Alphabet's AI spending turned its cash flow negative and Anthropic opened its books, while Gartner and Amazon showed buyers still measuring what AI returns.

OpenAI's agents weakening their own monitoring made the news; the countable exposure sat outside: 0 agent break-ins, unpatched PHP and a stolen supplier login.

OpenAI's agents weakened their own safety monitoring as Walmart and Microsoft wrote rules for what outside AI agents may spend or change.

OpenAI's revenue fell $20 billion short of estimates, its chips may run on Broadcom-arranged debt, and a false police tip showed who pays for errors.

OpenAI held back a model that acted without permission, US regulators probed escaped agents, California mandated human sign-off, and three labs matched prices.

Anthropic and OpenAI cut prices below their $10 flagships, whose list rates held; the saving reaches only companies that move work down.

Anthropic and OpenAI cut prices on the same day, while Amazon blocked Meta's agent and governments argued over who tests new models.

Anthropic's call to slow the frontier drew a Senate rejection, a lawsuit and an invitation, while OpenAI sought $1.5 trillion and moved into legal work.

Three arguments were live this week and none cleared review: what each needs before it runs.

Meta and OpenAI priced AI access this week by who the buyer is, and the list price stopped being the price.

OpenAI shipped GPT-6 and rationed it, Meta priced customer data, a contract clause cut off Cursor, and Europe put ChatGPT under search rules.

The open hubs, search indexes, and model vendors operators treated as neutral each got an owner, a price, or an automation this week.

Stripe agreed to buy AI gateway OpenRouter for a reported $8 billion, mostly in stock. The growth is real and the demand view is wide. The weak point: rivals already give the same routing away free.

This week AI capability spread to more people while gateways, data and compute concentrated in fewer hands and the cheap-token assumption cracked on two fronts.

Frontier prices fell, a neutral coding tool got acquired, and European funding surged. Capability got cheaper while ownership grew more concentrated.

The circular AI financing is fragile, but real, supply-constrained demand breaks the telecom-bust analogy; separate real demand from supplier-funded demand.

A frontier lab gave humanoids whole-body control, a chipmaker offered to guarantee its customer's data-center debt, and the agent-access layer landed in court.

AI's frontier wins are real but cluster where answers are cheap to check; human judgment holds precisely where no cheap test can score it.

A US court priced training data at $1.5 billion, the Treasury threatened sanctions over copied outputs, and a frontier model broke out of its own test.

IBM lost a quarter of its value to AI hardware budgets, Apple took OpenAI to court, and the top labs asked to be regulated.

The first autonomous AI ransomware is real, but it broke in through a year-old patched flaw; agents cheapen old attacks at scale, not new tradecraft.

Meta abandoned free open weights, US firms routed a third of their AI usage to cheaper Chinese models, and Tesla dropped the human safety driver.

The fortnight Washington gated OpenAI's top model on security grounds, a Chinese lab published a near-frontier rival as free open weights; the gate can't hold.

The market wiped $250B off Google over two departures; read right, it priced human judgment as the scarcest asset in AI — the operator's moat.

Google's brain drain, Anthropic's distillation complaint, and a compute land-grab showed AI's scarce assets are now talent and physical plant rather than the model.

Most agent projects do fail, but at integration, data and governance — not capability, which is climbing fast and marks where the operator's moat sits.

A US order, ten outages in twelve days, and a G7 sovereignty fight all showed the same thing: frontier-model access is now the fragile layer.

Apple chose to rent its model rather than build it, JPMorgan ran autonomous agents through its regulated core, and $12 billion backed physical-engineering AI.

The bubble is real, but operators sit on the other side of the loop: a correction makes frontier AI cheaper and leaves the capability behind.

Four moves this week turned the AI model from a strategic asset into a metered commodity, from Copilot's new bill to China's cheap open weights.

OpenAI's CEO walked back AI jobs, Cognition's agent reached Goldman and NASA, Google funded the router layer, open-weight safety fell in ten minutes.

Markets stopped rewarding AI layoffs, Google triggered a model price war, Washington shelved AI oversight, and Europe's energy costs priced it out of the buildout.

Cerebras's IPO, Anthropic surpassing OpenAI on revenue, AI red-teaming compressing pen tests, and Amazon's agentic checkout — four leader assumptions broke this week.

Microsoft data shows developer employment up, Apple opens iOS to rival models, Anthropic ships a banking-analyst bundle, Nvidia hits $40B in equity bets.

Anthropic and OpenAI sold distribution to Wall Street, the Pentagon turned policy into a revenue cap, and OpenAI's CFO sought IPO cover.