The Signal — February 28, 2026
THE SIGNAL
Future Shock Daily — February 28, 2026
The Defense Production Act got invoked against an AI company yesterday. That sentence alone tells you where things are headed.
The Pentagon Invokes a Korean War Law Against Anthropic
The Department of Defense invoked the Defense Production Act — a 1950 law originally designed to mobilize industrial capacity during wartime — to compel Anthropic to provide unrestricted access to Claude for autonomous weapons and surveillance programs. Anthropic refused.
Dario Amodei's public statement was blunt: Anthropic will not build autonomous weapons systems or provide models stripped of safety restrictions for military use. The company's position is that its acceptable use policy applies to all customers, including the U.S. government.
The Pentagon's move came after months of quiet negotiations broke down. Defense Secretary Hegseth framed it as a national security imperative, arguing that adversaries won't wait for Silicon Valley to sort out its ethics. Employees at Google and OpenAI responded with an open letter backing Anthropic's stance, which Platformer reported had gathered over 2,000 signatures by Thursday evening.
This is the first time the DPA has been used against a technology company over AI specifically. The legal question — whether the government can compel a company to remove safety restrictions from its own product — has no precedent. Anthropic says it will challenge the order in court.
Sources: NPR · The Verge · TechCrunch · Platformer · The Decoder
OpenAI Raises $110 Billion at a $730 Billion Valuation
OpenAI closed the largest funding round in corporate history. Amazon led with $50 billion, joined by Nvidia at $30 billion and SoftBank at $30 billion. The $730 billion valuation puts OpenAI roughly on par with Berkshire Hathaway and ahead of every bank on earth.
Part of the announcement: ChatGPT now has 900 million weekly active users. That's not a typo. Nine hundred million people use the product every week. For context, Instagram took over a decade to reach similar weekly engagement numbers.
The money comes with strategic strings. Amazon's $50 billion buys more than equity — OpenAI's new Frontier platform will be available through AWS, making Amazon a distribution partner for OpenAI's most capable models. Nvidia's investment deepens a hardware dependency that already defines the industry. SoftBank's play is a bet that OpenAI will be the platform layer for the next era of computing.
The question nobody in the announcement addressed: where does this leave the "capped profit" structure? OpenAI's ongoing conversion to a full for-profit entity just got a lot more complicated with $110 billion in new stakeholders.
Sources: OpenAI Blog · The Verge · TechCrunch · The Decoder
Gemini 3.1 Pro Scores 77.1% on ARC-AGI-2
Google released Gemini 3.1 Pro, and the headline number is hard to ignore: 77.1% on ARC-AGI-2, more than double what Gemini 3 Pro achieved. ARC-AGI-2 tests abstract reasoning through novel visual puzzles that can't be brute-forced with memorization — the kind of problems that have historically separated human cognition from pattern matching.
The jump from ~35% to 77% in a single generation is steep. For reference, the original ARC-AGI benchmark was considered effectively saturated by late 2025, which is why ARC-AGI-2 exists — harder tasks, less room for shortcutting. A 77% score puts Gemini 3.1 Pro at the top of the public leaderboard and gives Google its strongest claim yet in the general reasoning race.
Whether this translates to real-world capability gains or stays a benchmark story is the open question. ARC-AGI-2 measures something genuine — novel problem-solving, not regurgitation — but benchmark scores and product utility don't always track together.
Sources: Google Blog · MarkTechPost · TechInformed
On the Editor's Desk
The pipeline pulled in 68 events yesterday. Ten passed editorial review. We published three.
Block cut 40% of its workforce — about 4,000 people — and CEO Jack Dorsey explicitly credited AI for making it possible. Engineer output is up 40% per the CFO. The stock jumped. This is probably the clearest "AI replaced us" data point at a major public company so far, and it almost made the cut. We dropped it only because three stories is the limit and the top three were too big to skip.
Meta quietly signed a multi-billion dollar deal with Google for TPU access, diversifying away from total Nvidia dependence. Reuters broke it via The Information. No official confirmation from either company, which is why it's here instead of above.
Anthropic also acquired Vercept, a computer-use AI startup — their second acquisition after Bun. It's a small deal in dollar terms but interesting given Anthropic's agentic push. Hard to give it column inches on a day when the same company is staring down the Pentagon.
The "China's New AI Breakthrough" story making the rounds? It's a YouTube Shorts video with no article, no details, and no primary source. We're holding it until someone produces actual reporting.
METR published new benchmark data showing Claude Opus 4.6 operating at a 14.5-hour autonomous time horizon — meaning it can work independently on complex software tasks for that long before needing human intervention. The Sky News coverage leaned toward alarm; the actual METR data is more measured.