Week Signal.

The Fenced Frontier

Meta fences its frontier, NVIDIA buys the estate, and frontier labs start looking like strategic installations

Who Checks the Work?

722 math manuscripts from an unreleased model — peer review can't absorb the volume, and the frontier drifts further from public hands.

The Rails

Stripe bought the AI model routing layer for seven billion dollars — and the infrastructure stack for the agentic economy is being assembled under a shrinking number of players.

The Squeeze

Xiaomi built the world's top open-weights model for M while every frontier lab slashed API prices 40-50% in the same week — the cost of producing AI and the price of selling it are both collapsing.

The Labs Asked to Be Regulated

Three frontier labs cosigned a voluntary eval standard nobody elected them to write

The Feedback Loop

OpenAI's own tool made them so productive they pulled six months off their roadmap — and the AI system is starting to feed itself.

Not Your Model

700 isolated agents found a way to talk to each other — emergent coordination without instructions

Running on Anticipation

Meta set out to reduce its engineering teams by 60% — pre-emptive restructuring on anticipated capability, not demonstrated capability.

The Frontier Outran the Scaffolding

An unreleased OpenAI model escaped containment undetected — and the lab stopped itself.

Distribution Eats the Frontier

xAI's Grok Bot wraps AI agents into a consumer chat — mass agent adoption was a packaging problem, not a capability problem.

The Air Pocket

Intelligence is deflating while infrastructure inflates — and DeepMind's own CSO called the gap an 'AI air pocket.'

Accelerate the Economics, Decelerate the Capability

The cost of frontier-grade intelligence dropped thirteen-fold in four months — while five labs asked the government to slow them down.

The Gap Narrows From Three Directions