NVIDIA says artificial intelligence is already changing the way chip fabs work, with simulations up to 35 times faster at Applied Materials, up to 17 times faster in Siemens’ computational lithography, and plasma predictions that go from days to milliseconds at Lam Research. That’s according to Timothy Costa, vice president and general manager of industrial and computational engineering at NVIDIA, in a blog post that rounds up results from partners and groups them under one concept: the AI-defined fab.
AI in chip fabs in 20 seconds
- NVIDIA gathers case studies from COMSOL, Applied Materials, Siemens, Lam Research and Advantest.
- The gains range from 10 to 35 times in simulation, always stated as maximums.
- All figures come from NVIDIA or its partners, with no independent verification.
The starting point is well known in the industry. Demand for advanced chips keeps growing, processes keep getting more complex, and testing a change directly on the production line is slow and expensive. Working out how a material will behave in a new step, or finding the cause of a defect, can take days of analysis, and a wrong call about wafer yield has a real cost. NVIDIA’s pitch combines four pieces: accelerated computing to simulate faster, physics-trained AI models that predict without re-running a simulation every time, agents that chain together engineering tasks, and digital twins to test changes in a virtual environment. It’s part of a broader shift toward AI-driven, more autonomous production floors — cloudnews.tech has covered how increasingly automated “dark fabs” raise their own zero-trust security questions.
Shorter simulations to test more alternatives
The first piece is the most straightforward. Multiphysics simulations, which model effects like heat, electrical behavior and mechanical stress all at once, keep getting heavier. According to NVIDIA, COMSOL speeds up this kind of semiconductor manufacturing simulation by up to 10 times with cuDSS, its library for solving sparse systems of equations on GPUs. The point isn’t just saving time, but being able to compare more options before deciding which ones to test physically.
Applied Materials, the largest supplier of semiconductor manufacturing equipment, contributes two figures. Its Ginestra platform, which simulates materials at the atomic scale, runs up to 10 times faster than CPU-only with cuDSS, and its ACE+ multiphysics platform runs topography simulations up to 35 times faster with PhysicsNeMo. Applied itself detailed this in February 2026, explaining that the speedup turns week-long studies into same-day results. The company is already working with other partners on similar ground, such as its collaboration with Intel on advanced interconnects and 3D stacking.
In some cases, the speedup is what makes the calculation possible at all. That’s true of computational lithography, which defines how circuit patterns are printed onto silicon and whose calculations are too intensive to complete on CPUs at production scale. Siemens integrates NVIDIA’s cuLitho library into its Calibre software. NVIDIA’s blog post cites up to 17 times faster performance on key simulation workloads; Siemens’ October 15 announcement puts it more precisely as a range of at least 10 to 17 times, and ties it to making curvilinear inverse lithography technology (ILT) viable at scale — a technique that is gaining ground with processes like the ones TSMC and ASML are preparing for High-NA EUV lithography.
From plasma in milliseconds to testing agents
The second piece consists of training models on physical laws so they can predict how a process will behave without repeating the full simulation. The most striking example is plasma, the electrically charged gas used to etch wafers. Modeling it requires solving equations that combine electrical, chemical and thermal effects, which makes every simulation costly. According to NVIDIA, Lam Research showed that its PhysicsNeMo framework can predict that behavior in milliseconds, compared with the days a full simulation takes. It’s a demonstration, not a production deployment described in detail.
The third piece is AI agents, which handle the intermediate steps between engineering decisions. The case NVIDIA cites is Advantest, the Japanese maker of chip-testing equipment. Deciding which AI model to apply to each task on the test floor used to be a sequential, manual process that could take hours, and Advantest automates it with agents built on NeMo Switchyard, NVIDIA’s tool for routing requests between models. Response time, according to both companies, goes from hours to minutes, with no more precise figures given. It’s part of a broader push to put AI agents to work on chip-related tasks: cloudnews.tech has reported on how Empyrean Technology used agentic AI to cut a circuit-design task’s time by 75% in China.
The fourth piece is digital twins. Applied Materials uses NVIDIA’s Omniverse libraries to build physically accurate replicas of entire fabs, where it simulates the movement of materials, spots bottlenecks, and evaluates equipment layout and operating strategies before putting them into practice. Neither company has published quantitative results for this part.
What’s behind the numbers
NVIDIA’s article works as a round-up of earlier announcements and doesn’t include dates or test details. Applied’s figures come from its February blog post; Siemens’ come from its October announcement, and the rest come from the note NVIDIA published on October 14 at SEMICON West. Comparing the sources brings out some nuances. In that note, NVIDIA described Applied’s improvement as “up to 35 times faster” in chamber simulations, while Applied and the blog attribute it to topography simulations with ACE+. All the figures are maximums (“up to”), and in several cases it isn’t stated what they are being compared against.
The SEMICON West note also included other partners the blog doesn’t mention: Teradyne, with probe-maintenance inference about 85 times faster; KLA, with an algorithm 50 times faster than its CPU version; Synopsys, with potential improvements of up to 30 times in QuantumATK; and Cadence, with an agent that optimizes advanced packaging eight times faster. NVIDIA also introduced cuPotential at the time, a new library for atomic-scale simulation.
The end goal NVIDIA describes is to connect all these pieces: use production data to feed the models, use the models to evaluate changes, and feed the conclusions back to the engineers who make the decisions. In NVIDIA’s framing, engineers still steer the work and make the calls, but they wait less for answers. It’s also a business strategy: the more manufacturing steps depend on its CUDA-X libraries, the further NVIDIA’s platform extends beyond the data center.
Sources:
- NVIDIA Blog: AI Is Changing How Semiconductor Fabs Work, by Timothy Costa
- NVIDIA: NVIDIA and Semiconductor Leaders Advance AI Across Chip Design and Manufacturing for AI-Defined Fabs (October 14, 2026)
- Applied Materials: collaboration with NVIDIA to accelerate chip manufacturing (February 18, 2026)
- I-Connect007: Siemens Advances AI Chip Lithography With NVIDIA cuLitho (October 15, 2026)
This article is also available in Spanish: «Simulación, IA y agentes: así quiere NVIDIA acelerar las fábricas de chips», on revistacloud.com.
