Nvidia’s new AI physics model can help design chips and a whole lot more

Nvidia hopes that its new open-source AI model for physics, Apollo, will find application in a wide variety of high-tech scientific and industrial fields.

It unveiled the new model family at SC25, the International Conference for High Performance Computing, Networking, Storage, and Analysis, just a month after unveiling four others: Nemotron for agentic AI, Clara for biomedical AI, Isaac GR00T for robotics, and Cosmos for other physical AI applications.

Nvidia said that the Apollo family of models will allow developers to integrate real time capabilities into their simulation software in areas such as defect detection, computational lithography, and electrothermal and mechanic design for electronic devices and semiconductors, structural analysis, weather forecasting and simulation, computational fluid dynamics, electromagnetics, and simulation in nuclear fusion, plasma simulation, and fluid structure interaction.

Apollo will provide pretrained checkpoints and reference workflows for training, inference and benchmarking, allowing developers to customize them for their applications. It is, said Nvidia, “coming soon,” and will be available on HuggingFace, build.nvidia.com, and as Nvidia NIM microservices.

Sanchit Vir Gogia, CEO of Greyhound Research, said that Apollo stands out as the intellectual centerpiece of SC25. “Nvidia has turned AI-driven physics into a fully industrialized model family spanning semiconductors, structural mechanics, materials science, weather, climate, automotive aerodynamics, and more. These are not research curiosities. When tsunami forecasting models run billions of times faster, or when petabytes of materials data are folded into real-time inferences, the scientific method itself shifts. Apollo ensures that this shift occurs inside Nvidia’s ecosystem. Once engineers, climate researchers, and materials scientists base their workflows on these models, the surrounding software, hardware, and infrastructure decisions become inevitably Nvidia-aligned. This is the most powerful form of lock-in: dependency created through genuine breakthrough performance.”

Yet more Nvidia supercomputers

Some of those models could perhaps be put to use in new supercomputers being built with Nvidia chips.

Japanese research institute RIKEN is building two of them, one providing AI for scientific research, and the second dedicated to research in quantum algorithms, hybrid simulation and quantum-classical computing methods. Both use the GB200 NVL4 platform and are interconnected by NVIDIA Quantum-X800 InfiniBand networking.

Dion Harris, senior director, HPC and AI infrastructure solutions at Nvidia, said the second system will integrate GPUs directly into RIKEN’s quantum HBC hybrid infrastructure, linking quantum computers with accelerated computing systems and classical supercomputers like Fugaku.

In the US, Dell and the Texas Advanced Computing Center are announcing the 300 petaflop Horizon supercomputer, which will, Nvidia said, be “America’s largest academic supercomputer.“ Due to come online in 2026, it will contain 4,000 Nvidia GB200 GPUs and 9,500 Nvidia Vera CPUs.

Lock-step launches lead to lock-in

However, Gogia expressed concerns over the plethora of new Nvidia-based supercomputers — more than 80 announced this year alone. “This is not market success; it is architectural dependence,” he said. “National science agencies are aligning their multi-year roadmaps with Nvidia’s cadence, effectively transitioning from vendor selection to vendor reliance.”

Overall, he was impressed with Nvidia’s featured technologies, although he worries about the future as its dominance continues to increase.

“The breakthroughs showcased at SC25 are extraordinary,” Gogia said, but, “They come with a governance cost. When the entire lifecycle of scientific computation, spanning simulation, AI, data movement, networking, storage, orchestration, and quantum control, becomes anchored to a single vendor’s architecture, autonomy diminishes. CIOs, national labs, and research agencies must now decide whether they are comfortable with a future where the acceleration of science is extraordinary, but the ecosystem shaping it is extraordinarily narrow. Nvidia has offered the world a path to unprecedented capability. It is up to the world to decide whether that path should also be the only one.”