Cadence Stops Renting Intelligence — and the Market Stops Paying Up
A record backlog, a raised year, and a 24% drawdown: Maestro is Cadence's bid to own the model layer, but Wall Street is already discounting the agentic AI trade.
CDNS · Earnings Call · 2026-10-26
The quarter that arrived in a drawdown
Cadence's Q3 was, on the numbers, the best quarter it has ever reported. Revenue of $1.633 billion grew 22% year over year, “Backlog reached a record of $8.2 billion... we are raising our revenue outlook for the year to 21% growth and our EPS outlook by $0.30 to $8.40” — Anirudh Devgan, President and Chief Executive Officer · 2026-10-26. All product groups grew double digits — core EDA 20%, IP 24%, SD&A 30% — and the prior filed quarter already showed the ramp: revenue of $1.584 billion, up 24% year over year with GAAP operating margin at 32.2%, up 10.2 points. And yet the tape says something else entirely. CDNS closed 2026-10-23 in one unbroken 17-week decline of −18.8%, 24% below its June 2 peak of $416.39. That gap — accelerating fundamentals, contracting multiple — is the real subject of this report. 17.7x price-to-revenue is well off the 20.8x cycle peak, and price-to-earnings sits near 75x.Maestro: from renting models to owning them
The genuinely new material in this call was not another super-agent endorsement — there were ten of those, and Renesas' 34x productivity, Samsung Foundry, Socionics' 5x, Nokia's 5x on Aura Stack are all now familiar furniture. The new thing is Maestro: a portfolio of domain-specialized models, post-trained by Cadence on Cadence data, wrapped in an intelligent router. Anirudh Devgan was blunt about why this is only possible now: “there is this emergence of very, very good open weight models... now it's finally possible with open weight models, and it's actually very cost effective to post train these models” — Anirudh Devgan, President and Chief Executive Officer · 2026-10-26. The claimed payoff is up to 30% better quality and 5x better token efficiency versus general-purpose foundation models. This is a strategic reversal dressed as an upgrade. In July, Devgan's framing was that customers were demanding choice in their LLMs — “Can the agent be more intelligent in choosing the right model for the right task given the rapid progress in the LLMs?” — Charles Shi, Analyst · 2026-07-27 — and that Cadence would orchestrate other people's models. Now Cadence is building its own specialized models and routing between them and third parties, with model routing as the explicit cost lever. The frontier model is now a component to be sourced, not a partner to be deferred to — a direct answer to a competitor's frontier-model-centric partnership announced the same day, which one analyst pointedly labelled "GPT Synopsys." The confluence is worth noting: open weight models entered the market-wide top-75 keyword list in the last two quarters, so this is not a Cadence-only realization — it is the industry discovering that post-training a good open model is cheaper than renting a great closed one. The difference is that Cadence's version of "cheap" comes with a physics-accurate simulation engine attached.The monetization fog, and the competitor's royalty
What has not changed across four quarters is the answer to "how do you charge for this." In April it was "subscription plus consumption." In July, John Wall said “generally, our economics scale with customer adoptions” — John Wall, Senior Vice President and Chief Financial Officer · 2026-07-27 while cautioning the guide assumed no step function. On this call the phrasing was again about optionality: “our strategy remains centered on delivering high-value engineering software rather than building a large general-purpose AI infrastructure business” — John Wall, Senior Vice President and Chief Financial Officer · 2026-10-26, with customers free to bring their own LLM infrastructure and absorb their own model compute costs. That is a deliberately conservative posture, and it is also why recurring revenue growth slipped to 18% from 20%. Devgan attributed the gap to China comps and pointed to the longer arc — 11% last year, 18% now — but the more interesting exchange was Vivek Arya asking directly about a competitor pitching mid-teens growth, 50% margins, and a royalty-based model worth nearly $1 billion of incremental top line. Devgan's answer — that royalties are "just one vector" and Cadence already has them in Tensilica and design IP — was the least specific moment of the hour. There is a second, quieter signal: the company has stopped treating hardware as a lumpy afterthought and now sells it as a structural design size story. Devgan relayed TSMC's roadmap claim that chip sizes multiply 48x over five years.Hardware inventory rose again in Q3, which management framed as pre-building against backlog rather than a demand signal — but it is the cleanest volume-based lever Cadence has, and it does not require anyone to believe a word about agents.hardware is normally sold or consumed based on design size. So if the design size is bigger, then you need more and more hardware... TSMC showed their process roadmap and said in the next 5 years, the chip size will go up by 48x.