The compute bill can come back quickly.
SpaceX's call put payback, contract duration and the ability to reprice compute beside the capex headline. The question is how quickly deployed capacity can help finance what comes next.
The publication
Writing.
Technical work, economic consequences and the space between the two. Selected articles and dispatches, with dates and links to their original publications.
Published writing
The bill gets our attention. What it buys takes longer.
Meta and Anthropic expose two sides of the same investment question: spending can create new capability, while constrained capacity can limit the revenue used to judge that spending.
Micron: below past cycle-low valuations.
A comparison with documented memory-cycle lows asks what investors are paying for expected earnings—and what has changed in the contracts supporting their duration. Late 2018 is an explicit exception, not a detail to hide.
AI is also becoming a source of funds.
Semtech connects lower-power optical engineering with higher supplier value and operating cash. Its turnaround shows why the funding discussion also needs to examine businesses already generating money from the build-out.
The sell-off. The build-out.
A violent sell-off is set against earnings, contractual funding and engineering disclosures from OCP. The analysis follows the specific mechanisms that can change deployment, productivity and cash rather than making the price chart do the explaining.
AI doesn’t need everyone to get it.
Disclosures from five banks show how a small group can build systems that benefit far more people than ever touch a frontier model. Some productivity gains improve the systems needed for the next round of adoption.
The bull market had a bear inside it.
The broad index hid a severe semiconductor drawdown. Looking beneath it changes the sustainability question: what kept earning, what improved, and what did the share price already require?
AI is rewriting the software bill.
Jane Street and the emerging agent stack show how customers can buy less packaged software while becoming much larger infrastructure users. The application may survive; its old economics do not get the same guarantee.
AI can work. The gains can go elsewhere.
Lower costs can reach customers, fund more work or displace an old supplier without raising the adopter's profit. That makes an unchanged earnings estimate an incomplete test of whether AI is earning a return.
The rate changed. So did the business.
A Treasury move is not a complete funding model. Cash generation, customer advances and contractual commitments alter the mechanism through which rates reach an AI business.
The answer was right. Then the world changed.
A good answer can become expensive when the world changes and nobody reopens it. Continuing delegation requires evidence, durable state, permissions and the ability to recognise when an old conclusion has stopped being safe to use.
There is no memory cycle.
An AI deployment plan is also a memory deployment plan. Follow the connection from learned weights and persistent agents to the rack, the factory and the customer contract.
These pieces record an argument at a particular time. Figures and market observations belong to that date; the original publication carries the full evidence and qualifications.