The impact arrives in three stages: harvest existing hardware, change the bill of materials, then design for the scale control used to prevent.
Physics was never the only blocker. The energy, bandwidth, and parallelism of wave hardware have been visible for decades. What capped designs was the cost of keeping thousands of drifting analog parameters true. Engineers sized the system to what control could hold. Remove that ceiling and three things happen, in order.
Duty cycle converts to revenue. In controlled simulation matched to published device specifications, conventional servicing left hardware offline for much of the operating window at scale while the runtime held specification continuously at one acquisition per cycle. For anyone selling machine time, uptime is the top line. This is where pilots live today.
simulatedPower-on to specification in a handful of parallel reads can reduce per-unit commissioning and make fabrication spread easier to absorb. Looser tolerances and shorter test time can change product economics. The claim remains conditional on the component response, control authority, and envelope.
simulatedOnce the contract is assumed the way noise margins are assumed, “how many channels can we hold?” can leave the design review. Very large apertures, wafer-scale photonic weight banks, and dense sensing arrays become design candidates rather than automatic control failures.
projectedAI inference is dominated by linear algebra, which wave propagation can implement physically. The recurring deployment problem is not the matrix multiply alone. It is per-unit commissioning, drift during operation, and accuracy between service passes.
In controlled simulation, a 100,352-weight analog inference layer reached its 97.5% digital baseline from cold start and held it at full duty under drift. The uncontrolled system fell to 73%, while stop-and-recalibrate servicing delivered one ninth the useful work. The physical accelerator transfer remains unverified.
The stack is hybrid: digital orchestration above, the wave substrate performing the forward pass, and the stability contract below. On-device AI is a strong candidate because a part in a pocket must survive thermal change and recover without a service visit.
simulatedTraining, semantics, routing, planning
Physical linear computation and sensing
Maintain the qualified response profile
Wave hardware does not replace digital computing. Training remains digital and inference becomes heterogeneous. Conversion into and out of the wave domain costs energy the runtime does not remove. Analog precision has floors. Ecosystems move slower than hardware. Some device families on the class map remain untested.
The sober claim is that a large class of computation and sensing can migrate into previously difficult physical substrates, one qualified device and one design win at a time.
Nobody thinks about drift, the way nobody thinks about signal restoration in a laptop. That is the highest state infrastructure reaches.
Check your hardware