Physical design software: the only moats in SaaS

Alternative Titles

• Why CAD, EDA, and simulation incumbents survived every cloud disruption
• The 40-year oligopoly that cloud-native startups never cracked
• How Autodesk, Cadence, and Synopsys built the deepest moats in software
• Generative AI may finally crack physical design's impenetrable vendor lock-in

Key Takeaways

• Physical design software (CAD, EDA, simulation) is the rare software category untouched by cloud-era disruption. The same oligopoly formed in the early 1980s — Autodesk (1982), Cadence (1983), Bentley (1984), PTC (1985), Synopsys (1986), plus Mentor Graphics (1981) — still dominates today, largely via consolidation through acquisition rather than new entrants.
• The industry's modern structure emerged when workstations and PCs decoupled hardware from software. Incumbents like Applicon and ComputerVision sold $150K–$300K integrated machines; new entrants sold software-only at ~20% of the cost with 80% of the functionality, outsourcing sales to channel partners — a textbook Christensen-style disruption that hasn't been repeated since.
• Four structural moats explain the lack of disruption: (1) technical complexity — EDA tools have millions of LOC, and startups need 5+ years just to reach feature parity; (2) switching costs — decades of IP, retraining, and downtime, with design flaws far costlier than software bugs; (3) GTM — Fortune 100s are locked in, Autodesk is the school/certification default for SMBs, and there's no consumer bottom to wedge in from; (4) hardware engineers hold expensive licenses but lack purchasing power, and hardware orgs have low novel-tech absorption.
• The general CAD market is a four-way oligopoly. Autodesk's biggest business is actually AEC/BIM (Revit, AutoCAD Civil) — nearly 2x its CAD revenue — and it owns mid-market/SMB via AutoCAD and Inventor. Dassault leads automotive/aerospace (CATIA enterprise, SolidWorks for students/mid-market). Siemens leverages PLC integration and has the largest PLM business (NX used by Apple and SpaceX). PTC is smallest; its 2010s AR/VR and factory-analytics bets via early acquisitions largely underperformed.
• EDA has consolidated to Synopsys, Cadence, and Siemens EDA, and Synopsys's pending Ansys acquisition will consolidate simulation further. Synopsys and Cadence are among the highest-valued software companies globally (Cadence ~$78B). Cost-of-chip-design inflation has actually slowed alongside performance gains, and verticalization/chiplets/custom-silicon demand is rising — but incumbents hold massive proprietary training-data advantages that make AI-based startup entry risky from an IP standpoint.
• The physical design value chain spans five stages: requirements/conceptual design (Alias, Rhino3D), detailed design (general CAD plus domain-specific tools for chips, PCBs, buildings, infrastructure, plants), simulation (Ansys leads — FEA, CFD), design for manufacturing and verification/validation (still largely manual, especially V&V in chip design where flaws waste millions), and PLM (parts trees, BOMs, CAD version control) dominated by Siemens, Autodesk, and Dassault.
• Challengers have almost universally failed to break in. The most notable exit is OnShape, founded by a SolidWorks co-founder in 2012 and sold to PTC in 2019 for $470M — not a unicorn outcome. No startup has reached the top tier. The author's conclusion: gimmicky generative-AI features won't scale in risk-averse hardware organizations; a challenger needs to focus on mid-market/enterprise and deliver improvements of multiple orders of magnitude, not incremental gains.
• Three forward-looking areas where the author sees real openings: (1) generative design / topology optimization — Autodesk is already training a 3D foundation model called Bernini; realistic targets are simple components like brackets and nails, not whole vehicles; (2) design for manufacturing and constructability — digitized supply chains plus additive manufacturing growth create demand for platforms that check manufacturability earlier in the design loop; (3) automated physical design QA — LLMs summarizing requirements, flagging design issues against manufacturing/material data, and proposing optimizations as the most tractable near-term AI application.

Readwise Summary (short)