The investor list alone tells you something is shifting. Temasek writing a check large enough to lead. Applied Materials and NVIDIA both in. Siemens and Atomico rounding out the roster. When a London-based artificial intelligence startup announces a $300 million Series C at roughly $2.4 billion valuation—as PhysicsX did this June—and the backers read like a cross-section of the semiconductor equipment stack, it's worth asking what they see that the rest of us don't yet.
Founder Jacomo Corbo has a phrase for it: "Large Physics Models." Think foundation models, but instead of parsing language or generating images, they're trained to accelerate engineering simulation. The pitch is audacious—speed gains of several orders of magnitude over conventional finite-element methods. The company says it grew its team from 150 to over 300, opening offices in the Bay Area and Singapore to complement existing hubs in New York and London.
And the timing, well. Subtle it is not.
Semiconductor equipment spending is barreling toward record territory, propelled by an AI infrastructure buildout that shows no sign of fatigue. SEMI projected in April that global 300mm fab equipment spending would climb 18 percent to $133 billion this year, then notch another 14 percent gain to hit $151 billion in 2027. Gartner, meanwhile, forecasts worldwide semiconductor revenue will surpass $1.3 trillion in 2026. TSMC alone approved roughly $45 billion in capital expenditures back in February, earmarking the lion's share for advanced nodes and another 10 to 20 percent for advanced packaging and masks.
But beneath that capital flood, a bottleneck is taking shape. The computational complexity of designing and fabricating chips at 2nm nodes and beyond has become genuinely daunting. Multi-patterning optical proximity correction can chew through weeks of CPU time. Defect inspection spits out terabytes of data per wafer. PhysicsX, which counts semiconductors among its target sectors alongside aerospace and automotive, argues that conventional numerical simulation simply can't keep pace. That, at least, is where physics-informed AI is supposed to come in.
Whether it actually does—at scale, in production fabs, with the kind of ROI that semiconductor equipment buyers demand—remains a more open question than the valuation might suggest.
The Incumbents Are Already Moving
The semiconductor industry didn't wait for a startup to tell it that AI might be useful. Layering machine learning onto design and manufacturing workflows has been underway for years, and the pace has accelerated sharply in recent years.
NVIDIA's cuLitho, announced in production at TSMC and Synopsys back in March 2024, delivers a documented 15× speedup for optical proximity correction on H100 GPUs. Samsung and other foundries have since adopted it. Synopsys reported in March 2025 that its Proteus OPC platform, running cuLitho, clocked 15 times faster on NVIDIA hardware; broader TCAD and verification suites saw similar gains on Grace Blackwell systems.
Equipment vendors, for their part, are embedding AI deeper into the stack. Applied Materials introduced its SEMVision H20 e-beam defect review system with AI image recognition in January 2025, then followed up this June with packaging-grade e-beam systems—VeritySEM 7AP and SEMVision G7AP—targeting sub-optical defect regimes in 3D architectures. KLA and SPTS lean heavily on AI-powered process control across inspection and metrology. Lam Research's Sense.i etch platform marries in-situ sensors with real-time analytics to optimize chamber performance on the fly. PDF Solutions rolled out Exensio AI/ML for yield ramp last October, adding agentic capabilities via Exensio Studio AI this past March.
Then there's the consolidation wave. Synopsys completed its $35 billion acquisition of Ansys in July 2025, stitching together an end-to-end chip-to-system simulation platform. Cadence answered in February by acquiring Hexagon's Design & Engineering business to bulk up its multiphysics footprint, then unveiled ChipStack—an AI "super-agent" for chip design that the company says has assisted in more than 1,000 tape-outs to date. Siemens, not to be left out, has embedded generative and agentic AI into its Solido custom IC flow and announced a partnership with PhysicsX to integrate physics AI into Simcenter X for computational fluid dynamics workflows.
In other words, the incumbents are neither asleep nor standing still.
Why Physics AI Might Actually Matter Now
Three forces are converging, and they create an opening—perhaps a necessity—for startups like PhysicsX.
First, computational lithography at the leading edge is hitting economic and throughput walls. High-NA EUV systems, which ASML began shipping to Intel and imec in late 2025 and early this year, promise finer resolution but carry eye-watering price tags. TSMC said publicly in April that the cost-to-ROI doesn't justify near-term adoption. Which means multi-patterning and baroque OPC workflows stay in play—exactly where GPU-accelerated and physics-ML surrogates can compress weeks of compute into hours. A January preprint on computational lithography claimed 57× acceleration using cuLitho primitives combined with machine learning, underscoring that there's headroom left for AI-native approaches.
Second, inspection and metrology data volumes are exploding. Advanced packaging, chiplets, heterogeneous integration—each multiplies the number of interfaces and failure modes engineers have to monitor. Applied Materials' new e-beam systems for packaging defects and KLA's process-control platforms generate datasets so massive that human review becomes a bottleneck in itself. AI triage and classification—ideally physics-aware enough to distinguish real anomalies from noise—start to look essential rather than nice-to-have. BCG's "Factory of the Future" report this May, surveying 1,000 manufacturers, found that AI-enabled process control and digital twins are reshaping manufacturing economics across industries, semiconductors very much included.
Third, the industry is inching toward reusable physics intelligence. Traditional finite-element or CFD models can take days or weeks to solve. A task-specific neural surrogate might speed that up, but it requires retraining for each new geometry or material. PhysicsX claims that Large Physics Models, pre-trained on vast corpora of simulation data and augmented by real-world measurements, can generalize across design variants and slot directly into engineering workflows. The company claims acceleration of 10⁴ to 10⁶ times versus conventional simulation in certain scenarios.
