{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidc3gtmkk6ugmzgwzb5jhlafp4eoywhic2udpzs7evkvm25rlhjwi",
    "uri": "at://did:plc:ls6rbbwjyqeakittcsi3k6x3/app.bsky.feed.post/3mkkarsaoixv2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreic2x46kpxg74tsdmwhtnoqtlhbu6f4vavnyjtghe5gzwpo6m2koe4"
    },
    "mimeType": "image/png",
    "size": 471622
  },
  "description": "Cadence's CEO told the Street something almost no other software CEO has the conviction to say in 2026 — agentic AI doesn't replace the base tools, it calls them 10-100x more often. So if that's true, where exactly is it showing up — and why is it showing up there first, not somewhere else?",
  "path": "/cadence-1q26-rev-agents-dont-replace-the-base-tools-they-call-them-harder-cadence-1q26-agent-bu-hui-ti-dai-base-tool-zhi-hui-ba-ta-jiao-de-geng-qin/",
  "publishedAt": "2026-04-28T09:16:20.000Z",
  "site": "https://www.jasonandjarvis.org",
  "tags": [
    "The Software Industry's Value Migration in the AI Era",
    "Chip Design's Agentic Moment",
    "Cadence's €2.7bn Acquisition of Hexagon D&E",
    "Synopsys 4Q25: Ansys to the Rescue",
    "Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文",
    "Subscribe now"
  ],
  "textContent": "Subscribe\n\nCadence's 1Q26 print is the kind of report that's easy to misread.\n\nRead the headlines and the story is unusually clean: revenue $1.474bn, Non-GAAP EPS $1.96, both clearing the upper end of management's own guidance and sell-side consensus. Management lifted FY26 revenue growth guidance from +11%–+13% straight to +16%–+18%, and for the first time wrote Rule of 60 into a forward commitment. Backlog hit a record ~$8.0bn.\n\nBut the part of the call that actually deserves to be lifted out isn't any of those numbers. It's what Anirudh Devgan said when Charles Shi pressed him with the sharpest question of the night — a sentence almost no other software CEO would have the conviction to deliver in the 2026 environment:\n\n\"What's different about agentic AI is that it doesn't replace the core EDA engines. It calls them more often, and it calls them intelligently.\"\n\nIf that sentence is true, then most of the market read on this quarter is missing the plot. Everyone's busy doing the arithmetic on Hexagon's $0.28 of dilution and the +$0.08 of organic EPS raise. What's actually changing the underlying physics of Cadence's business model is the extra $200mn that landed in backlog — that, not the P&L, is the first live microscope on agents calling base tools 10-100x more often.\n\nLet me walk through 1Q26 in that order.\n\n### I. The sharpest question of the night: if vibe coding can already replicate a SaaS, why is EDA an exception?\n\nThe single highest-information exchange on the call came from Needham's Charles Shi.\n\nHe didn't ask whether agentic AI would drive more consumption — that's the consensus that's already been pre-priced into the stock. He asked the sharper inverse question: could the base tools themselves be rewritten by AI? And he hooked it with a deliberate jab — \"there are also EDA startups happening at the same time.\" This is the reflex skepticism software investors have learned over the course of 2025: when vibe coding can already clone a vertical SaaS feature set in a fortnight, why on earth would EDA's moat be the exception?\n\nI worked through that question in The Software Industry's Value Migration in the AI Era, where I argued that what gets commoditized is the \"selling finished features\" layer of SaaS, while the moat shifts upward to orchestration and governance. Devgan's answer is essentially that same logic, instantiated inside the EDA vertical — only he calls it the \"three-layer cake.\"\n\nCadence Three-Layer Architecture and AI Super Agent Portfolio\n\nsource: Cadence 1Q26 Earnings Presentation\n\nThe bottom layer is accelerated compute and data, carried by GPU-accelerated platforms like Millennium. The middle is principle simulation and optimization — the solver engines accumulated over decades. The top layer, finally, is agentic AI: ChipStack (RTL design and verification), ViraStack (analog and custom), and InnoStack (digital implementation and sign-off), together spanning the full chip design flow. Devgan's core judgment is that the value comes from the tight coupling across these layers, not from any single one.