Nicolas Sauvage has a theory that most investors don’t want to hear: the best bets take four years to look obvious. It’s a thesis he’s been testing since 2019, when he founded TDK Ventures, the corporate venture arm of a Japanese electronics conglomerate most people associate with magnetic tape. Today, the fund manages $500 million across four funds, and its portfolio is studded with the kind of infrastructure bets that suddenly feel inevitable in hindsight.
Take Groq, the AI chip startup currently valued at $6.9 billion. In 2020, when Sauvage cut a check into the company, generative AI wasn’t yet a household obsession. The infrastructure boom hadn’t started. But Groq’s founder, Jonathan Ross (who engineered Google’s Tensor Processing Units), was focused on a specific problem: inference, the computational work that happens every time an AI model answers a query.
Ross didn’t just build a chip. He built the compiler first, then stripped the architecture down until, as Sauvage describes it, “you can’t remove one part and have it still work.” It looked niche. It was niche. But Sauvage saw something else: asymmetry. Unlike consumer hardware with natural ceilings, demand for inference compounds with every new application, every new model, every new AI agent that needs to think and act across dozens of steps.
The gamble paid off. Sauvage couldn’t predict that inference demand would explode this year. But he positioned TDK to catch it anyway.
Why a Japanese Tape Company Funds AI Chips
Here’s the thing that makes this story worth paying attention to: TDK Ventures shouldn’t exist. A Japanese electronics conglomerate best known for consumer electronics and magnetic tape has no obvious reason to build a venture fund in Silicon Valley. Yet Sauvage, a Frenchman with no Japanese language skills and no standing at TDK headquarters, pitched the idea after delivering two Stanford lectures on corporate VC. The first made the case for it. The second catalogued every reason it fails.
He refused to take no for an answer.
When Sauvage finally got the green light, TDK gave him a mandate framed as a question: What’s the next big thing for us, and what might kill us? That constraint turned into discipline. The portfolio now reflects it: solid-state grid transformers, sodium-ion batteries for data centers, alternative battery chemistries that sidestep lithium and cobalt supply chains, and increasingly, physical AI applications that can do one hard thing reliably.
The through-line across all of it is the same. Identify the bottleneck four years out. Find the founders already working on it. Invest.
Physical AI Doesn’t Try to Do Everything
Sauvage is watching robotics closely now, but with a specific lens. He’s not betting on humanoids or general-purpose machines. He’s betting on robots with a single, clearly defined job.
Agility Robotics, in his portfolio, moves things from one place to another in warehouses. ANYbotics, a Swiss company he’s backed, builds ruggedized robots for hazardous environments where humans can’t safely go. These aren’t robots trying to do everything. They’re solving one problem with precision and reliability. There’s clarity of purpose baked into the bet.
That clarity matters because it reduces the scope of the problem. A robot that moves boxes doesn’t need to understand philosophy or navigate social nuance. It needs to move boxes reliably, day after day, in conditions that make human labor impractical.
The Compute Stack is Shifting Again
Sauvage tracks an interesting pattern in the technology stack. GPUs dominated training (the massive parallel computation of teaching models). Inference chips like Groq’s are now reshaping what happens when models respond. But the next layer is already shifting, he argues: CPUs are due for a renaissance.
GPUs are powerful and fast. CPUs aren’t the most powerful or fastest. But they’re flexible, and flexibility matters when you’re orchestrating. When an AI agent delegates tasks, checks progress, and loops back across dozens of decisions, something has to manage the choreography. That something increasingly looks like a CPU handling the branching logic and decision-making that training and inference don’t address.
It’s not the flashiest bet. No one writes venture checks because CPUs are having a moment. But that’s partly the point. The obvious plays get crowded. The asymmetric ones don’t.
Manufacturing is About to Speed Up
Sauvage has been reading research from Eclipse, a venture firm focused on the intersection of AI and hardware. They documented what he calls “vibe manufacturing”: rapid, AI-assisted iteration of physical prototypes. Chinese manufacturers are compressing the design-build-test cycle in ways Western supply chains aren’t equipped to match.
It’s a bottleneck signal. And Sauvage is already moving on it through TDK’s various bets.
The remaining unsolved problem, he says, is dexterity. Physical AI models are improving fast enough that the technology feels inevitable. What’s still missing is physical fluency. The countries and companies that figure out how to iterate on atoms as fast as others iterate on code will have a structural manufacturing advantage.
That’s the wave Sauvage is positioning TDK Ventures to catch today. Whether he’s right about the four-year horizon is something we’ll only know in hindsight, which is precisely why the bet works.
