The AI buildout is creating unprecedented demand for compute, memory, and the hardware needed to support it. As that demand grows, the ability to design and quickly validate hardware is becoming an important bottleneck. We believe physics is that layer needed to close that gap, making it the next foundation model frontier.
That’s why we’re thrilled to invest in Vinci‘s $250M Series B, alongside Advent and Temasek and other industry leaders like Khosla, Eclipse, and AMD.
Vinci is building a foundation model for physics, akin to an LLM for the physical world. Their vision is to become the intelligence layer through which hardware is validated, optimized, and eventually designed, enabling faster, cheaper design cycles and earlier iteration with fewer costly mistakes.
The physics bottleneck in hardware
Everything we interact with in the world has physics at its foundation. If we can have better physics models, the results flow downstream and upstream into everything – engineering, product development, and our daily lives. These new models can increase innovation, creativity, efficiency, speed, quality, safety, and lower costs.
More specifically, in hardware, physics is critical for every step of the design process. Before a chip, battery, or other physical product can be manufactured, engineers must first design it and then understand how it will behave in the real world. A mistake here can be enormously timely and expensive, yet the process used for design and simulation still runs the way it did twenty years ago.
Legacy tools use a separate solver for each physics domain, relying on specialized simulation tools and expert engineers, with hours of setup before a simulation can run. Engineers prepare the geometry, configure the simulation, wait for the result, and interpret it before making the next design decision. Simulation is therefore an episodic part of the engineering workflow rather than something engineers can continuously use as they explore a design.
Newer AI approaches have addressed the problem by building custom models for individual customers or use cases. However, those models do not scale easily, generalize, or improve across deployments. Karan’s own years working in chip design and go-to-market inside the semiconductor industry gave us early conviction in just how real this bottleneck is.
One general model, built to compound
Physics is universal. A good physics result is a useful ground truth, regardless of who performed the measurement – and it doesn’t matter where and when it was done. Physics is physics – and it is quite literally the ground truth for everything around us.
Therefore, Vinci’s bet, and ours, is that a single general model can replace both legacy point solutions and newer custom models. The same model weights transfer to new designs and new use cases, and therefore the model can generalize across geometries, industries, and physics domains rather than being rebuilt for each one.
The model has three key aspects: a machine learning model, analogous to LLMs for words and text; physics equations that act as truth and guardrails; and physics solvers such as FEA and Monte Carlo models that serve as spot checks and quality control. The architecture pairs a field-based physics foundation model with a proprietary GPU numerical solver. The model predicts the physical solution, which is then checked against governing fundamental equations. It is trained on proprietary data and operates entirely inside each customer’s environment, so customers can protect their data advantages.
The single model also enables the model to run zero-shot, without fine-tuning or customer-specific data and retraining. That makes simulation dramatically faster and cheaper. Engineers can simulate and iterate earlier in the design process, catch costly mistakes before production, and accelerate the design cycles that gate new hardware development. It also means simulation can move from an occasional, late-stage check into something engineers can do constantly and early, catching costly mistakes long before they reach production and enabling first-time-right silicon.
The moat should deepen as Vinci adds more proprietary data, physics capabilities, and embeds itself into more workflows. Every new physics domain the model learns, and every new deployment, feeds the same underlying architecture, building an advantage that’s difficult to replicate with a narrower point tool.
The roadmap runs from simulation today toward a broader physics intelligence layer that reasons across thermal, mechanical, and electrical behavior on one platform, and eventually toward tools that help engineers edit and optimize designs directly, with the model proposing and validating changes alongside them.
A team built for the problem
Building a foundation model for physics requires deep technical expertise alongside an understanding of how engineering software gets built and deployed. Vinci has assembled that combination. Hardik Kabaria, Vinci’s founder and CEO, holds a Stanford PhD in mechanics and computation and spent eight years in industry, most recently as SVP of Software Engineering at Carbon. He has assembled a world class technical team, complementing himself with experts in machine learning, robotics, and semiconductors, and close to half the team holds PhDs.
The future of hardware engineering requires dramatic acceleration of the design and verification processes while pushing the outer edges of capabilities to meet the insatiable need for intelligence powered by AI. Vinci is ushering in a new era of hardware design and verification by removing friction from the process while still being grounded in the immutable laws of physics.
We’re thrilled to partner with Hardik and the entire Vinci team as they power this new foundation model frontier and radically transform how every physical product gets validated and eventually designed. Let’s go!
Related Insights