\n\nI covered ChipStack's nine-sub-agent architecture in Chip Design's Agentic Moment — at the time it was still a product-launch posture. At CadenceLIVE Silicon Valley this quarter, Cadence added two more super agents to the family: ViraStack pulls analog into the agentic envelope (a notoriously hard category to automate, which is why Cadence also built something called AvedaStack, a method of embedding \"skills\" into the agent itself), while InnoStack handles digital implementation and sign-off. Sitting above all three is AgentStack, a head-agent framework that does the orchestration.\n\nStand back from the architecture and what Cadence is doing looks almost diametrically opposite to what Salesforce or Workday are doing. Those companies are wrapping the \"sell seats, charge by seat\" SaaS layer inside an agent. Cadence is treating its own base tools — the solver engines — as the agent's computational substrate, so that as agent deployment density rises, base tool usage gets stepped up rather than displaced.\n\nThis is the most fundamental difference between EDA and a generic SaaS in the AI era. A SaaS that gets replaced is essentially \"turning a database into a UI\" — a job that agents are getting cheap enough to do. EDA, by contrast, is \"solving a complex physical equation numerically until convergence\" — and the agent will never \"understand\" that solver on its own. It can only call it more frequently, more intelligently.\n\nDevgan offered a remarkably concrete way to picture this — the single passage from the call I'd most want to lift out and frame:\n\n\"Imagine a chip with 100 blocks and 100 engineers. When each engineer runs 1 block, they typically run only 1 or 2 experiments to compare settings. When agents run those blocks, they might run 10 or 100 variations.\"\n\nTranslated into investor language: **agents don't replace base tools — they raise the calling density of base tools by one to two orders of magnitude.**\n\nAnd once base tools get called more often, the economics flow back to whoever sells the base tools.\n\nAs for Charles Shi's hook — \"EDA startups are happening at the same time\" — Devgan's answer never quite confronts the counterfactual of \"could AI write a better base tool?\" head-on. He uses elimination: \"I'm not worried that some other party will be able to write any better base tools. Our competitor of the base tool is anyway best in class.\" That competitor is Synopsys, not some vibe-coded EDA startup. This is a sufficiency argument, not a necessity argument. But for a company with 10,000 R&D engineers, 1,000 of them top-tier PhDs, sufficiency is enough to underwrite the current valuation multiple — at least the kind of multiple this +18.7% top-line print supports.\n\n### II. The dual-track monetization: sell agents on one side, wait for base tools to be called several times more on the other\n\nIf base tools get called orders of magnitude more often by agents, how does Cadence actually collect on it? That's the real business-model question buried inside 1Q26.\n\nDevgan's answer is two tracks.\n\n**Track one is the new agent tools themselves.** This track runs on subscription-plus-consumption pricing (similar to other leading AI tools), and what's being delivered replaces the analog/RTL/sign-off work clients used to do with humans. On this track, Cadence's customers have given Devgan an extraordinarily persuasive pricing anchor. He quoted a big customer verbatim: \"Whether it's analog or digital, every new design needs 2x more engineers — and that kind of headcount growth simply can't be hired.\"\n\nThat's the pricing anchor for ChipStack/ViraStack/InnoStack — it isn't being benchmarked against a customer's EDA budget, it's being benchmarked against engineers the customer literally cannot hire. A tool that lets clients \"bend the headcount curve\" structurally has a pricing ceiling that doesn't sit inside the EDA budget line.\n\n**Track two is base tool usage growth itself.** This track runs on Cadence's traditional strength — a hybrid of seat-based and consumption-based licensing. Devgan said it directly: \"base tool usage is going up pretty significantly in the current environment.\"\n\nThe boundary condition here is worth pulling out — could a customer write its own super agent and route around Cadence's layer entirely?