannot be enforced carry roughly the weight of declarative text — and that is precisely why they may become the template for future government AI contracts, letting each vendor cooperate “principledly.” A subtler shift runs underneath: responsibility for misuse migrates from contract terms into stated intent. A vendor that “does not intend” mass surveillance is not a vendor that has prevented it by contract. The buyer gains flexibility, the vendor gains cover, and no one gains an enforceable boundary. Three, for startups and suppliers hoping to sell AI into government, the episode posts the price of admission in plain figures: accept all lawful uses.
How Anthropic’s lawsuit ultimately resolves will decide whether the “supply-chain risk” label matures into a routine procurement weapon. For any AI company that counts government among its customers, that is the thread to keep pulling.