Pentagon Inks $3.4B AI Deals With Nvidia, Microsoft, AWS — Excludes Anthropic From Classified Networks

By Jaspal Singh May 1, 2026 Updated: May 4, 2026

The Pentagon's biggest AI announcement to date landed yesterday: contracts with Nvidia, Microsoft, and AWS to deploy classified-grade large language models across Department of Defense networks. One name was conspicuously absent — Anthropic. Days after the White House circulated a draft workaround for Anthropic's Acceptable Use Policy, the DoD's procurement choices read like a deliberate signal. The Pentagon wants AI horsepower without the policy strings, and it's willing to spend big to get it.

The combined ceiling on the awards — disclosed in budget filings and reporting from Bloomberg and Defense One — exceeds $3.4 billion across multiyear options. Nvidia's slice covers GPU clusters and reference architectures for two classified compute environments. Microsoft is integrating Azure Government Top Secret with custom-tuned variants of GPT-class models. AWS is providing both Bedrock-hosted models and the underlying secret-cleared infrastructure. Together, the three firms now blanket the DoD's classified AI stack in a way that, until last quarter, Anthropic seemed positioned to share.

What's actually in the contracts

The structure is unusual for DoD AI work. Rather than narrow task orders, these are open-ended capability vehicles — closer to the JEDI/JWCC cloud awards than to a single-application contract. Each prime can be tasked across the SECRET and TOP SECRET networks (SIPRNet and JWICS), with Microsoft additionally cleared to operate at the highest compartmented levels through its existing IL6+ accreditations.

Three workloads are explicitly named in the unclassified summaries: signals intelligence triage, logistics and supply-chain forecasting at INDOPACOM, and large-scale document review for the Office of the Secretary of Defense. The latter is the one most observers find revealing — it's exactly the kind of "long-context reasoning over sensitive text" workload that Anthropic's Claude has been pitched on for two years.

Why Anthropic was cut out

The official line from the Pentagon is procurement-neutral: best-of-breed selection, contract competition, no comment on individual vendors. The unofficial line, which I've heard echoed across three different defense-tech investor calls this week, is blunter — Anthropic's AUP language around "national security applications" was treated as a commercial and operational risk that DoD lawyers didn't want to litigate inside a multiyear award.

That tracks with the timeline. On April 28, news broke that the White House was drafting a workaround for Anthropic's AUP restrictions on classified AI use. On April 30, OpenAI publicly criticized Anthropic for "restricting Mythos." On May 1, OpenAI walked back its own cyber tool — but the optics had already shifted. By the time these contract awards were finalized, Anthropic was the company DoD had to negotiate around, not with. Nvidia, Microsoft, and AWS all sell models or model-hosting infrastructure that can run frontier AI without an Anthropic-style usage policy attached.

My Take

This isn't an indictment of Anthropic's safety work — it's a procurement consequence of it. The company has built its commercial differentiation on having tighter usage policies than its peers. That position is internally coherent and probably the correct long-run bet for enterprise customers worried about brand and liability exposure. But it is a bet that scales poorly in markets where the buyer is the U.S. government and the use case is "kill chain decisions, but make it AI."

The deeper story is what this means for the AI industry's bifurcation. We are watching a clean split form between vendors who treat their models as products (Anthropic, parts of Google) and vendors who treat them as infrastructure (Nvidia, Microsoft, AWS). Infrastructure wins government contracts because infrastructure doesn't say no. Products lose them because products have opinions. Whether you think that's a good thing depends on whether you trust the operator more than you trust the policy. I'd argue most safety-aligned arguments rest on the assumption that nobody trusts either — which is why the AUP existed in the first place.

What this means for AI in defense procurement

Three things will shift in the next 12 months because of this announcement. First, expect "AUP-clean" to become a procurement criterion in classified RFPs — not stated openly, but baked into the technical evaluation factors. Second, expect a wave of fine-tuned, smaller open-weight models trained on government data to fill the gap where commercial frontier APIs are deemed too restrictive. Third, expect Anthropic to either soften the AUP for narrow government carve-outs or formalize a separate gov-only product line — the latter being more likely given how their compliance team has historically operated.

For the broader market, the contracts also reset the ceiling on what "AI for defense" looks like in dollar terms. $3.4B across three primes is roughly the size of Palantir's full annual U.S. government revenue. The DoD has just told the market that AI infrastructure is now a tier-one defense category, not a sidebar to cloud or analytics spend. That is a structural shift that will pull more pure-play AI vendors into the defense-tech ecosystem in 2026 and 2027.

Frequently Asked Questions

Why was Anthropic excluded from the Pentagon AI contracts?
Officially, Pentagon procurement decisions are based on best-of-breed competition. Unofficially, Anthropic's Acceptable Use Policy restrictions on national-security applications were considered a contract-execution risk. The White House had been drafting a workaround for the AUP, but contract awards were finalized before any policy resolution, leaving Anthropic outside this round.

What's the total contract value?
The combined ceiling across Nvidia, Microsoft, and AWS exceeds $3.4 billion when multiyear option periods are included. Initial obligations are smaller — typically less than 30% of the ceiling in year one — with task orders driving the rest.

Are these LLM-only contracts?
No. The vehicles cover GPU compute, model-hosting platforms, fine-tuning services, and secure inference infrastructure across SIPRNet and JWICS. LLMs are the headline workload, but the scope includes the full classified AI stack.

Could Anthropic still win classified work later?
Yes — the contract structure leaves room for additional vendors to be onboarded as subcontractors or as task-order awardees. But it would require either a softening of the AUP for government use or a new product line designed for classified deployment.

The Bottom Line

Anthropic built its brand on usage policies. That brand just cost it a seat at the Pentagon's biggest AI table. Nvidia, Microsoft, and AWS have each grabbed a third of a $3.4 billion infrastructure pie that didn't exist two quarters ago — and the message to the rest of the industry is that AI for defense is being structured as infrastructure, not as software with opinions. Expect every frontier vendor without a government carve-out to rethink that posture before the next round of awards lands.