ACQ2: The Software Behind Silicon (with Synopsys Founder Aart de Geus and CEO Sassine Ghazi)

2025-03-05 - source - Read full transcript

Ben Gilbert (host)David Rosenthal (host)Aart de GeusSassine Ghazi

Key insights

Synopsys was born from an accidental layoff, not a planned startup.

Aart de Geus and a team of mostly summer students at General Electric had built innovative synthesis tools; when GE was hit by the 1985 semiconductor downturn and exited the business, de Geus was told he'd be laid off. Rather than take the technology, the team told GE about their startup idea in full transparency and negotiated a spinout with GE support, later paying GE $23 million in stock value when Synopsys went public.

founding-story-and-company-longevity

Logic synthesis succeeded because it delivered undeniable, immediately verifiable results to skeptical early customers.

Early customers who handed over circuits they'd spent weeks hand-optimizing were shocked when Synopsys's software returned versions 30% smaller and 30% faster within hours. Customers initially disbelieved the results, checked for two weeks, then became evangelists - their critical feedback (pointing out where the tool 'wasn't that great') directly improved the product, turning the first two-to-three dozen customers into de facto co-developers.

founding-story-and-company-longevity

EDA tools had to earn the trust to autonomously change a circuit, which was previously taboo.

Before synthesis, the field was called 'computer aided design' because tools only assisted human decisions; software that changed a circuit outright was considered evidence of bugs. De Geus described Synopsys as having 'license to kill' - the first tools trusted enough to autonomously modify a design, a trust threshold each subsequent wave of automation (including AI in 2017-2018) has had to re-earn.

trust-and-verification-in-ai-tools

Even when AI-driven design tools produced measurably better results, engineers resisted them for roughly two years until they could accept opaque optimization.

Synopsys introduced AI for synthesis and place-and-route around 2017-2018. Ghazi says results were consistently better from the start, but users refused to trust or adopt the tool because they wanted to understand exactly what the AI changed - an explainability demand that has since faded as AI became broadly normalized in the industry.

trust-and-verification-in-ai-tools

EDA's AI use case is fundamentally different from generative AI because it tolerates zero functional error.

De Geus argues the industry's AI optimization differs from consumer generative AI (which can be '90% accurate') because a single violated design rule can zero out chip yield entirely. Synopsys therefore layers extensive verification steps after any AI-optimized output before committing to manufacturing, since a shipped bug is catastrophically expensive to fix post-fabrication.

trust-and-verification-in-ai-tools

EDA has stayed a two-company market because the barrier to entry is decades of cumulative, compounding domain knowledge, not just a trainable model.

Ghazi explains that new EDA entrants can't just 'train a model and create an output' - the field requires accumulated learning across 25+ years of revolutionary but evolutionarily-delivered technique changes (citing 1997-era crosstalk capacitance lessons as still load-bearing today). This cumulative-knowledge moat is why the market consolidated to essentially Synopsys and Cadence, paralleling why no one can simply recreate TSMC.

eda-as-critical-infrastructure

Moore's Law is not a natural law - it is the output of continuous, deliberate cleverness by companies like Synopsys.

Ben Gilbert makes the point explicitly: unlike a physical law, Moore's Law only continues because companies keep re-inventing how to sustain it. De Geus frames Synopsys's contribution as roughly 10 million X cumulative productivity gain, and argues the next order of magnitude (10-100-1000X) will require fundamentally new approaches beyond single-chip density scaling.

moores-law-and-systemic-complexity

Chip scaling has shifted from a manufacturing-limited problem to a physics-limited (thermal, mechanical) problem.

Ghazi describes how, even when a transistor works correctly in isolation, packing billions together (e.g. NVIDIA Blackwell's 208 billion transistors) creates severe thermal issues that can cause mechanical warpage and cracking - forcing physics-level analysis (Synopsys's rationale for acquiring Ansys) into what used to be a purely electronic design problem.

moores-law-and-systemic-complexity

Continued performance gains now depend on connectivity between chiplets, not just transistor density on a single die.

