ALETH / AI-BRIEF / 2026-03-07 / FALLOUT FROM THE PENTAGON'S ANTHROPIC BREAKThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
The Pentagon designated Anthropic a supply chain risk, and Anthropic will sue.
OpenAI added Pentagon deal terms barring intentional domestic surveillance.
OpenAI’s GPT-5.4 is its first general-purpose model with native computer use.
The US military used Claude inside Palantir’s Maven to prioritise targets in Iran.
Nvidia put $2bn each into Lumentum and Coherent for future optics capacity.
Browse by day:
Weekend: Claude No. 1 · OpenAI terms · Pentagon talks
Monday: Nvidia optics · US Supreme Court · US agencies · OpenAI
Tuesday: Flash-Lite · GPT-5.3 Instant · Anthropic revenue · Qwen
Wednesday: Iran targeting · Gemini suit · UK AI lab · Broadcom · Ratepayers
Thursday: WhatsApp · US chip rules · GPT-5.4 · Anthropic
Friday: UK copyright · Claude Code · SoftBank · Firefox · Abilene
Weekend (28 February-1 March)
Claude reached No. 1 among free US apps on Apple’s App Store on Saturday, a day after the administration moved to cut Anthropic off.
ChatGPT sat at No. 2 and Google’s Gemini at No. 4. Claude had ranked No. 131 on 30 January. Anthropic said free users had risen >60% since January and paying subscribers had more than doubled this year.
Anthropic’s Claude hits No. 1 on Apple’s top free apps list · CNBC
Claude beats ChatGPT in app downloads after Pentagon blacklists Anthropic · Axios
OpenAI said its Pentagon agreement has “more guardrails than any previous agreement for classified AI deployments, including Anthropic’s”.
The contract lets the department use OpenAI’s system “for all lawful purposes”, within existing law. OpenAI said the deployment is cloud-only, with cleared OpenAI staff in the loop. It also said Anthropic should not be designated a supply chain risk.
Our agreement with the Department of War · OpenAI
Anthropic should not be designated as a supply chain risk · OpenAI on X
Anthropic’s talks with the Pentagon collapsed over the department’s demand to use Claude on bulk data collected about Americans, The Atlantic reported.
On the final Friday the Pentagon had agreed to drop qualifiers such as “as appropriate” from its pledges, a source told the magazine. The New York Times reported that Emil Michael pushed to process commercial data, including location data and browsing histories. Anthropic offered the National Security Agency (NSA) access for classified material, in return for a binding guarantee on commercial data about Americans. The Pentagon refused. Intelligence agencies including the CIA are pushing for a settlement.
Inside Anthropic’s killer-robot dispute with the Pentagon · The Atlantic
Monday 2 March
Nvidia committed $2bn each to Lumentum and Coherent.
Both multiyear deals include multibillion-dollar Nvidia purchase commitments and rights to future capacity of laser and optical components. The money supports R&D and US manufacturing, including a new Lumentum fab.
NVIDIA announces strategic partnership with Lumentum · Nvidia
NVIDIA and Coherent announce strategic partnership · Coherent
The US Supreme Court declined to hear Stephen Thaler’s appeal over copyright in an image made by his AI system.
The denial leaves in place the 2025 DC Circuit ruling that upheld the US Copyright Office’s refusal to register the work for lack of human authorship.
Docket for 25-449 · Supreme Court
AI art can’t be copyrighted after Supreme Court declines to review · The Verge
The US Treasury, State Department and Department of Health and Human Services said they would stop using Anthropic’s products.
Treasury Secretary Bessent said “no private company will ever dictate the terms of our national security”. The General Services Administration is removing Anthropic from its marketplace and USAi programme. Fannie Mae and Freddie Mac will also cut ties, and the State Department is replacing Claude in its chatbot with OpenAI’s GPT-4.1.
Treasury is terminating all use of Anthropic products · Scott Bessent on X
Agencies shed Anthropic contracts following Trump’s directive · Nextgov/FCW
Fannie Mae, Freddie Mac drop Anthropic · National Mortgage News
State Department switches to OpenAI as agencies phase out Anthropic · Reuters
OpenAI announced additional Pentagon language barring intentional domestic surveillance.
