Nicolas Sauvage has a theory: the best bets take four years to look obvious. It is a principle he shared last week at StrictlyVC’s San Francisco event, co-hosted by TDK Ventures, and one he has been testing since 2019, when he founded the corporate venture arm of Japanese electronics giant TDK. The fund now manages $500 million across four funds, with AI chip startup Groq, valued at $6.9 billion in its most recent round, as its highest-profile holding.
Sauvage wrote his check into Groq in 2020, well before generative AI made infrastructure bets fashionable. The company, founded by Jonathan Ross, one of the engineers behind Google’s Tensor Processing Units, was focused from the start on inference: the computation that happens every time an AI model responds to a query. Ross designed the chip by building the compiler first, stripping the architecture down to its minimum viable form. Sauvage saw asymmetry where others saw a niche play. Unlike consumer hardware, demand for inference compounds with every new application and model, and the explosion of AI agents has since validated that read.
The portfolio Sauvage has assembled around TDK Ventures follows the same discipline throughout: identify the bottleneck four years out, then back the founders already working on it. Current holdings include companies working on solid-state grid transformers, sodium-ion batteries for data centers, and alternative battery chemistries that sidestep the supply chain fragility of lithium and cobalt. In physical AI, Sauvage is watching robots with highly specific jobs rather than general-purpose machines. Agility Robotics, a portfolio company, focuses on the single task of moving objects in warehouses facing workforce shortages. Swiss firm ANYbotics builds ruggedized robots for environments too hazardous for human workers. The through-line is clarity of purpose.
Sauvage is also tracking what he sees as the next shift in the compute stack. GPUs dominated the training era. Inference chips like Groq’s are reshaping what happens when a model responds at scale. His next call is that CPUs are due for a renaissance, not because they are the most powerful or fastest chips, but because they are the most flexible and best suited to the branching, decision-making logic that AI agent orchestration requires. When an agent delegates a task, monitors its progress, and loops back across dozens of steps, something has to manage the choreography. Sauvage believes that something increasingly looks like a CPU.
On manufacturing, Sauvage is closely watching a recent Eclipse Ventures report documenting what he describes as “vibe manufacturing” in China: the rapid, AI-assisted iteration of physical hardware prototyping that mirrors what vibe coding did for software. Chinese manufacturers, the report found, are compressing the design-build-test cycle in ways Western supply chains are not yet equipped to match. Sauvage sees it as a bottleneck signal. The unsolved problem he is watching most closely is dexterity: AI models are improving fast enough that physical AI feels inevitable, but the physical fluency to match that intelligence is still missing. The companies and countries that learn to iterate on atoms as fast as others iterate on code, he argues, will hold the next manufacturing advantage.
