Moonshot's Kimi taped out a chip in 48 hours and knocked 9% off the EDA duopoly

Moonshot AI says its open-weight Kimi K3 model took a chip from RTL to tapeout simulation in 48 hours: 4mm², 100MHz, 13 modules, on the public Nangate 45nm cell library, with no licensed Cadence or Synopsys software anywhere in the flow ¹². The task, as the demo's write-up states it, was a nano-K3: a chip built to run a nano-scale version of the model itself ².

The market reached for its DeepSeek template. On 17 July Cadence and Synopsys fell roughly 9% and the Nasdaq about 1%↗, and holders of both were suddenly pricing one question: what a tool licence is worth when an agent does the designing ².

Twenty-year-old silicon, by the CEO's own arithmetic

Anirudh Devgan priced the demo down on the 27 July earnings call↗: a small block, on technology about 20 years old, running 20 to 30 times below current frequencies, and even at that old node the designers needed EDA tools to finish it ³. Open-source EDA has existed for decades and lives in universities; for real designs, he said, people use Cadence ³.

Asked about a future model that emits GDSII straight to a foundry, his answer was that "the ground truth will prevail," and that the commoditization he does expect sits at the model layer, where customers want choice among LLMs, not in the tools ³.

The sell side lined up behind him. BNP called the selloff unwarranted, noting NVIDIA already runs Cadence's ChipStack agent internally at scale ⁴. Mizuho called the risk "way overblown," reading K3 as an orchestrator of open-source flows such as OpenROAD rather than a replacement for deterministic solvers ⁵.

The quarter that argued back

The numbers behind the calm are company-claimed but on the record: Q2 revenue of $1.584 billion, up 24%, a record $8.1 billion backlog, full-year guidance raised to 19% growth, IP up more than 40% ³. ChipStack, the verification agent, counts more than 20 customer engagements and has cut a typical five-week verification cycle to under a day; ViraStack, for analog, counts more than 25, at 2x to 10x gains ³.

On 22 September Cadence shipped an agent that writes RTL↗, claiming 24% less area and 18% less power than raw foundation-model code, with Honda R&D evaluating ⁸. None of it is yet a separate agent line in the reported numbers; the CFO called the raise broad-based, with no sudden step function assumed from agents ³.

The tell is the pricing

The rebuttal carries its own commercial shift. Per 36Kr, the only outlet reporting the 9 September Goldman fireside↗, Devgan is moving agents to a consumption-plus-subscription model while the tool layer keeps its licences, and his own arithmetic concedes the shrinking base: a CPU that once took 500 people five years now takes 30 to 40 people six months ¹.

Analyst Karl Freund reads the same economics moving the bill to verified design throughput rather than classic seat licensing ⁶. Mizuho frames it as expanding EDA monetization from tools into engineering labour, potentially tripling the core addressable market ⁵. Bloomberg Intelligence adds the caution: EDA and design IP are about 15% of chip R&D spend, and AI as the front door "may shift software value up the stack" ².

If consumption pricing turns the physics layer into a metered commodity, the licence base under both duopolists' multiples is what erodes, and no agent revenue has been reported yet to replace it ³.

What decides it

The weights landed on 27 July with a 47-page technical report↗, open for replication with one caveat built in: Moonshot's licence requires a separate commercial agreement from any Model-as-a-Service operator above $20 million in annual revenue ⁷. Cadence guides Q3 revenue to $1,595 to $1,625 million ³. The question for shareholders of both duopolists is not whether agents design chips; it is which layer sends the invoice.