\n\nThe historical answer was: they always could. In the pre-agentic era, Cadence already gave customers TCL or Python interfaces to write their own flows, and customers always want differentiation. But Devgan offered a striking field report: after Cadence walked customers through demos of InnoStack, ViraStack, and ChipStack, the reaction was —\n\n\"Oh, there's no point writing these kind of agents.\"\n\nThat single line carries more narrative weight than any sell-side moat description. The customers chose not to build it themselves — not because they couldn't write agents, but because they couldn't write the tight coupling between agent and solver. The lower-level interaction Cadence builds at the API call layer is precisely the surface a customer's own agent can't reach. The solver, after all, isn't theirs.\n\nStack the two tracks together and what agentic AI does to EDA TAM is multiplicative, not additive. Devgan provided a long-arc anchor: EDA used to occupy 7% of customer R&D, now it's around 11%, \"and R&D spend itself will go up significantly\" — so EDA is taking a larger share of a growing pie. He stayed deliberately measured on the TAM question: \"I tend to print things rather than predict things.\" But he did mention, almost in passing, that \"all the big CEOs\" he speaks to are \"not only willing, they want to see that happen\" — customers' own CEOs are actively pushing for more spend on automation and compute.\n\n### III. Backlog is where agentic AI shows up on Cadence's books for the first time\n\nBacklog and cRPO Six-Quarter Trend — Both Hit Record Highs in 1Q26\n\nsource: Cadence 1Q26 Earnings Presentation\n\nIf Devgan's \"agents calling base tools 10-100x\" story is real, where on Cadence's financial statements should it have shown up first in 1Q26?\n\nNot on the P&L — the P&L was largely consumed by the accounting drag of Hexagon. The $0.28 dilution and the +$0.08 organic raise both look — and read — like deceleration.\n\nThe cleanest place to look is backlog.\n\nA few attributes set the $8.0bn number apart from anything on the P&L. First, Hexagon contributed only ~$75mn to 1Q26 backlog — meaning core organic backlog managed a QoQ net add even in the seasonally soft Q1 window. Second, Q1'25 backlog was $6.0bn, Q4'25 was $7.8bn, and 1Q26 came in at $8.0bn — the typical Q1 seasonal step-down pattern just got broken. Third, 2026 is, by Cadence's own disclosure, a light renewal year (similar to 2022 on annual-value terms). And yet, as Wall put it directly: \"those are some of the strongest growth years for us, because of all the add-on activity.\"\n\nStitch those three facts together and the conclusion is hard to escape: customers are adding capacity outside the renewal cycle.\n\nWhich maps perfectly onto Devgan's \"agent runs 100 variations rather than 1\" picture. The base tool capacity baked into a historical contract isn't sufficient under an agentic workflow — so customers come back and add on. Goldman Sachs put it more bluntly in their write-up: in their view, agentic AI is the single largest medium-to-long-term revenue acceleration catalyst for Cadence.\n\nBut Wall paired it with a critical hedge:\n\n> \"We're not assuming a sudden step function in AI monetization in the guide... we're obviously being disciplined in our 2026 outlook.\"\n\nDevgan added: \"I used to say two contract cycles is generally true — but because this is a new category of labor productivity expansion-type demand, with base tool usage rising in parallel, agentic AI monetization could happen sooner than two contract cycles.\" Then he immediately walked it back: \"I don't want to predict too much, and like John said, we are not putting it in our guide.\"\n\nThat last sentence — \"not putting it in our guide\" — is the line from this call I'd most want to remember. Management's posture toward agentic AI is \"willing to show it in print, unwilling to commit to it in a forecast.\" Which is why the report has a strange aesthetic on the page: the guide is conservative, the P&L is accounting-loaded, the backlog is record-breaking, and the call narrative is unmistakably agentic. Those four things won't be priced into the stock simultaneously — they'll cash out at different speeds.\n\nCadence itself seems clear on the order of operations: digest Hexagon over 2026, then let agentic AI do the talking in 2027.\n\n### IV. Physical AI is the second curve that arrived inside the Hexagon box\n\nIt's tempting to flatten the Hexagon D&E acquisition into \"$160mn of revenue plus $0.28 of EPS dilution.