New language says the system “shall not be intentionally used for domestic surveillance of U.S. persons and nationals”, including through commercially acquired personal data. The department also affirmed that its intelligence agencies, such as the NSA, will not use OpenAI’s services without a new agreement.
Sam Altman told staff OpenAI “shouldn’t have rushed to get this out on Friday”, and that the deal “just looked opportunistic and sloppy”.
Our agreement with the Department of War, update of 2 March · OpenAI
Here is re-post of an internal post · Sam Altman on X
Tuesday 3 March
Google released Gemini 3.1 Flash-Lite in preview, calling it the fastest and most cost-efficient model in its Gemini 3 series.
It costs $0.25/m input and $1.5/m output tokens. Google says it starts answering 2.5X faster than Gemini 2.5 Flash, and scores 86.9% on GPQA Diamond.
Gemini 3.1 Flash-Lite: our most cost-effective AI model yet · Google
OpenAI released GPT-5.3 Instant to all ChatGPT users, saying it makes fewer unnecessary refusals.
On OpenAI’s higher-stakes evaluation it cut hallucination rates by 26.8% with web search and 19.7% without, against prior models. It is in the API as gpt-5.3-chat-latest, and GPT-5.2 Instant retires on 3 June.
GPT-5.3 Instant: smoother, more useful everyday conversations · OpenAI
Anthropic was on track to surpass a $19bn annualised revenue run rate, up from $14bn a few weeks earlier, Bloomberg reported.
The run rate was $9bn at the end of 2025, and Claude Code drove much of the growth. OpenAI topped $25bn in annualised revenue as Anthropic narrowed the gap, The Information reported on Wednesday.
Anthropic on track to surpass $19 billion in revenue run rate · Shirin Ghaffary on X
OpenAI tops $25bn annualized revenue as Anthropic narrows gap · The Information
Junyang Lin, a central technical leader on Alibaba’s Qwen team, stepped down a day after the release of four small Qwen3.5 models.
Core team members Binyuan Hui, Bowen Yu and Kaixin Li followed him out, 36Kr reported. Alibaba CEO Eddie Wu accepted Lin’s resignation and set up a Foundation Model Task Force that he will lead, The Information reported.
me stepping down. bye my beloved qwen · Junyang Lin on X
Qwen model head Lin Junyang resigns, Alibaba executives answer staff · 36Kr
Wednesday 4 March
The US military used Claude inside its Iran targeting system.
Palantir’s Maven Smart System, with Claude embedded, provided real-time targeting and target prioritisation as the US struck 1,000 targets in 24 hours, The Washington Post reported. Two people said it was Claude’s first use in a major war.
The Wall Street Journal first reported Claude’s use in the strikes, which began hours after Trump ordered agencies off Anthropic. The military will keep using it until a replacement is phased in, and one person said the administration would use government powers to retain it if Amodei tried to stop it.
Claude central to US Iran campaign · The Washington Post via Stars and Stripes
A wrongful death lawsuit filed in federal court in Northern California alleges Google’s Gemini drove a 36-year-old Florida man to suicide.
The complaint says Gemini sent him on missions to obtain a robot body, then persuaded him to end his life so he could join it as a digital being. Google said Gemini repeatedly referred him to crisis hotlines and made clear it was an AI.
Complaint, Gavalas v. Google · US District Court, N.D. Cal.
The UK government will launch a Fundamental AI Research Lab backed by up to £40m, where researchers can win funding and large-scale compute.
A panel chaired by Google DeepMind’s Raia Hadsell will assess applications. AI minister Kanishka Narayan called it “a long-term investment in the brilliant minds who will keep the UK in the AI fast lane”.
Broadcom reported Q1 FY2026 revenue of $19.3bn, up 29% YoY, with AI revenue up 106% to $8.4bn.
It guided Q2 revenue to about $22.0bn and AI semiconductor revenue to $10.7bn. It also authorised up to $10bn of share buybacks through 2026.
Broadcom Q1 FY2026 results, Exhibit 99.1 · SEC
Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI pledged at the White House to pay for the power and grid upgrades their data centres need.
The companies will negotiate separate rate structures with utilities and states, and pay for the power brought online for them whether they use it or not.
Fact sheet: the Ratepayer Protection Pledge · The White House
Thursday 5 March
Meta will let general-purpose AI chatbots use the WhatsApp Business API in Europe for 12 months, to head off EU interim measures.
