The Signal — March 24, 2026
The Anthropic-Pentagon fight just picked up a U.S. Senator, Sam Altman bet $375 million on fusion timing, and a startup wants to end NVIDIA's inference monopoly.
A Senator, a Hearing, and the Biggest AI-Government Fight of the Year
The Anthropic-Pentagon saga gained a new dimension yesterday when Senator Elizabeth Warren (D-MA) sent letters to Defense Secretary Pete Hegseth and OpenAI CEO Sam Altman, calling the Pentagon's decision to designate Anthropic a supply-chain risk "retaliation" for the company's refusal to loosen safety guardrails on military AI use. Warren also demanded details about OpenAI's defense contracts, drawing a direct line between Anthropic's blacklisting and OpenAI's expanding Pentagon relationship.
The timing matters. Today, U.S. District Judge Rita Lin hears arguments in San Francisco on whether to grant Anthropic an injunction halting the ban. The case has been building for weeks: sworn declarations filed Friday from Anthropic's Sarah Heck and Thiyagu Ramasamy disputed the Pentagon's core claims, while a leaked email from Under Secretary Emil Michael — sent one day after the March 3 supply-chain designation — told Anthropic CEO Dario Amodei the two sides were "very close" on the exact issues now cited as national security threats.
Warren's intervention shifts the frame from procurement dispute to political accountability. A Democratic senator, tech worker amicus briefs from OpenAI/Google/Microsoft employees, and Anthropic's own legal team are now aligned against the designation. Whatever Judge Lin decides today sets precedent for how the government can use supply-chain risk designations against AI companies.
Sources: TechCrunch · CNBC · Washington Examiner
Altman Leaves Helion's Board as OpenAI Eyes Fusion Power
Sam Altman stepped down as board chair of Helion Energy on Monday, the fusion startup he has backed since 2015 with over $375 million of his own money. The reason: OpenAI is in early-stage talks to purchase fusion power from Helion, and Altman needed to clear the conflict of interest.
Axios reported the potential deal's scope: 5 gigawatts by 2030, scaling to 50 GW by 2035, with OpenAI securing 12.5% of Helion's total production. Those are staggering numbers — 5 GW alone could power roughly 3.7 million homes. Helion, which has existing agreements with Microsoft and steelmaker Nucor, declined to confirm the OpenAI discussions, saying it "has not announced any new customer agreements."
The move signals how seriously AI companies are treating the energy constraint. Training and running frontier models requires power at a scale that existing grid infrastructure can't easily provide. Fusion remains unproven at commercial scale — Helion hasn't generated net energy yet — but the bet is that by the time these data centers need the power, the physics will have caught up. It's a $375 million wager on timing.
Sources: Reuters · Axios · TechCrunch · GeekWire
Gimlet Labs Raises $80M for Multi-Chip AI Inference
Gimlet Labs, founded by Stanford adjunct professor and CEO Zain Asgar, closed an $80 million Series A led by Menlo Ventures, bringing total funding to $92 million. The pitch: software that runs AI inference workloads across NVIDIA, AMD, Intel, ARM, Cerebras, and d-Matrix chips simultaneously.
Most AI infrastructure today is locked to a single chip vendor — overwhelmingly NVIDIA. Gimlet's approach treats heterogeneous silicon as a unified compute pool, distributing model inference across whatever hardware is available. The company claims this produces better performance per dollar than running on any single chip type alone.
The investors backing this thesis include Prosperity7 Ventures (Saudi Aramco's tech fund), which called multi-silicon orchestration "the missing layer in the stack" for large-scale AI deployments. If the approach works at production scale, it loosens NVIDIA's grip on the inference market by making chip choice less of a lock-in decision.
Sources: TechCrunch · GlobeNewsWire · SiliconANGLE
On the Editor's Desk
Reuters dropped an exclusive yesterday about OpenAI and Anthropic both courting private equity firms for joint ventures aimed at enterprise AI distribution. The headline number: OpenAI is reportedly offering PE firms preferred equity with a guaranteed 17.5% minimum return. Forbes pegged OpenAI's projected 2026 losses at $14 billion against $20 billion in annualized revenue. The structural insight — using PE firms as distribution channels into portfolio companies rather than just capital sources — is genuinely interesting, but the story didn't clear our pipeline in time for today's edition. We're watching it for tomorrow.
The council also flagged ClawWorm, a self-propagating worm attack demonstrated against production agent frameworks. The arXiv paper showed a 64.5% attack success rate across 1,800 trials. Agent security is a running thread since CanisterWorm last week, and this deserves a standalone piece rather than a newsletter mention.
Also in the pipeline: NC State researchers published a technique called SN-Tune that identifies and freezes safety-critical neurons in language models, preventing jailbreaks without degrading capability. Reduced harmful output scores from 65.5 to 2.0 on Llama 3 8B. Needs more space than a Signal slot to explain properly.
