The economics of local intelligence
Inference has a cost curve. Where it crosses the cost of electricity on hardware you already own, the market for software changes shape.
Two cost structures
Cloud intelligence rents compute by the token. The price is legible, the capacity is elastic, and the marginal cost of a query never reaches zero. Local intelligence has the opposite profile: a fixed cost in hardware you may already own, a marginal cost near the price of electricity, and capacity that is strictly bounded.
These two structures produce software with different economics. Rent-priced intelligence favors products that meter usage. Locally-priced intelligence favors products that sell capability — the way offline software once did — because the vendor's cost no longer scales with the customer's usage.
The crossover
For any given workload there is a crossover point: a volume of inference at which owning the computation is cheaper than renting it. The point moves every year — down, as model efficiency improves and consumer silicon absorbs accelerator blocks, and up, as frontier capability demands more compute per query.
Mapping that crossover for realistic workloads is empirical work: measuring real token throughput per watt on ordinary machines, tracking it over hardware generations, and understanding which categories of software cross over first. Repetitive, private, latency-sensitive workloads cross first. Novel, public, bursty ones cross last.
Why a lab studies markets
Because the technology is only half of any product decision. Knowing when local inference becomes the economical default tells the lab where to direct engineering — and where a small team can build something a rent-scale organization cannot easily copy: software whose costs do not grow with its users.