Spend pushes vertical integration into the customer’s stack — and they respond.

Follow the buyer arc in compressed time. Experimentation and token-maxing became a CFO line item, then a harder arithmetic: cost per token against cost per human hour, and how much extra output the gap actually buys. An outsized repositioning, in a short margin of time. This week the data on the record started drawing implied strategies to the surface — and the largest companies in the world are not going to be pushed around.

Apple is the clearest read. The complaint names two defendants by name — Tang Tan, OpenAI’s chief hardware officer, and Chang Liu — and alleges they took confidential hardware information on their way out, and more than four hundred former Apple employees have since joined OpenAI.

The second comes from the person with one of the best seats in the market. Satya Nadella, quoted in the Journal, wondering aloud how an enterprise retains value if it depends entirely on someone else’s foundation model. Understand it from where it comes. Nadella is a formidable operator; he will position Microsoft where it is strongest — and Signal 008 already read that play as trust and governance, not frontier lock-in: model-agnostic orchestration, cost discipline underneath, plugging AI into customer workflows only where the margin works. The move is valid. It is not proof that the largest enterprises have suddenly become frontier-model dependent. The frontier promise is efficiency, cost, and innovation. Dependency on anything you do not control is a house built on sand.

The third is the Nadella playbook, three ways. Palantir on governance. Snowflake on open-weight routing inside the perimeter. Databricks on agents where the install base and margin already are. Not the seismic moment the media wants you to react to — world-class operators competing for the most consequential market on the planet, and the strategies are starting to rhyme.

The capex line is priced. The new data point is how Amazon’s US$25 billion bond issue cleared. Bloomberg reported investors sold SpaceX, Alphabet, Nvidia, Meta and Oracle paper to make room — new debt funded by exiting old debt, the well metered rather than closed. Barron’s read cheap credit and called it fine. Both can be true for an afternoon. What tightens the loop is the lender asking who repays — ahead of the customer finishing the dependency sum.

GPT-5.6 Sol hit the market at roughly a third of Anthropic’s Fable 5, a model whose independent benchmarks rank slightly ahead, with ChatGPT Work in the same release. Anthropic expanded Cowork; Meta pushed agents into coding. The week is data, not drama.

This is the competitive mechanic we already flagged. H6 — China catching up — is redefining inference price outside the West and, through open weights and subsidised models, setting the floor Western labs must answer to. Barron’s reports GLM-5.2 and DeepSeek on state-supported terms; Zhipu raised US$4 billion the same week, per the Journal. Commodity inference gets conceded; margin moves to workflow and per completed task. Any industry would expect that shape. What is remarkable is the speed — a global repricing compressed into months, which is the measure of how fast the capability is actually moving.

Scale is the escape route — and scale runs through power. PJM capacity moved from roughly US$29 to US$329 per megawatt-day between 2024 and 2026, per Ars Technica; the Journal reports gas-turbine backlogs for data centres as long as eight years. Steelmakers on the same grid say the bill is unsustainable. Signal 002 flagged this thread; it is a constituency fight now, not a wholesale-market abstraction.

Australia makes the pattern portable. The AFR reports Anthropic seeking roughly 1.4 gigawatts of local capacity — near the scale of the entire existing industry — while negotiating copyright exemptions ahead of an IPO, as the Reserve Bank names the build-out in the inflation fight. Enterprises are pricing lab dependency. Governments run the same arithmetic with more leverage — power, labour, culture, timed to a listing.

That arithmetic will not stay in filings. Data-centre jobs pitched against higher electricity bills, cost of living and inflation is tailor-made for local media — misinformation, disinformation and legitimate grievance in the same cycle. Expect the fight for power access to run socially and politically, in public, until the stack finds something closer to energy abundance.