A dedicated chip normally takes 18–36 months from architecture definition to tape-out. On July 18, 2026, Moonshot AI released Kimi K3 with a demonstration: an agent-driven pipeline ran autonomously for 48 hours and completed the design, optimization, and verification of an inference accelerator prototype using only open-source EDA tools and the Nangate 45nm process library. A month later, on August 25, 2026, OpenAI presented the first measured data for its in-house inference chip Jalapeño at Hot Chips — TSMC 3nm, co-developed with Broadcom, nine months from architecture to tape-out. Both events point to the same thing: the capability boundary of AI coding is moving from "writing application software" toward "building physical hardware."
⏱️ Two Timelines: 48 Hours and 9 Months
The Kimi K3 line is officially an "early proof of concept": 48 hours of continuous autonomous operation, input of an open-source EDA toolchain plus the Nangate 45nm cell library, output of an inference accelerator prototype capable of running Kimi's own Nano model.
The Jalapeño line is a real chip. Project started mid-2024, taped out November 2025, first measured data delivered August 25, 2026. Engineering samples are already running OpenAI's own models in the lab, with small-scale deployment planned for late 2026 and scale-up in 2027.
The two numbers are not on the same scale. Kimi K3 used a 45nm open-source library, generations behind the current 3nm/2nm frontier, and the company itself states there was no tape-out, no returned silicon, no test board, and no independent reproduction. Jalapeño is real silicon on TSMC N3P, already back from the fab and running. Reading this as a "who is faster" contest gets it wrong — these are two different routes.
🔬 Kimi K3's Chip: 1.46 Million Cells in 3.98 mm²
| Dimension | Value | |---|---| | Process library | Nangate 45nm open-source cell library | | Area | ~3.981 mm² (often rounded to 4 mm²) | | Modules | 13 | | Standard cells | ~1.46 million | | On-chip memory | 0.277 MB SRAM | | Compute array | INT4 MAC array with on-chip dequantization | | Timing | Converged at 100 MHz | | Simulated throughput | 8721 tokens/s |
The model itself is a 2.8-trillion-parameter MoE with a 1M-token context window, activating 16 of 896 experts per inference. Beyond the chip, Kimi also wrote a Triton-like GPU compiler from scratch — MiniTriton (an MLIR-based tile-level IR, optimization passes, PTX code generation) — matching or exceeding official Triton and torch.compile on Roofline benchmarks.
💥 After 48 Hours: EDA Stocks Fell for a Day
On the first trading day after the announcement, Synopsys fell as much as 12% intraday, closing down ~8%; Cadence fell 7.85%. The panic spread along the supply chain: Zhipu dropped 28% in Hong Kong, MiniMax 16%, and SoftBank 9% in Tokyo.
Sell-side interpretations split into two camps. Union Bancaire Privée's Vey-Sern Ling argued that if US enterprises shift to Chinese models and reduce Anthropic spending, capex and ultimately chip demand would shrink. Morgan Stanley's Gary Yu framed it as the result of long-term accumulation in China's AI industry rather than an overnight disruption. Bloomberg Intelligence's Robin Zhu called the sell-off a reasonable reaction.
A month later, the narrative reversed. When OpenAI revealed Jalapeño's measured data and the "LLMs will topple EDA" narrative peaked, Synopsys and Cadence shares actually rose. Yicai's September 9 retrospective explained why: K3 attempted to prove "AI can work without EDA giants' tools," while Jalapeño — a chip with real deployment — remains deeply tied to commercial EDA. Physical verification, sign-off, and advanced packaging were still done by commercial and deterministic tools. The market switched from "replacement panic" to "assistive value-add."
🔥 Jalapeño's Measured Data: 1.5–1.9x per Watt, 1.7–3.6x Lower Latency
Testing used SemiAnalysis' public InferenceX benchmark, comparing against NVIDIA GB200 and GB300 systems on three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.
| Metric | Jalapeño result | |---|---| | Peak AI throughput per watt | 1.5–1.9x the compared systems | | End-to-end latency | 1.7–3.6x lower (about 28%–59% of rivals) | | High-interactivity workloads | 2.1–4.1x higher | | Kimi K2.5 1T | ~1.5x per watt, 3.4x lower latency | | Rated power | 700 W; sustained ≤550 W under test loads | | Memory | 6 HBM4 stacks, 216 GB total, 15.4 TB/s bandwidth |
The GB300 reference uses 288 GB HBM3E and is rated at 1400 W. OpenAI VP of hardware Richard Ho said AI systems usually trade latency against throughput, and Jalapeño aims for both.
