The economics of local intelligence
Cloud inference rents compute by the token. Local inference prices it at the cost of electricity. The crossover between those two curves is the most consequential number in software economics right now.
Two price systems
Every query against a cloud model carries a metered price: per token, per minute, per some unit the vendor defines. Every query against a local model carries a price too — the electricity to move the bits — but it is three orders of magnitude smaller and paid to a power company, not a software one.
These are not just different prices. They are different price systems, and they produce different software. Metered intelligence pulls products toward metering: usage limits, seat pricing, throttling, and the entire apparatus of software that must watch its own consumption. Fixed-price intelligence permits the opposite: software that sells capability outright, the way software did before the service era, because the vendor's cost no longer scales with use.
The crossover point
For any workload there is a volume at which owning computation becomes cheaper than renting it. The crossover depends on four variables: model efficiency, silicon, electricity, and what the frontier demands. Three of the four move in local intelligence's favor — models get smaller per unit of capability, consumer silicon keeps absorbing accelerator blocks, and the same watts do more work each year.
The fourth variable runs the other way: frontier capability is expensive, and workloads that genuinely need the frontier stay in the cloud longer. The honest summary is that local intelligence wins from the middle of the capability distribution downward, and the middle keeps rising.
What crosses first
The pattern we observe in the lab's own work: repetitive workloads cross first — the query you run a thousand times, where the meter never stops. Private workloads cross next, because sending data to a cloud has a cost that never appears on any invoice. Latency-sensitive workloads cross after that. Public, novel, bursty workloads cross last, maybe never.
None of this is an argument that local replaces cloud. It is an argument that the boundary is economic, moving, and knowable — and that the businesses built on the winning side of the boundary will look structurally different from the ones built on the other.