Price per token times volume was roughly right for chat. It is now wrong in three directions: who consumes it, how often you read memory, and where inference happens.
The hardware ceiling rises over years. The software floor falls over months.
Memory went vertical because the bus ran out of room. A large software company cut its AI bill 90% by routing away from frontier models.
One raises the ceiling over years. The other drops the floor in months.
Unlimited free chats, open weights with revenue sharing, and Chinese labs competing on price. Capability at the top stopped being the product and cost of serving became it. Memory and power are the real constraints now.
Jeff Dean left Google for AI in science. Move's designer joined Anthropic for AI security. Meta shipped a codebase agent and reported a sandbox escape. Four items, one pattern: the hard problems moved to verification and containment.
Texas halted new data center approvals, SpaceX is targeting 10 gigawatts by 2027, Anthropic signed $10B with a startup cloud, and a bitcoin miner leased $4.7B of Norwegian capacity. The AI constraint moved from silicon to power.
🔧 The model race gets the headlines. The compute contract decides who is still in it. On Anthropic's $10B deal, modular inference pods, rack scale as the new unit of purchase, and the one foundry underneath all of it.
Amazon EKS now gives Kubernetes upgrades a seven-day undo window. That will not replace testing, but it changes the operational math: teams can validate under real traffic without treating every minor-version upgrade like irreversible surgery.