Engineering simulations have been one of manufacturing's most stubborn bottlenecks. Running the complex mathematical models that predict how an aircraft wing will flex, how heat flows through an engine block or how a semiconductor will behave under load could take days, even weeks, of computation. Entire specialist teams were required simply to set up, run and interpret the results. 

That bottleneck now has a solution. NVIDIA is a California-based technology company best known for designing graphics processing units, the specialised chips originally built to render video game visuals at high speed. Those same chips, it turned out, were well suited to a much broader class of problems: any calculation that can be broken into thousands of parallel threads running simultaneously. Scientific computing, weather forecasting, drug discovery and artificial intelligence all benefit from this architecture.

The company has spent more than a decade building a software ecosystem around its hardware. That ecosystem, known as CUDA-X, is a collection of libraries and tools that allow engineers and developers to write programmes that run on NVIDIA chips efficiently. It is the foundation on which much of the world's accelerated computing now runs.

Leading the effort to connect this platform to the engineering and manufacturing industries is Tim Costa, Vice President and General Manager of Computational Engineering at NVIDIA. His role, as he describes it, is to connect accelerated computing and AI to engineering and design industries. His goal is to "cut engineering and simulation cycles from days to hours and put workflows that used to require hero teams into the hands of every designer and every engineer".

See the full story in the June 2026 edition of Manufacturing Digital.

From maths to manufacturing

Tim's background began in mathematics. His doctorate at Oregon State University focused on multiscale methods: techniques for modelling systems where behaviour at the microscopic level, such as the movement of atoms in a material, determines what happens at the engineering scale, like whether a component fractures under load. 

After graduating, he moved to Sandia National Laboratories, a US government research institution focused on science and engineering for national security, where he worked on simulation techniques for solid mechanics. From there, he spent years as an engineer building the mathematical libraries that form the numerical foundation for most of the world's high-performance and scientific computing.

He joined NVIDIA in 2019 to lead the development of the CUDA-X libraries. His remit now extends to quantum computing and the full span of computational engineering. Tim explains there has been a consistent thread throughout his career: "Transforming our ability to compute the fundamental mathematics that describes the physical world, so engineers and scientists can ask bigger questions."

Each chapter of his career, Tim explains, prepared him for a different layer of the challenge he now addresses. His doctorate grounded him in how mathematics drives engineering workflows, from the underlying physics through to the numerical methods engineers rely on in practice. His time as a library engineer taught him to translate from mathematics through computer architecture all the way to real industrial applications. Then, at NVIDIA, he says: "Leading CUDA-X put me at the intersection of every industrial sector and accelerated computing."

His current remit, which extends to quantum computing as well as the full span of computational engineering, has "closed that loop” giving Tim “the chance to support and collaborate with these ecosystems to transform the landscape of computing and revolutionise industrial design and engineering".

Reliability vs innovation

While Silicon Valley's culture prizes speed and novelty, manufacturing prizes repeatability and safety. Tim explains: "An AI method 10 times faster but 1% off from the ground truth is useless to a manufacturer." In software, a small error could cause a slightly strange movie recommendation. In manufacturing, this could mean an aircraft wing fails or a semiconductor shorts out. He feels that this reliability concern should be reframed rather than dismissed. 

Innovation, he argues, should not just meet the reliability bar, it should raise it. Greater computational speed creates room for higher-fidelity methods, simulating physics that was previously out of reach. That improved fidelity translates directly into better, safer products.

Tim explains: "The biggest lesson is that these are not actually opposed. Reliability is the bar innovation has to clear to be useful on an industrial floor." He believes the computational space that accelerated computing creates opens the door to more than just speed. "The performance and speed that accelerated computing and AI unlock create room for methods that have higher fidelity, simulating physics that was previously out of reach, and improve the actual reliability and capability of manufactured products," he says. "Innovation at its best does not undermine reliability. It compounds it."

Physical AI and the factory floor

Physical AI spans "the full arc, from accelerating the simulation and engineering of a new aircraft, vehicle, or chip to operating the line that builds it,” Tim says. “For a factory, that shows up in three concrete ways.”

The first is products arriving on the production line. Tim explains that those products "were themselves shaped by AI upstream, designed and validated against far more robust physics than was possible a few years ago." The simulation tools used during design, powered by CUDA-X libraries and AI physics models, produce better-engineered products before a single physical component is made.

The second impacts the factory itself. NVIDIA provides infrastructure for building digital twins through a platform called Omniverse, which is built around an open data format called OpenUSD. A digital twin is a detailed virtual replica of a physical facility, kept in synchronisation with the real asset using sensor data. Tim says this allows operators "to test a new line layout, a robot cell or a process change in the virtual factory before touching the real one". 

The third layer is operational AI systems that handle inspection, quality control and predictive maintenance. Tim describes vision models that "inspect every part on every shift", sensor models that "predict failures before they happen" and robots that "can be retrained for new tasks rather than rebuilt". He is candid about its state, though: "It is still early for some of this, but it is real, and it is running in production today at leading manufacturers around the world."

The hardest problems are human

If the tools exist and early production deployments are under way, why has adoption not moved faster? Tim's answer is that the limiting factor is rarely the AI model itself. He says: "It is the surrounding environment: fragmented data, disconnected IT and OT systems and workflows that are not linked across design, simulation, manufacturing and operations."

Tim explains that many companies "still lack a unified digital representation of their products, processes or factories, which makes simulation, validation and scaling much harder". Without that foundation, even capable tools are difficult to deploy consistently across multiple sites, whether the application is on the factory floor or upstream in engineering.

Many of the hardest remaining problems are, in his words, "human and workflow-driven”. These include setting up simulations, managing studies and interpreting results. The tools to run individual simulations faster exist. What has not yet arrived at scale is the layer of automation that connects those tools into a continuous workflow.

That is where Tim sees the next wave of transformation arriving. He expects autonomous engineering agents, AI systems capable of configuring and launching simulation runs and proposing subsequent steps, to take over more of that orchestration work. As Tim explains: "Every engineer will orchestrate a team of agent engineers and become limited only by their imagination."

See the full story in the June 2026 edition of Manufacturing Digital.