Four caveats go together: the comparison excludes Rubin; data came mainly from OpenAI, with SemiAnalysis verifying only some InferenceX runs in the lab and no AgentX agentic results; not all production workloads were covered; and Jalapeño uses single-token prediction while some public NVIDIA production data uses multi-token prediction, so the lead could change under uniform settings.
🤖 The More Interesting Part: AI Writing Kernels for the Chip
Engineers wrote kernels in OpenAI's in-house language Gluon; a kernel can span thousands of lines and needs repeated manual tuning against the chip's structure. OpenAI then let internal Codex work with GPT-Astra on this task, porting three open-weight models that were not on the production plan to Jalapeño within two months. In some GPT-OSS attention and MoE modules, AI-generated implementations ran 1.5–1.8x faster than the original human-expert versions (OpenAI notes this applies only to selected modules and cannot be extrapolated). Other reports cite a 56% area optimization for BF16 multipliers.
This loop touches something deeper than performance. ASICs have always delivered strong performance around specific workloads, but every new model means rewriting kernels, adapting operators, and maintaining compilers and debug tooling. GPUs are expensive, but CUDA's value lies in spreading those engineering costs across a huge developer ecosystem. OpenAI is showing another path: rather than first replicating CUDA, let models automatically generate, test, and optimize low-level code, lowering the adaptation cost of new models on its own silicon.
This path is internal-only for now. Jalapeño mainly serves OpenAI's own models and workloads; external developers cannot yet freely deploy and maintain applications as they would with CUDA. It shows OpenAI can reduce CUDA dependence internally — not that it has built a market-facing alternative ecosystem.
🧭 The NVIDIA Relationship Isn't Over
Two weeks before the data release, on August 17, NVIDIA announced credit support for SB Energy's PORTS-Pike campus in Ohio, reportedly up to $105 billion per the Financial Times. NVIDIA isn't directly lending to OpenAI: SB Energy builds and operates, OpenAI signs a 20-year lease, and the campus will exclusively use NVIDIA's compute platform.
Jalapeño does inference only, not training. Richard Ho told Bloomberg: "NVIDIA is a great partner, and we still need a lot of NVIDIA." OpenAI and Broadcom announced a 10 GW multi-generation custom accelerator partnership in October 2025, targeting deployment from H2 2026 and completion by end of 2029. Second-generation Jalapeño is in late development with tape-out expected in months; third-generation concept design has started.
⚖️ Three Unresolved Hurdles
| Unverified item | What it is | |---|---| | Kimi K3's chip | 45nm open-source library; no tape-out, no silicon, no test board, no independent reproduction; officially an early proof of concept | | Jalapeño's math | Per-watt figures use publicly rated package power, not same-lab measured consumption; mass production, yield, HBM/packaging costs, and long-term operations remain uncounted | | Software ecosystem | Every new model, operator, and runtime requires continued stack adaptation; Jalapeño hasn't yet proven this long-term software cost comes down |
A more basic boundary: the three test models are large but not the newest open models from each lab. How quickly newer, more complex models adapt to Jalapeño — and how they perform — remains to be seen.
🔁 Where This Sits in AI Coding History
Over the past two years, AI coding's main storyline has been whether models can handle repo-scale changes, with evaluations moving from single-file completion to SWE-bench to long-horizon autonomous tasks. Kimi K3's 48 hours and OpenAI's nine months push that line one step further: when an agent can work continuously for two days without breaking down and iterate inside a verification loop, it can take on not just software repos but physical design flows described by deterministic toolchains.
But this extrapolation has a hard constraint. Chip design is full of physical verification, sign-off, and advanced packaging steps that rely on commercial EDA's deterministic tools — places a probabilistic language model cannot carry the load. Jalapeño's nine-month tape-out depended just as much on Broadcom's deep involvement and the commercial EDA foundation.
Three numbers to watch over the next 12 months: real inference cost reductions after Jalapeño's large-scale deployment, whether second-generation tape-out timelines compress further, and the ratio of human labor displaced by Codex-like tools in chip kernel adaptation.
📚 References
1. Yicai/Sina Finance: LLMs enter semiconductor design — will EDA "tremble"? 2026-09-09 2. NetEase Intelligence/Tencent News: OpenAI's first chip stuns, first generation hits the frontier, 2026-08-26 3. Zhidx: Kimi K3 technical report and weights release, 2026-07-28 4. Cailian Press/Cnyes: Tape-out in 9 months — OpenAI's custom ASIC unveiled, 2026-08 5. SemiAnalysis InferenceX benchmark / Hot Chips 2026 (2026-08-25) 6. MartechAI: OpenAI's Jalapeño AI Chip Shows Higher Speed and Power Efficiency in First Benchmarks