\" But Devgan offered a second reading — Hexagon as the \"middle layer fill-in\" for Physical AI.\n\nHe's been carrying the conviction that \"physical AI will be bigger than data center AI by a long shot\" for 5 years now. By \"physical AI\" he means AI extending out of pure digital domains into autonomous systems, automotive, drones, robotics — a world where system simulation gets redefined as a workload. The leading structural and multi-body dynamics technologies that Hexagon D&E brings — combined with Cadence's earlier acquisitions of Millennium, Cascade, and BETA CAE — finally complete the principal simulation + optimization layer in physical AI's middle tier.\n\nIn Devgan's framing, this stack is what lets Cadence \"narrow the critical sim-to-real gap\" — closing the distance between simulation and reality, enabling customers to build and train fundamentally new AI world models. That isn't the same story as the $160mn top-line line item. That's SD&A buying a ticket onto a new curve in multi-physics simulation.\n\nDevgan also walked through a second-order transmission worth flagging: physical AI doesn't only feed SD&A and the AI layer — it loops back to support Cadence's EDA and IP businesses. He cited Tesla management's public remarks about \"silicon shortage, physical AI being part of the cause,\" then added: \"Cadence has always had analog and digital capabilities together in our solutions — that's why we've been strong at every major automotive semi customer and at every major systems and OEM customer. When that translates over to drones and robots, it'll turbocharge the silicon business all over again.\"\n\nThat, more than anything else, is the meaning of the Hexagon deal beyond the P&L — it doesn't just make SD&A $160mn bigger, it expands Cadence's addressable demand pool from chip design into physical-world simulation. I worked through the electromechanical convergence logic in Cadence's €2.7bn Acquisition of Hexagon D&E, and physical AI is precisely the demand sink that logic was always pointing toward.\n\nA small product-level detail Devgan kept returning to: Jensen Huang signed a Millennium box on stage at CadenceLIVE. It's a moment of theater, but also a piece of narrative shorthand — under the expanded NVIDIA partnership, Cadence is now layering agentic AI design capabilities across three target arenas at once: chip design, physical AI systems, and hyperscale AI factories.\n\nCadence's other ecosystem move is the joint optimization with Google: ChipStack on GCP, paired with Gemini — combining LLM reasoning with cloud-scale compute, delivered as a cloud-native chip development platform. MediaTek, separately, has done a \"wide-ranging expansion\" covering agentic, core EDA, 3D IC, and system analysis end to end. Read the three partnerships together and the signal is: at the agentic layer, Cadence isn't expecting customers to grow the capability themselves. It's stitching together hyperscaler compute + foundation models + Cadence base tools into a three-way assembly, and bringing the whole stack to the customer's doorstep.\n\n### V. A short financial coda at the end\n\nI had a much longer Hexagon-financials section in the first draft. Re-reading, I realized that section isn't really the protagonist of the 1Q26 call — it's more like the cleanup work that needs to happen across all of 2026. So I'm going to compress it.\n\nHexagon D&E's Immediate Q1 Balance-Sheet Impact\n\nsource: Cadence 1Q26 10-Q\n\nThe short version: of the $225mn FY26 revenue raise, $160mn is Hexagon and $65mn is organic. The $0.20 EPS cut is the net of Hexagon dilution at $0.28 and an organic raise of +$0.08. GAAP OPM took a one-time hit of 425bps and Non-GAAP OPM was cut 125bps. The bulk of that comes from acquired intangibles amortization stepping from $105mn in FY25 to $286mn in FY26E (yoy +172%). Wall himself characterized 2026 as an \"integration year,\" and explicitly used BETA CAE's 2024-2025 absorb-and-rebound pattern as the template for Hexagon, with 2027 expected to return to accretion.\n\nIt's worth flagging the accounting nature of the EPS dilution. Hexagon's consideration was 30% stock and 70% cash — meaning the largest source of short-term EPS dilution isn't operating margin drag, it's the cash interest income that disappears when 70% of the consideration is paid out of the cash on the balance sheet. That's exactly why the BETA pattern is the right template: once cash interest income comes back, most of the physical form of the dilution simply evaporates.