Synthesis history

Three substantive additions this pass. Moreh's 21K aggregate tokens/s benchmark on AMD MI300X [^25133] introduces a per-request vs. aggregate throughput distinction that sharpens a core tension. Qualcomm's CEO projectin…

Kog AI's primary technical blog post (23206) surfaced in the feed, confirming the monokernel approach has published documentation — though no specific claims were extractable from it in this pass. Community discussion s…

Kog AI, an inference startup, has claimed approximately 3,000 tokens per second on 8× AMD MI300X GPUs and 2,100 tokens/s on 8× NVIDIA H200 with a 2B parameter model — roughly 20–30× faster than the ~100 tokens/s typical…