The money behind frontier AI moved to the front of the story today: a bubble warning from one of the field’s founders, a Google challenge to Nvidia, and a $13 billion bet on cheap inference. Add a near GPT-5.6 launch, agents that finally remember, and a policy analyst defecting into OpenAI, and the through line is clear. The question is no longer whether AI is capable. It is whether the business behind it adds up.
The Bubble Clock Starts Ticking: AI’s Economics Under Pressure
The math behind frontier AI is getting harder to ignore, from a blunt warning about the business model to the scramble for cheaper chips and the talent that runs them.
- LeCun warns of a bubble explosion and calls xAI a failure. Yann LeCun said labs must raise prices or cut costs to avoid a reckoning, because users pay far less than it costs to serve them, and called Musk’s xAI unable to compete. The structural pressure he describes is real even if the messenger runs a rival.
- Google opens its TPU business to rivals, starting with Anthropic. Google is renting thousands of its in-house chips to Anthropic, turning TPUs into a credible second option to Nvidia and reshaping how compute deals get financed. A real number two changes the negotiating math for anyone building at scale.
- Baseten closes a $1.5B round at up to $13B on cheap inference. The startup is betting that serving open models cheaply, not proprietary frontier APIs, is the next infrastructure war, with revenue tripling to roughly $600 million. A $13 billion valuation is a market vote that the quality gap is closing.
- DeepSeek made investors promise not to poach its staff. A loyalty clause on its first funding round bars backers from recruiting DeepSeek engineers or seeding their spinouts, a sign of how fierce China’s AI talent war has become. The condition reveals leverage on capital and weakness on retention.
Frontier Moves: New Models, New Training Tricks
The labs keep pushing capability and timing, with a new flagship reportedly days away and fresh ideas for how models actually learn.
- OpenAI eyes a GPT-5.6 launch next week with a 1.5M token context. Reporting points to a base model plus Mini and Pro variants, faster Codex responses, and pricing set to undercut Anthropic while Claude Fable 5 availability stays uncertain. The next two weeks are the clearest comparison window in months.
- OpenAI says training on beneficial traits generalizes broadly. A small slice of beneficial-behavior data shifted model conduct across 44 evaluations the training never targeted, and the gains held under adversarial pressure. The composition of RL data may matter more than previously recognized.
- NVIDIA’s ZPPO fixes the hardest-question problem in RL. A replay buffer keeps drilling a model on the hard prompts standard RL throws away, beating GRPO across text, vision, and video benchmarks. The approach sidesteps the on-policy violations that teacher injection introduces.
Agents Grow Up: Memory, Collaboration, and New Leaks
Agentic systems are gaining persistence and shareable context, and researchers are mapping the failure modes that come with them.
- Perplexity builds persistent memory into its agent platform. Brain gives agents a source-linked memory graph so they stop reconstructing context from scratch every session. If it works, every other agentic platform faces pressure to close the same gap.
- Claude Code gets artifacts: shareable live pages from a session. Anthropic can now publish a work session as an auto-updating page, pushing the tool from individual productivity toward team collaboration. The page replaces the standup summary, not just the chat reply.
- Deep research agents leak private data through their own queries. ServiceNow’s MosaicLeaks benchmark shows the leak is real, and that privacy-aware training cuts it from 34% to 9.9% without losing task accuracy. Any team running research agents over private plus public data is exposed today.
- Mistral’s Le Chat gains a Code tab and an app-building surface. The Paris lab is chasing the territory ChatGPT, Claude, and Gemini already hold, moving Le Chat from chat toward a full product platform. Open weights remain its clearest differentiator.
Governing the Machines: Who Sets the Rules
The people writing AI policy are moving inside the labs, and the labs are building security that assumes their own agents might go wrong.
- Google DeepMind treats its own AI agents as insider threats. A new control framework layers containment on top of alignment, assuming a deployed agent might not share the company’s goals and building for that case anyway. A lab that builds for alignment failure does not fully trust its alignment training.
- Dean Ball leaves the think-tank world to run policy inside OpenAI. The Hyperdimensional analyst is joining OpenAI to lead a new Strategic Futures team, arguing the easy phase of AI governance is over. The periodization is the durable signal: old playbooks are stale.
Today’s Quick Hits
- Midjourney pivots to medical hardware with a full-body scanner. The image generator says its ultrasonic scanner maps the body in under 60 seconds, to be housed in company-built spas, with FDA clearance still ahead.