\n\nRule of 60 Path Over Five Years\n\nsource: Cadence 1Q26 Earnings Presentation\n\nDevgan turned FY26's Rule of 60 expectation into an explicit management commitment. From 57 in 2022 to 61 in 2026E, Cadence took four years to walk that line. The quality of the line is worth examining from both sides: on the constructive side, the 2024-to-2026E leg was driven by Non-GAAP OPM rising from 42% to ~44% alongside revenue growth rebounding from +14% to +17% — a margin-and-growth dual driver crossing 60. On the cautionary side, FY26E Non-GAAP EPS growth lands at +10%-+11% — the first single-digit print in five years (FY21–FY25 ran 18%/30%/21%/16%/20%).\n\nIf you're underwriting Cadence at 45x forward P/E, you're buying the revenue acceleration. If you're indexing on EPS growth, what you're holding is mid-stretch deceleration. These two perspectives diverged for the first time after 1Q26 — which is why sell-side has been quietly rolling its valuation anchor from CY26E to CY27E. BofA lifted its price target to $400 while bumping the multiple from 40x to 42x and re-anchoring to CY27E EPS of $9.47 — essentially discounting a year of \"accounting noise\" forward into the future. That kind of \"anchor roll\" is trusting the framework, not the quarter — which only works if you also trust the BETA pattern and agentic monetization both deliver.\n\n### VI. A few things I'll be tracking\n\n**First, the cadence of H2 organic upside.** Wall is already telegraphing \"appropriate prudence.\" JPM is already calling for another raise in July. Cadence's historical pattern is \"conservative early, raise progressively through the year\" — but 1Q26 has already burned $65mn of organic raise upfront. How much ammunition is left, and how quickly it gets fired, is the central data point at the July Q2 call.\n\n**Second, whether agentic AI's \"visibility\" inside backlog continues to be validated.** Wall already framed 2026 as a light renewal year that nonetheless tends to be among the strongest growth years — that's pulling add-on activity to the front of the narrative. If Q2/Q3 backlog keeps printing record highs in a soft renewal year, agentic AI converts inside Cadence's mechanics from a narrative into a quantifiable demand stream.\n\n**Third, whether base tool defensibility gets re-priced by the sell side.** Charles Shi's question this quarter was a concentrated projection of the 2025–2026 software-investor mindset — it's not going away, it's coming back in sharper forms. Cadence answered with a \"we use our own AI to write our own base tools\" elimination argument; sooner or later the market will demand an answer that satisfies both sufficiency and necessity. Synopsys's head-on contest in multi-physics post-Ansys integration will be the next stress test Cadence has to absorb — I traced part of that contour in Synopsys 4Q25: Ansys to the Rescue.\n\n**Fourth, the customer landing pace and pricing anchor for ChipStack/ViraStack/InnoStack.** Devgan's \"2x engineer headcount\" is a remarkably strong pricing anchor, but it remains, at this point, narrative. If H2 produces the first explicitly outcome-priced super-agent contract of meaningful size, that would be the moment Cadence's agentic monetization moves from \"in print\" to \"in guide.\"\n\nThe actual value of an agent is that it swaps the human engineer's mental model — \"1 block, 1 or 2 experiments\" — for the electricity bill of \"100 blocks, 100 variations each.\" That race has no finish line, because the solver can never enumerate every variation. It only steps the EDA consumption curve up by an order of magnitude as agent deployment density rises.\n\nCadence is the first company to articulate the mechanism cleanly. The $8bn backlog in 1Q26 almost certainly isn't the terminal value — but it is the first place on Cadence's own books where that curve became visible.\n\n#\n\n### Earnings Call Recap\n\n\n   Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文\n\n\nI will now share a more detailed analysis of this earnings result behind the paywall:\n\nUpgrade your subscribtion?\n\n### This post is for subscribers only\n\nBecome a member to get access to all content\n\nSubscribe now",
  "title": "Cadence Design 1Q26 Review: Agents Don't Replace the Base Tools, They Call Them Harder",
  "updatedAt": "2026-04-28T09:18:14.021Z"
}