Back to HomeA few watts.
Gated edge AI
A few watts.
Make each one count.
For drones, radar and machines that cannot carry a data centre.
Research & prototypes
20W
A human brain
14,300W
One eight-GPU AI server
945TWh
Data centres worldwide by 2030
Sources: NIST; NVIDIA DGX B200, maximum; IEA, projection.
How it runs
Quiet until it counts.
- Gate
Relevance
A tiny always-on check. 98% of background sensor noise stops here.
- Budget
Specialists
Only the models the decision needs, inside the platform power budget.
- Output
Report or abstain
The verifiable evidence and energy cost travel with every decision.
Measured
Count the whole cost.
- •Selection, sensing, movement and inference.
- •Savings reported beside the misses.
- •No valid path, no forced answer.
STAGED FLIGHT TRIAL TELEMETRY
0.83%
of frames woke the expensive specialist step on staged flights (12 of 1,443 frames). 0.35% on empty flights (7 of 2,010), reporting nothing.
Read preprint findings →Applications
Where it runs.
Which decision has to run at the edge?
Bring the platform and the power budget.