AI News: May 2, 2026

1. Pentagon Signs Eight AI Companies for Classified Networks, Locks Anthropic Out

Pentagon. The Department of Defense announced agreements with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS, and one additional partner to deploy AI models on classified Impact Level 6/7 networks for mission planning and weapons targeting. Anthropic is conspicuously excluded after refusing to drop usage restrictions on autonomous weapons and mass domestic surveillance, and was designated a supply-chain risk by the administration. Anthropic has filed two lawsuits in San Francisco and Washington, D.C. seeking to overturn the order while talks with the White House continue, but the day-one footprint of US classified AI is now competitor-heavy. Source

2. Big Tech’s 2026 AI Capex Hits $725 Billion

Big Tech. Combined planned AI capital expenditure for Google, Amazon, Microsoft, and Meta in 2026 now totals roughly $725 billion across data centers, accelerators, and adjacent infrastructure, according to consolidated guidance compiled this week. The number is significant for two reasons: it ratchets up the structural floor for AI compute supply through 2027, and it tightens the capital math for any non-hyperscaler trying to compete on owned infrastructure rather than hyperscaler rentals. Expect renewed pressure on power, land, and HBM supply chains in the second half of the year. Source

3. News Publishers Push Common Crawl to Block AI Training

News/Media Alliance. The News/Media Alliance submitted a formal demand on behalf of NBCUniversal, CNN, McClatchy, Vox Media, Ziff Davis, USA Today, and dozens of others, asking Common Crawl to revise its terms of use to explicitly prohibit AI training and to add an opt-out registry warning that unauthorized AI use breaches those terms. The push targets Common Crawl directly — long an upstream source for major model pretraining corpora — rather than only individual model providers, and frames the archive’s continued operation as “data laundering” if it does not act. Press Gazette has separately reported that roughly 80% of the world’s largest news sites now block AI training bots at the robots.txt layer. Source

4. Chinese AI Startups Begin Abandoning Offshore Structures to Register Domestically

Moonshot AI, StepFun. Following Beijing’s rejection of Meta’s attempted acquisition, Chinese regulators signaled that AI companies pursuing public listings should remain domestically registered. Moonshot AI and StepFun are reported to be unwinding their VIE-style offshore holding structures to register directly in China, the first concrete examples of a regulatory regime that has been telegraphed for months. The move tightens Beijing’s grip on its top labs and limits foreign-investor optionality at a moment when those labs are also some of the most active open-weight model publishers globally. Source

5. Replit’s Amjad Masad Comments on Reported Cursor / SpaceX Deal

Replit. Replit CEO Amjad Masad addressed the M&A air at TechCrunch’s event on May 1, distancing his company from acquisition speculation following reports that rival Cursor is in talks to be acquired by SpaceX for a substantial sum. The reported deal would land Cursor inside Elon Musk’s orbit alongside xAI, an unusual structural alignment for a developer-tool incumbent and one that would meaningfully reshape competitive dynamics in the AI coding space if confirmed. Masad signalled that Replit’s preference is independence rather than a sale. Source