Thompson distinguishes access to open model weights from the cost of serving them: avoiding training costs does not eliminate inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → expense.
His central financing concern is timing: infrastructure investment may outrun the cash flows needed to sustain it, even if the underlying technology keeps improving.
Treat his views on platform winners and capital constraints as an investment thesis, not as measured evidence of future model performance.
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
Separate model progress from the capital and operating costs required to deliver it. Thompson’s argument offers a framework for evaluating infrastructure availability and the durability of subsidized AI pricing.
