The October 6, 2026 filing says femtoAI opened its developer community after a beta program. The wire headline and the filing body do not say the same thing. The weaker line is "10-fache Einsparung". The document I trust is the PR Newswire filing, and the number that decides it is 10.

Die Brief reads the tenfold figure as a vendor target, not a measured boundary. The release lists an SPU evaluation kit, compressed models, a compiler, and Sparsity Studio, but it does not tie the 10x claim to a board, power rail, wattage, memory size, SKU, or benchmark condition. It also says the platform can serve OpenAI Whisper with tenfold lower memory, but the Whisper variant is not named. For a buyer, the useful question is whether Sparsity Studio can reproduce the ratio on a named workload; for a fab, the useful question is whether the SPU kit exposes enough timing and power data to separate sparsity from compiler effects.

What remains unmeasured is the actual power draw, memory footprint, and accuracy delta under a named workload. Legato and Marshall are named, but their products, models, and test setups are not. No public number confirms the ratio.

Die Brief treats the tenfold ratio as a vendor target, not a measured silicon boundary. The buyer should ask for a named workload, board, and power rail before treating the SPU kit as proof.

The document leaves out measured power draw, memory footprint, accuracy delta, board, power rail, wattage, memory size, SKU, and benchmark conditions.

The filing gives a tool path, not a measurement path. Without board, rail, wattage, memory size, SKU, and workload, the 10x claim cannot be separated from compiler, model, and sparsity effects.