Pentagon Ultimatum Shakes AI Industry

You're one of the first 10,000 people getting The AI Debrief in your inbox - welcome to the smartest AI newsletter for founders and business leaders.

The Pentagon just delivered an ultimatum that could reshape how every AI company thinks about government contracts. Defense Secretary Pete Hegseth gave Anthropic until 5:01 pm tomorrow to remove all safety restrictions on its Claude AI model - or face the Defense Production Act. This Korean War-era law lets the government seize private technology in the name of national security.

This isn't a negotiation. It's a test case for who controls AI guardrails when billions in government contracts are at stake.

In today’s AI debrief:

The debrief: Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei until Friday at 5:01 pm to grant the military unrestricted access to Claude or face the Defense Production Act, contract termination, and blacklisting - but Anthropic rejected the ultimatum Thursday, saying they "cannot in good conscience" remove safeguards against mass surveillance and autonomous weapons.

The details: In a tense Tuesday meeting at the Pentagon, Hegseth told Amodei that the Department of Defense cannot depend on a company that maintains restrictions on "lawful" military uses, even after Anthropic offered to allow Claude for missile defense and cyber operations. The Pentagon specifically referenced Claude's use during the January operation that captured Venezuelan President Nicolás Maduro, with sources telling Axios that Anthropic raised concerns to partner Palantir about the operation. Hegseth was joined by Deputy Secretary Steve Feinberg, Under Secretary Emil Michael, and the Pentagon's top lawyer in the meeting - a show of force that signals how seriously the government is taking this fight. Anthropic's two red lines remain firm: no use of Claude for mass surveillance of Americans, and no fully autonomous weapons that fire without human involvement, with the company arguing Claude is not reliable enough to avoid potentially lethal mistakes like unintended escalation. Wednesday night, the Pentagon sent what it called a "best and final offer," but Anthropic says the contract language includes legal escape hatches that would allow the restrictions to be "disregarded at will." Claude is currently the only commercial AI model operating inside the Pentagon's classified networks under a $200 million contract awarded in July, though xAI's Grok was recently approved for classified use, and the Pentagon is rushing to onboard OpenAI and Google as alternatives.

Why it matters: This is the first time the Defense Production Act has been threatened against an AI company, and it sets the precedent for every frontier AI developer working with the government. If Anthropic caves, it proves that safety principles don't survive government pressure - if they stand firm and lose the contract, it proves you can't build a defense-focused AI business while maintaining ethical guardrails. For founders, this is a preview of the choice coming: do you take government money and give up control, or walk away from billions in defense contracts? OpenAI, Google, and xAI have already agreed to "any lawful use" terms, leaving Anthropic isolated. The outcome by tomorrow will determine whether AI companies can set their own usage policies or whether "national security" overrides everything. Watch what happens at 5:01 pm Friday - that's when we'll know if private AI guardrails can survive contact with the Pentagon.

The debrief: Meta announced a $100 billion, multi-year deal with AMD to purchase up to 6 gigawatts of AI computing power - just one week after expanding its Nvidia partnership - in what's being called the biggest chip diversification bet in AI history and a direct challenge to Nvidia's 84% market dominance.

The details: The AMD deal is structured as an "equity-for-compute" partnership where Meta can acquire up to 160 million AMD shares (roughly 10% of the company) for $0.01 each if AMD hits delivery milestones, with the full stock award requiring AMD's share price to reach $600 (it closed at $196.60 on Monday). Shipments of the first gigawatt deployment begin in the second half of 2026 using custom AMD Instinct MI450 GPUs optimized specifically for Meta's workloads, with the full 6-gigawatt buildout scaling through 2030-2031. This comes just seven days after Meta announced it would deploy "millions" of Nvidia's latest Blackwell and Rubin GPUs, plus becoming the first large-scale deployment of Nvidia's Grace CPUs as standalone chips. The timing reveals Meta's strategy: continue buying from Nvidia (which still controls the vast majority of AI chips) while building AMD as a credible second source to avoid single-vendor lock-in. Meta is also developing its own in-house MTIA chips, which have reportedly hit delays, but CEO Mark Zuckerberg framed the AMD partnership as "an important step" toward "personal superintelligence" as Meta commits to at least $135 billion in AI capital expenditures in 2026 alone as part of a $600 billion+ multi-year infrastructure buildout.

