AI Safety Standards, Corporate Spinoffs, and Infrastructure Battles Define Day

Big Tech

• Following the White House's imposition of export controls on Anthropic, which forced the company to suspend access to its powerful AI models Fable 5 and Mythos 5 over a perceived security flaw (a jailbreak), discussions have shifted toward establishing industry-wide rules for AI safety. The White House and Anthropic are now working on a framework designed to assess the severity of security flaws in new AI models, aiming to guide potential government intervention (13). These talks seek to develop common benchmarks—including how safeguards were bypassed and the practical consequences of a breach—to standardize the evaluation of future jailbreaks across the industry (13). This shift reflects an effort by the administration to create guardrails for powerful frontier models while acknowledging that no AI model can be completely immune to hacking (13, 5). ^[13][5]
• AI talent wars intensified as Google AI leader Noam Shazeer left the company to join OpenAI, marking a significant shift in the industry's top labs [49, 36]. Shazeer, who was previously co-lead of Gemini and credited with helping close the gap between Gemini’s capabilities and ChatGPT’s, announced his departure from Google on Wednesday [49, 36]. Notably, two years prior to this move, Google had paid $2.7 billion to poach Shazeer and part of his team from Character.AI [49, 36]. Furthermore, OpenAI is also strengthening its policy credentials with the addition of former Trump White House AI policy official Dean Ball, who will lead the new Strategic Futures team focusing on matters like catastrophic risk and government relations [36]. ^[36][49]
• Amazon is positioning itself as a major competitor to Nvidia by planning to sell its homegrown AI chips, Trainium, to third parties for use in data centers, potentially challenging Nvidia's market dominance [50]. This ambition stems from Amazon CEO Andy Jassy’s assessment that the company’s annual run rate for this chip business could reach approximately $50 billion, which would place it directly into the competitive sphere with Nvidia’s current $326 billion revenue run rate [50]. While AWS has historically resisted selling its chips due to the complex nature of its service revenue model (which includes charges for storage and security alongside AI tokens), spokesperson Doron Aronson confirmed that Amazon may sell these racks of chips in the future, echoing Jassy’s earlier statements about the high demand for Trainium capacity [50]. ^[50]