{
  "id": 12421062,
  "title": "AI is now capable of developing its own inference hardware",
  "url": "https://urgent.news/2026/10/06/ai-is-now-capable-of-developing-its-own-inference-hardware",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-06T16:23:25.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://github.com/FeSens/openTPU"
  },
  "original_language": "en",
  "account": "An open-source AI accelerator called openTPU has been developed by AI. This accelerator is capable of designing and building its own hardware for inference purposes. It addresses two key questions: how far can AI agents go in hardware design, and can they create the chip that runs their own inference?\n\nThe openTPU accelerator is designed to run ten modern models with their real weights on an Inspur YPCB-00338 card, which has a Xilinx Kintex-7 xc7k480t processor and two DDR3 channels. The accelerator produces the same tokens as the simulator, bit for bit. The performance of the models on the card was measured using the production image deploy_champ_e698dcd7, with LFM2-2.6B, SmolLM3-3B, and Phi-4-mini models showing faster decoding times compared to the previous production image.\n\nThe design of openTPU is simple and explained in detail through its instruction set architecture (ISA), which is described in docs/isa.md. The accelerator consists of a sequencer, a matrix unit, a vector unit, and a quantizer, with no cache and no hidden scheduling. Every data movement is an instruction, allowing for a detailed trace of the run and its speed.\n\nThe openTPU accelerator runs an AI model called Qwen3.5-4B, which has an int8 image over 4 GiB. The logits are streamed back while the card runs, with 4-bit decode being faster than the previous production image. The open-source nature of openTPU allows for changes to the ISA, simulator, and RTL, while maintaining passing tests with python3 -m pytest -q.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}