Ten weeks in: four tracks, one convergence.

Watch how the customer was asked to pay, because it changed three times. First the consolidation read: AI would compress the software stack the way every capability wave does, and the labs would push into their customers' businesses to find revenue. Then tokenmaxxing — more tokens looked like more value until the invoice arrived. By midsummer, enterprises were reining in employee AI as metered costs bit; one executive said flatly they had created a monster. Then came the correction in late June and early July: the number that matters is not the token, it is the completed task. Cursor benchmarked Opus at roughly $11 per finished coding job against its own model at 55 cents. The buyer pays for work done, not sticker price.

Under all three sits the read we filed in the first financing window: capital got certain before returns did. Money flowed down the stack into infrastructure while the deployment layer stayed contested, and that contest never resolved. Application revenue was the most-worked hypothesis in the register — the sign of a layer where cash is real, pricing keeps moving, and staying power is untested.

The wider record adds the concession that keeps this honest. Real books of business already exist in pockets: Suncorp scaled AI into claims workflows, Rio Tinto into logistics, and Databricks reached a reported $188 billion valuation on enterprise data it already holds. Those outcomes are genuine. They are not yet widespread, and not what most of these valuations assume. Moving fast was never the question. Moving fast on a foundation no pricing cycle has tested is.

The frontier arc ran in four moves. It began as a race — whoever had the best model won — and at the opening of the arc we read the contest as cost-per-watt at frontier capability. Convergence followed: by early June the top Chinese and US coding models sat within a point of each other, and by midsummer the labs had stopped claiming the model was the constraint, naming their own capability overhang. Open weight came next, with GLM-5.2 shipping at roughly one-sixth of closed-model pricing in late June. This week it sits at open source as deliberate strategy: Kimi K3, open-weight by design, reportedly beating a Western flagship on coding benchmarks and moving markets as it did.

One honest wobble sits in the middle of that arc. Inference cost read both ways for four straight weeks. The China gap logged both ways in the export-control fortnight — not through weakness, but because the order cut two directions at once, concentrating capability in US hands even as it fractured the global market. Calling those weeks clean in hindsight would be the easy lie. The read stayed ambiguous until the evidence closed it.

The wider record adds a qualifier that matters. Open models now lead on volume — nearly half of tracked business tokens on one major router running through Chinese systems, per reported industry data — yet frontier labs still capture most spending, because their prices are higher and buyers keep paying for the hardest jobs. Capability commoditised before revenue did. That is the same shape as the price-versus-proof gap one track over: activity racing ahead of the money underneath it. Once frontier capability is a download, the model stops being a defensible advantage, and that fact drives most of what the other three tracks are sorting out.

Most coverage frames this as the United States against China. Nine weeks of coverage say it is three-cornered, and the third corner explains this week. Inside the West there are two engines. The labs compete on the frontier — raw capability, speed, and the push into customers' businesses that the compute bill demands, visible this week in Anthropic's services venture. The application incumbents compete on something older: hold on data, customers, and workflow. They spent the week defending it, with Microsoft coaching sellers to knock the labs on cost and trust rather than capability. The week before last we read the labs as cornered rather than predatory, and the incumbents answering.

The third engine is Chinese state coordination, and it plays by neither Western rulebook. It is patient where competition is fast, directional where incumbency is defensive, and it needs no return on the price floor it sets. Kimi K3 and Xi's "AI for all" keynote are not a lab competing on metrics. They are a state playing a longer game: concede cheap inference, set the global floor, and let Western capital strand itself against a price that does not have to make money.

Two concessions keep the frame honest. The corners are porous, not walled. The same companies that fight also partner: Snowflake with the clouds, the labs with the hyperscalers they rent from, OpenAI and Anthropic increasingly with each other. The frame leaves Europe out, which spent the window chasing partial sovereignty rather than picking a side. Neither breaks the read. None of the three engines has won: Western labs bet speed decides it, incumbents bet customer hold outlasts the frontier, China bets a plan outlasts them both. Anyone telling you which wins is guessing.

Every track above assumes the machines stay on, and says nothing about who pays when they do. Power was the steadiest call across the run, and it escalated on a schedule. Late May it entered through the ratepayer thread, when PJM wholesale capacity had roughly doubled on data-centre demand. Early June it became a permitting queue, with more than 60 percent of 2027 US capacity not yet under construction. Mid-June it became federal policy when FERC voted to fast-track connections on the operators' terms. Two weeks ago it became a constituency fight — capacity pricing up roughly tenfold across two years, steelmakers on the same grid calling the bill unsustainable.

This week it became law. New York halted large data-centre construction outright: the first statewide moratorium in the nation, and not an outlier but the first to land of many. More than a hundred proposals have surfaced at local, county, and state level, with bipartisan support across fourteen states, and roughly seven in ten Americans say they do not want a data centre nearby, per recent polling.

This is where the story stops being about markets and lands on a person. The electricity a data centre needs is the electricity the household already pays for, and the split is now in the tariffs. In Arizona, the largest utility proposed a 45 percent rate rise for extra-large users that are mostly data centres — and a 14.5 percent rise for ordinary residential customers in the same filing. Read those two numbers together and the politics of the whole build-out sits in one rate case: the industry's bill and the household's bill are climbing together. When a valuation outruns its earnings, the investor carries the risk. When a build-out outruns public consent, the public carries it, and that is the one no IPO can price.