AI stopped looking like a chat product this week and started looking like a fight over energy, access, and who gets to keep the work that teaches them.
The clean story would be that the models got better again.
They did. DeepSeek V4 arrived with a price tag low enough to make another year of expensive API spend look less automatic. Chinese open weights kept eating the local and hosted stack. The US lead narrowed enough that "American models are obviously better" became a habit worth testing, not a fact worth repeating.
But the week was not really about another benchmark jump. It was about the parts of the system that decide whether a model can matter at all.
Meta cut 8,000 people and called it an AI strategy. EY handed 130,000 auditors an agent. Oracle moved closer to owning its own power supply. Maine showed that a state legislature can say no to data-center megawatts. OpenAI quietly began the Assistants API sunset, reminding builders that today's abstraction can become tomorrow's migration ticket.
Put those together and the pattern is less glamorous than the launch notes. AI is becoming an operating assumption, and operating assumptions have bills.
For buyers, the important question is no longer "Which model is smartest?" It is "Which dependency can I survive?"
If your product relies on a specific OpenAI interface, you now have a migration date to manage. If your internal tooling assumes the expensive frontier model is the only credible option, DeepSeek and Qwen deserve a real benchmark on your own work. If your procurement policy treats Chinese model weights as just another open-source input, the Pentagon blacklist story says legal and operational reality may not agree.
The same split is happening in infrastructure. An ASIC cloud built for agents, Oracle's energy posture, and Maine's data-center politics all point to the same boring truth: inference is not weightless. The people promising unlimited agents still need chips, substations, water, zoning permission, and a power bill someone accepts.
That matters because the labor story is getting less abstract.
Meta's cuts make "AI-driven restructuring" feel like the standard memo, not a warning sign. EY's audit agent lands on the junior and mid-career work that used to teach people how to become senior. The humanoid pilot in Erlangen will not replace a warehouse this quarter, but it is collecting the motion data that lets management ask a colder question later: which tasks do we stop backfilling?
The chrome around the web is changing too. Chrome's second-brain pane means more readers arrive with a summary already loaded. They click through to verify, not to understand from scratch. That moves value toward clear sources, strong first paragraphs, and work that cannot be flattened into a pane.
The seller move is not to promise intelligence. Everyone is promising intelligence.
The seller move is to prove continuity. Show the migration path when an API dies. Show the cost-per-agent-run, not just the token price. Show where the workload can route if the default model becomes too expensive, too regulated, or too politically awkward. Show the human promotion ladder after the agent takes the training work.
That is the uncomfortable part of this week. The capability curve kept moving, but the real pressure came from dependencies becoming visible.
The week in one line: model advantage is becoming less durable than access to power, permission, and a clean exit plan.