De Geus argues that single chips are approaching a practical size ceiling (~1.5 inch square), so future gains come from splitting functionality across multiple dies connected via high-bandwidth, low-energy interposers - a shift he calls 'SysMoore,' extending exponential-improvement thinking from the chip level to the whole multi-die system level.

moores-law-and-systemic-complexity

Synopsys's relationship with foundries changed from one-way 'enablement' to deep, joint invention over the last six-plus years.

Ghazi explains that Synopsys used to simply take process specs from foundries (TSMC, Samsung, Intel, GlobalFoundries) and adapt its own tools ('enablement'). Now hundreds of Synopsys engineers sit embedded with foundry teams during process development itself, because pushing to smaller nodes requires co-inventing physics and design techniques together rather than handing off finished specs.

eda-as-critical-infrastructure

Synopsys's customer base has shifted from almost entirely semiconductor companies to nearly half system/OEM companies.

Fifteen years ago essentially 100% of Synopsys revenue came from chip companies; today roughly 45% comes from system companies (automakers, consumer electronics OEMs, etc.) that either design their own silicon or need to architect complex in-house electronics even without designing a chip - reflecting both AI-driven silicon differentiation and Synopsys's expansion into system-level virtualization tools.

silicon-to-system-expansion

The Ansys acquisition extends Synopsys from chip-level EDA to system-level multi-physics simulation.

Ghazi frames the January 2025 Ansys deal around two vectors: (1) core EDA increasingly requires deep physics simulation (thermal, structural) that Ansys leads in, and (2) system companies (e.g. automakers) need to simulate an entire product as a 'digital twin' spanning electronics, mechanical, and other physics domains before manufacturing - positioning the combined company as 'silicon to system.'

silicon-to-system-expansion

Companies

Synopsys - The $80-billion Electronic Design Automation (EDA) company at the center of the episode; spun out of General Electric in the mid-1980s by Aart de Geus and six colleagues after GE exited semiconductors.
General Electric - Aart de Geus's employer in the mid-1980s, where he and a team of mostly summer students developed the synthesis technology that became Synopsys; GE supported the spinout and later received $23 million when Synopsys went public.
Cadence Design Systems - Synopsys's main competitor in EDA; the two companies effectively make up the entire modern EDA market alongside Synopsys.
TSMC - Founded three months after Synopsys; discussed as a parallel case of an industry pivot toward fabless design plus dedicated foundries, and as a company Synopsys works with directly on process co-development.
NVIDIA - Described as mission-critical Synopsys customer; Jensen Huang appeared on stage with Sassine Ghazi at GTC (arriving from the parking lot minutes before going live) to affirm Synopsys's role in NVIDIA's chip design.
Ansys - The physics simulation and analysis company Synopsys announced acquiring in January 2025; framed as extending Synopsys from silicon-level EDA to system-level multiphysics simulation (thermal, structural, etc.).
ASML - Maker of EUV lithography equipment; discussed as the other essential link (alongside Synopsys and the foundries) in keeping Moore's Law-style scaling alive, with an estimated decade of runway left in EUV.
Sun Microsystems - An early Synopsys customer where NVIDIA co-founders Chris Malachowsky and Curtis Priem worked before founding NVIDIA; Jensen Huang was their LSI Logic contact at the time.
Intel - Sassine Ghazi's employer before joining Synopsys in 1998; also cited as a foundry partner and as a chipmaker with a multi-die architecture (like AMD) similar to NVIDIA's Blackwell.
Ford - Cited as an example of a traditional non-tech company now designing its own chips, illustrating how Synopsys's customer base has expanded beyond pure semiconductor companies.

Techniques and frameworks

Logic synthesis - The founding technology of Synopsys: automatically converting a functional circuit description into an optimized gate-level implementation, replacing manual circuit design and immediately delivering ~30% smaller and ~30% faster results than human designers.
Trust but verify (AI-assisted design) - Aart de Geus's framing for how Synopsys has repeatedly introduced AI/expert-system automation (starting with rule-based expert systems in the 1980s, then ML for synthesis and place-and-route from 2017) while still requiring exhaustive verification because a manufacturing error is enormously costly.
SysMoore - Synopsys's term for extending Moore's Law thinking beyond a single chip to an entire multi-die, multi-physics system, driven by connectivity between chiplets rather than transistor density alone.
Design technology co-optimization (DTC) - Practice of jointly optimizing chip design choices and manufacturing process/building-block choices, rather than treating either as fixed, which Synopsys leads in.
Tech-onomics - Aart de Geus's term for the idea that every technical decision in chip design is simultaneously an economic decision (e.g. gate count determines manufacturing cost, node choice determines whether a chip is commercially viable).