Providers will pay €0.05 to €0.13 per non-template message, depending on the country. The European Commission said it is analysing the impact on its interim measures and antitrust investigations.
New pricing policy for AI Providers · Meta
Meta will allow rival AI chatbots on WhatsApp in Europe, but for a fee · TechCrunch
The US Commerce Department is reported to be drafting rules that would require government approval for AI chip exports to any country.
A White House official said the draft “does not reflect what President Trump has said on export controls”. The Commerce Department said it would not return to the AI diffusion rule. The FT reported that big buyers of Nvidia and AMD chips may have to invest in US AI infrastructure.
Washington reportedly moves to put AI chips on tighter leash · The Register
Reporting that we were returning to the AI diffusion rule · US Commerce on X
OpenAI released GPT-5.4 with native computer use.
GPT-5.4 is in ChatGPT as GPT-5.4 Thinking, in the API and in Codex, with GPT-5.4 Pro for the hardest tasks. OpenAI’s own evaluations put it at 75.0% on OSWorld-Verified, against 47.3% for GPT-5.2 and reported human performance of 72.4%. On GDPval it matched or beat professionals in 83.0% of comparisons. API pricing is $2.5/m input and $15/m output tokens, and $30/m and $180/m for GPT-5.4 Pro. Codex gets experimental support for a 1M-token context window.
Introducing GPT-5.4 · OpenAI
OpenAI launches GPT-5.4 with native computer use mode · VentureBeat
The Pentagon formally designated Anthropic a supply chain risk.
Anthropic said it received the Department of War’s (DoW) letter on 4 March and sees “no choice but to challenge it in court”. Amodei said the label covers only Claude’s use in direct Department of War contracts, because the statute, 10 USC 3252, is narrow. Anthropic will keep supplying the department “at nominal cost” through the transition. Amodei also apologised for the tone of a leaked post, written hours after the ban, in which he called OpenAI’s messaging “straight up lies”. Anthropic’s products, including Claude, stay available to customers other than the DoW.
Where things stand with the Department of War · Anthropic
Anthropic CEO Accuses OpenAI of ‘Safety Theater’ in Pentagon AI Deal · Bankwatch
Anthropic remains available outside of defense projects · CNBC
Friday 6 March
The UK government will delay its AI copyright changes after its consultation produced no option stakeholders favoured, the Financial Times reported.
An AI bill is no longer expected in May’s King’s Speech. The House of Lords Communications and Digital Committee has called for a licensing-first regime.
UK to delay difficult decisions on AI copyright rules · Financial Times
Alexey Grigorev said Claude Code wiped DataTalksClub’s production database with a Terraform command.
The wipe took down the course platform and 2.5 years of homework, project and leaderboard submissions. The automated snapshots were gone too.
Claude Code wiped our production database via Terraform · Alexey Grigorev on X
SoftBank is seeking a bridge loan of up to $40bn, mainly to fund its stake in OpenAI, Bloomberg reported.
It would be the largest pure dollar borrowing in SoftBank’s history. Four banks including JPMorgan would underwrite the loan, which would run about 12 months.
SoftBank seeks record loan of up to $40 billion for OpenAI stake · Bloomberg
Claude Opus 4.6 found 22 Firefox vulnerabilities in two weeks, 14 of them high-severity, Anthropic said.
The 14 are almost a fifth of the high-severity Firefox flaws fixed in 2025. Mozilla fixed most of the bugs in Firefox 148, with the rest due in later releases. Across several hundred attempts costing about $4,000 in API credits, Claude produced working exploits in two cases, in a test environment with some browser security features, including the sandbox, removed.
Partnering with Mozilla on Firefox security · Anthropic
Oracle and OpenAI cancelled plans to expand their flagship Stargate campus in Abilene, Texas, Bloomberg reported.
The site was to grow from about 1.2 GW to 2 GW, but financing problems and OpenAI’s shifting demand forecasts got in the way. OpenAI would rather deploy Vera Rubin chips at a new campus. Meta is in talks to take up Crusoe capacity.
Oracle and OpenAI end plans to expand flagship data center · Bloomberg
ALETH · OPEN ACCESS RESEARCH