You can design a chip to run Kimi K3 at 87000 tps???
A technical deep-dive into Moonshot AI's Kimi K3 inference chip architecture—the bugs we found, the mistakes we made, and how we achieved ~10x higher frequency.
Christina Lee · August 5, 2026 · 20 min read Several amazing friends texted me after seeing Kimi K3's autonomous chip design demo in their technical blog.1 It's super creative to demo the recursive setup of "model designs its own chip to run itself." Although it is not a commercial setup, it reveals that LLMs can, in a certain setup, go through the full RTL-to-GDS flow end to end. It's quite a precursor of something bigger happening beyond the software development space. Reminder: similar to the Kimi blog post, the chip serves a ~1M-parameter "nano model," not the full 2.8T-parameter K3.…
saved by
related reading
- Redesigning the Inference Chip: From Nvidia GPU's Flaws to OpenAI Jalapeñozartbot.github.io
- AI Chip Architecturesjepeake.com
- The Inference Shift – Stratechery by Ben Thompsonstratechery.com
- An Interview with MatX CEO Reiner Pope About LLM Chipschipstrat.com
- AI Chip Architecturesjacobpeake.com
- Harness Engineering for Self-Improvement | Lil'Loglilianweng.github.io
- MatX: High-throughput chips for LLMsmatx.com
- OpenAI Jalapeño: Better Than Nvidia Blackwellnewsletter.semianalysis.com
- Taalas is what Etched should have been.zach.be
- KernelBench: Can LLMs Write GPU Kernels?scalingintelligence.stanford.edu
- Kimi K3 Tech Blog: Open Frontier Intelligencekimi.com
- Gimlet Labsgimletlabs.ai