Measured

The Energy Institute no longer uses the substitution method to estimate primary energy. Here's why that matters for our charts on energy.

I built a doorway so my AI partner could see, hear, and speak with me.

Two weeks of CPU inference optimization, and the thing that won wasn

One Protocol, Three Places to Keep StateOne rmcp v3 server - the SDK was three days old at measurement. Three state policies behind one trait (sticky memory, Redis store, sealed continuation), plus a held gRPC stream as reference (different transport and

Measured tokens/sec for open LLMs on Apple Silicon: Llama, Qwen, Mistral, DeepSeek via Ollama on M4 - median of 3 runs, full methodology and raw data. Updated as our fleet grows.

Same product, same download page, same week. Mac converted about 6 of 10 downloads into a first launch. Windows converted about 2 of 15. Here is the measurement.

10x more tokens per megawatt. CoreWeave has the first measured performance of NVIDIA Vera Rubin NVL72, showing 10x improvement in tokens per second per megawatt on DeepSeek-R1 compared to Blackwell.