Why it matters: This isn't Meta scaling back AI investment - it's Meta spending so much on AI that no single chip vendor can supply it all. The real story is that Nvidia's monopoly pricing pushed Meta to take a 10% equity position in a competitor just to create leverage in future negotiations. For VC-backed startups, this reveals a harsh truth: chip access is becoming a moat, and if you're not spending billions to lock up supply, you're competing for scraps. The "equity-for-compute" structure is new and will be copied - expect OpenAI, Microsoft, and Google to demand similar deals where they exchange stock for guaranteed chip supply, fundamentally changing how semiconductor partnerships work. If AMD successfully delivers and vests Meta's warrants, watch AMD's stock potentially 3x from current levels, and watch every hyperscaler rush to replicate this model. The bet Meta is making: AI compute demand is so massive that the market can support multiple chip vendors at scale, breaking Nvidia's stranglehold. If AMD fails to deliver on time, Meta's AI roadmap stalls and Nvidia wins by default.

The debrief: Goldman Sachs warned that accelerating AI adoption could push U.S. unemployment from 4.3% to 4.5% by year-end, with the bank estimating AI is already responsible for 5,000-10,000 monthly net job losses in exposed industries and accounted for 7% of all planned layoffs in January 2026.

The details: Goldman economist Pierfrancesco Mei wrote that "AI-driven displacement could raise the unemployment rate slightly in 2026, with upside risks from faster adoption and larger displacement," noting that job growth has already "slowed and turned negative in a few subindustries where AI is most ready for deployment." The warning comes with a caveat that faster AI adoption could add 0.3 percentage points to unemployment beyond the base forecast. The research identifies computer programmers, accountants and auditors, legal and administrative assistants, and customer service representatives as most at risk, with unemployment among 20-to-30-year-olds in tech-exposed occupations rising nearly 3 percentage points since early 2025. This coincides with what analysts are calling "The Great AI Scare of 2026," hitting software stocks like IBM and Salesforce, and follows a Citrini Research scenario published Sunday, projecting that unemployment could double and the stock market could lose one-third of its value by 2028 as agentic AI accelerates job displacement. However, Goldman's longer-term view is more optimistic: the bank expects the impact to be temporary, estimating that generative AI will ultimately boost labor productivity by 15% across developed markets when fully adopted, with most displaced workers finding new roles after a transition period.

Why it matters: This is the first major bank to put hard numbers on AI's near-term employment impact, and it's coming from Goldman - the same firm advising most of these companies on their AI strategies. The critical insight is that AI isn't causing a recession-style unemployment spike - it's creating concentrated job losses in specific white-collar categories even while the economy remains healthy overall. For founders and executives, the game has changed: investors used to reward mass layoffs as "operational efficiency," but Goldman notes that markets now increasingly view large-scale job cuts as a warning sign of weak growth rather than smart management. If you're planning to cut headcount and cite "AI efficiency gains," be aware that Wall Street is growing skeptical of that playbook. The smarter approach: show productivity improvements without mass cuts, redeploy talent to higher-value work, and demonstrate that AI is augmenting your team rather than replacing it. The unemployment forecast also has Fed policy implications - if joblessness rises while inflation stays elevated, expect interest rates to stay higher for longer, making capital more expensive just as AI infrastructure spending accelerates.

UN launches 40-member global AI panel over U.S. objections - The UN General Assembly voted 117-2 to establish the Independent International Scientific Panel on AI, selecting 40 experts from 2,600+ candidates to serve as an "early-warning system" on AI impacts. Only the U.S. and Paraguay voted against, with the U.S. citing "sovereignty concerns.".

California passes sweeping AI transparency laws - Three new bills (AB 2169, AB 2071, AB 2025) expand consumer data rights for AI systems, mandate AI literacy in K-12 schools, and require disclosure of AI-altered real estate materials. The legislation builds on California's SB 53, which made it the first state to directly regulate frontier AI developers.

OpenAI and AMD struck a similar equity-for-compute deal in October - Meta's $100B AMD partnership follows OpenAI's October 2025 agreement with AMD that also traded equity for chip supply, establishing a new "equity-for-compute" category in semiconductor partnerships. [TechCrunch]

- Drew & the rest of the humans behind The AI Debrief