INTEL PUBLISHED OPENINFER’S NUMBERS. MIRELO SHIPPED ALL SUMMER.

Intel publishes OpenInfer’s benchmark: 3x faster inference on “just a CPU”.

OPENINFER   Intel Xeon 6 SoC benchmark, published in Intel’s Data Center Content Library.

On 19 August, Intel published OpenInfer’s benchmark results in its own Data Center Content Library: “How OpenInfer Delivers 3x Faster AI Inference on Intel Xeon 6 SoC.” Intel’s own framing is the good part: OpenInfer turned “just a CPU” into a 3x faster AI inference engine. Three weeks earlier, OpenInfer had published the fleet-level version of the same argument. Two identical four-GPU fleets, the same four models, the same open-loop request traffic. The baseline pinned one model to each GPU on vLLM. OpenInfer pooled all four models across every GPU and let the scheduler place each request. Result: 53.9k tokens delivered inside SLA against 21.4k, zero rejected requests against 1,725, and peak fleet utilization up from 21.5% to 43.5%. Same hardware, 2.5x the sellable tokens. That table is the whole thesis: the binding constraint in agentic inference is not silicon, it is scheduling.

Mirelo shipped all summer: a new API, a viral tool, a model with Kyutai.

MIRELO   Speech-Preservation API · Audio-to-MIDI · MuScriptor with Kyutai.

Three things in eight weeks. The new Speech-Preservation API fixes the problem every video-agent developer hits: regenerate a clip’s ambient sound and the dialogue underneath it dies. Mirelo’s keeps the speech intact and rebuilds the effects around it, up to a minute of video. Then a major Audio-to-MIDI update: tempo map export, meter detection, score view, per-instrument tracks, and a mix that drops straight into a DAW. And MuScriptor, released with Kyutai, is now transcribing hundreds of songs a day on Kyutai’s server alone.

The Campus Scout Program opens a UK chapter.

CAMPUS SCOUTS   The new UK cohort. Kickoff on 12 August, VC 101 on 10 September.

Six new scouts joined over the summer, our first concentrated push into the UK, from Oxford to King’s College London, working alongside Nick, our venture partner in London. The reasoning is simple: the European companies we want to meet are usually being built in a lab or a lecture hall a year before anyone writes about them.