The Pulse
Flexcompute Wins Samsung and Applied Ventures Backing for Physics AI
Flexcompute announced strategic investment from Samsung Ventures, Applied Ventures and other investors to expand its Physics World Model. The company says its system solved a 5-billion-degree-of-freedom semiconductor warpage model within 2%
AI.info Team · Written with AI ·
“We are building the verification layer between bits and atoms: a Physics World Model any engineer or AI agent can ask whether a design will work, before it is built.”
Vera Yang, co-founder and president of Flexcompute
Flexcompute announced on October 12 that Samsung Ventures and Applied Ventures are among the investors backing its effort to build a Physics World Model, software intended to predict how engineered designs behave under real-world physical conditions. The company did not disclose the investment amount or terms, and said other investors also participated.
The Boston company plans to use the backing to expand its research team, increase GPU computing capacity and extend its models across fluids, electromagnetics, structures, heat and interactions among those fields. Its pitch is that AI predictions for engineering should be constrained by physical laws, rather than relying only on patterns in training data.
Flexcompute’s 5-billion-degree-of-freedom test
Flexcompute cited a semiconductor-package analysis as evidence of the approach. The company says it modeled warpage in a package containing more than 800,000 solder joints, an application-specific integrated circuit and 32 optical engines. It reports that the calculation involved 5 billion degrees of freedom and produced results within 2% of commercial finite-element results in minutes.
The comparison and timing are Flexcompute’s claims; the announcement does not name the commercial software used or provide an independent assessment of the result. Still, the example points to the scale of engineering problems the company wants its system to address: simulations involving many interacting parts, where design teams need to estimate how heat, materials and structure affect one another.
“Every industry is racing to put AI to work on real engineering problems, and the limit is how quickly each one can be validated against the laws of physics,” said Yu Cao, chief AI officer of Applied Materials. Cao said the technology could allow more of advanced-manufacturing design work to happen digitally, reducing time and cost.
From GPU solvers to “Physics Twins”
Flexcompute says its team includes more than 70 PhDs and has built GPU-based solvers for computational fluid dynamics, electromagnetics, structural analysis and thermal modeling. Its approach, which it calls Tokenized Physics, combines neural networks with first-principles physics in the model. The company says that design is meant to keep predictions consistent with conservation laws and useful for designs the model has not encountered before.
Customers currently use the system through what Flexcompute calls Physics Twins: models of individual products that engineers and AI agents can question, test and optimize before construction. In the workflow described by the company, engineers set objectives and approve results, while agents run iterations under physical constraints. Flexcompute says engineering teams at more than 250 companies across aerospace, automotive, semiconductors, photonics and advanced manufacturing use its products.
Strategic investors, undisclosed terms
The announcement gives Samsung Ventures and Applied Ventures strategic roles but does not specify investment size, ownership stakes or other financial terms. Samsung Ventures is Samsung’s corporate venture-capital arm; Applied Ventures operates as the venture-capital arm of Applied Materials, a semiconductor-equipment maker. Their participation links Flexcompute’s research ambitions to two companies with interests in hardware and manufacturing, but the release does not detail any product, customer or commercial commitments attached to the investment.
Flexcompute’s immediate plan is to expand the physics domains covered by its model and the computing capacity behind it. Its central test will be whether the reported speed and accuracy hold across the varied engineering problems that customers bring to their Physics Twins—not just the semiconductor package example highlighted in the announcement.