Summary

Ben Gilbert and David Rosenthal sit down with Synopsys founder Aart de Geus and current CEO Sassine Ghazi for a deep dive into Electronic Design Automation (EDA), a field Acquired had never covered directly despite dozens of episodes on the companies that depend on it. De Geus recounts the accidental founding: while at General Electric in the mid-1980s, he and a small team (mostly summer students) built logic synthesis technology - software that could automatically optimize circuit designs rather than merely assist human designers. When the 1985 semiconductor downturn led GE to exit the business and lay the team off, they spun out transparently with GE's support, eventually paying GE $23 million when Synopsys went public. Ghazi, who joined in 1998 after grad school and a stint at Intel, adds the customer-facing perspective on how synthesis results (circuits delivered 30% smaller and faster than months of hand optimization) converted skeptical early adopters into evangelists who fed continuous improvement back into the product.

A recurring theme is trust: Synopsys tools were the first in the industry given "license to kill" - the ability to autonomously change a circuit, something previously considered evidence of a bug. That same trust threshold reappeared when Synopsys introduced AI-driven synthesis and place-and-route around 2017-2018; even though results were consistently better, engineers resisted for roughly two years because they wanted to understand exactly what the AI had changed. De Geus and Ghazi both stress that EDA's AI use case is categorically different from generative AI because it tolerates zero functional error - a single violated design rule can zero out manufacturing yield - so every AI-optimized output still passes through extensive verification before a chip goes to fabrication.

The conversation turns to why the EDA market stayed consolidated to essentially two companies (Synopsys and Cadence): the barrier isn't training a model, it's decades of cumulative, compounding domain knowledge, where a lesson learned in 1997 (crosstalk capacitance) still matters today. De Geus makes the broader point that Moore's Law isn't a natural law but the continuous output of deliberate engineering cleverness by companies like Synopsys - and that scaling has shifted from being manufacturing-limited to being physics-limited, as packing hundreds of billions of transistors together (NVIDIA's Blackwell has 208 billion) creates thermal and mechanical problems that didn't exist at the individual-transistor level. This is driving the industry toward multi-die "chiplet" architectures connected via high-bandwidth interposers, a shift de Geus calls "SysMoore": extending exponential-improvement thinking from a single chip to an entire interconnected system.

That systemic shift also explains Synopsys's business evolution. Fifteen years ago essentially all its revenue came from semiconductor companies; today roughly 45% comes from system/OEM companies (carmakers, consumer electronics firms) that either design their own silicon or need to architect complex electronics even without doing so - a direct result of AI making custom silicon a competitive necessity across industries. Ghazi frames Synopsys's January 2025 acquisition of Ansys as a natural extension of this trajectory: Ansys's multi-physics simulation leadership (thermal, structural, 40+ years of cumulative trust) complements Synopsys's core EDA business and positions the combined company to simulate entire products - "digital twins" spanning electronics and mechanical systems - not just chips. The episode closes on a reflective note, with de Geus discussing how a company that has become genuinely mission-critical infrastructure (for NVIDIA, TSMC, and much of the semiconductor value chain) inherits responsibilities beyond the technical, from geopolitics to energy use to what role companies play as centers of value in a fragmenting world.

Notable Quotes

"We're the only ones that have license to kill... because license to kill means we can actually change a circuit. That was completely taboo before. If a tool did that, it means they had put some bugs in it." - Aart de Geus

"People take Moore's law as if it's derived from the natural universe property... It's not. It literally relies on companies like Synopsis getting clever again. Every time Moore's law happens, it's because somebody got clever again." - Ben Gilbert

"The learning is not just, hey, can I train a model, then create an output, and I'm there? The cumulative knowledge to get to the current state before you look at the future state is massive." - Sassine Ghazi

"Every technical decision is simultaneously an economic decision, be it for the build or for the use side of things." - Aart de Geus

"They who have the brains to understand should have the courage to act." - Aart de Geus