That's an extraordinary claim. And as of early July, PhysicsX hasn't published named semiconductor case studies with quantitative KPIs to back it up.
The Missing Reference Customer
PhysicsX has disclosed case studies in foundry casting quality and brake cooling optimization. It announced a partnership with the America's Cup sailing team in March. A collaboration with Siemens on data-center power infrastructure modeling—using physics-AI surrogates for busway thermal analysis—illustrates how the platform can plug into established simulation environments.
The company lists semiconductors as a deployment sector in its Series C announcement and throughout releases over the past year or so. But no public, metrics-backed fab or lithography case study appears on its newsroom as of this writing. Strategic investors like Applied Materials and NVIDIA suggest validation potential within the chip-making stack, certainly. Yet the absence of a named reference customer in semiconductors raises questions about whether we're looking at production readiness or something closer to proof-of-concept.
NVIDIA's own ecosystem, by contrast, offers a clearer picture of physics AI in semiconductor workflows. Beyond cuLitho, the company announced at GTC an expansion of its open models for "physical AI"—the Cosmos family—and PhysicsNeMo, a framework for training physics-informed neural operators. A PhysicsNeMo developer blog post on June 11 introduced a diffusion module focused on physics-AI applications. PhysicsX announced collaboration with NVIDIA on open standards for physics AI (the Opora framework) this past March, positioning itself within this broader NVIDIA-anchored ecosystem. Meanwhile, Alsemy, another startup, has a published NVIDIA case study deploying physics-informed AI for chip modeling on PhysicsNeMo in partnership with Korea's National Nano Fab Center.
The equipment and EDA majors, for their part, are further along in production deployment. TSMC's use of cuLitho is well documented. Samsung presented OPC acceleration results at GTC last year. Applied Materials' AIx platform, launched back in 2021, harnesses big data and AI for PPACt optimization—power, performance, area, cost, time-to-market—across its process tools. Lam's Sense.i etch systems ship with in-situ sensing and analytics baked in.
These aren't foundation models in the PhysicsX sense; they're more narrowly scoped, vendor-locked solutions. But they demonstrate that the chip industry is already betting real money on AI to manage complexity at scale.
What Comes Next
Market dynamics certainly favor continued investment in physics AI for semiconductors. SEMI's April forecast of $151 billion in 300mm fab equipment spending by next year, Omdia's upward revision of semiconductor growth to 62.7 percent for this year (driven by HBM and AI), and Gartner's projection of 55.8 percent growth in datacenter systems all point to sustained demand for EDA acceleration, computational lithography, and fab analytics. McKinsey estimates the semiconductor industry could reach $1.6 trillion by 2030, with AI datacenters as the primary engine. Equipment makers are expanding capacity—Applied Materials announced a Singapore manufacturing expansion this June to support AI chip demand—and that expansion pulls through software and AI infrastructure whether or not it comes from a startup.
Regulatory headwinds are real, if navigable. The EU AI Act entered force in August 2024, with general-purpose AI model obligations fully applying by August 2 of this year. Foundation models, including physics models, face transparency, risk assessment, and technical documentation requirements. PhysicsX, as a UK-based company with European operations (including the European Industrial AI Cloud deployment announced in February), will need compliance pipelines. Export controls on advanced lithography equipment—Dutch rules expanded in September 2024 require licenses for ASML's advanced DUV, and EUV has been restricted since 2019—don't directly target software, but geopolitical tensions shape ROI calculations for nodes and tools. Indirectly, that affects which simulation and AI platforms gain traction.
The strategic question, though, is whether specialized physics foundation models—Large Physics Models—will emerge as a distinct category alongside large language models, or whether the incumbents simply absorb the technology into their platforms. Cadence's ChipStack and Siemens' Solido agentic AI suggest the latter path may already be underway. PhysicsX's investor base—Applied Materials and Siemens on the cap table, NVIDIA as both partner and investor, Temasek writing a $300 million check—implies a hybrid outcome: startups provide the pre-trained foundation models and frameworks, established vendors integrate them into production workflows.
For founders eyeing specialized AI for scientific and manufacturing applications, the semiconductor opportunity is sizable but unforgiving. This industry values proven ROI over architectural elegance. A 15× speedup in lithography OPC—cuLitho's demonstrated gain—translates directly to mask-shop economics and time-to-market. A generalized physics model that needs case-by-case tuning faces skepticism unless it can show cross-domain transfer learning with minimal retraining.
PhysicsX's lack of a public semiconductor reference customer as of mid-2026, despite listing the sector prominently and securing strategic investors from the equipment stack, suggests that bridging the gap from research-grade acceleration to fab-qualified deployment is not a trivial exercise. Perhaps that's why the company raised $300 million.
The semiconductor equipment spending supercycle, projected by SEMI to peak next year, creates urgency. High-NA EUV adoption remains staggered—Intel leading, TSMC delaying on cost grounds—which keeps multi-patterning and computational complexity front and center. Inspection data volumes from advanced packaging and chiplets will only grow. If Large Physics Models can deliver on the promise of 10⁴ to 10⁶× acceleration with production-grade accuracy, the market is undeniably there.
Whether PhysicsX captures it, or whether NVIDIA, Synopsys, and Cadence subsume the category into their own platforms, remains an open bet. But at a $2.4 billion valuation and $300 million in fresh capital, the company has bought itself enough runway to find out. The hard part—proving it works where it counts—is what comes next.
