Nicolas Sauvage is betting on the boring part of AI

Nicolas Sauvage believes it takes four years for the best bet to be considered wise. I’m thinking about when I said this on stage last week at StrictlyVC’s San Francisco event, co-hosted by TDK Ventures.

It’s a theory he’s been trying to prove since 2019, when he founded the Japanese electronics giant’s corporate venture arm, which now manages $500 million across four funds. Groq, an AI chip startup valued at $6.9 billion in its latest funding round last fall, is the most high-profile example of this idea.

In 2020, long before the generative AI boom made infrastructure the go-to place for direct funding, Sauvage wrote a check to the company founded by Jonathan Ross (one of the engineers who built Google’s tensor processing unit). From the beginning, Groq focused on inference, the computational heavy lifting that occurs every time a model responds to a query. Ross first built a compiler and then designed the chip by stripping away the architecture to the point where, as Sauvage puts it, “you could remove parts and still make it work.”

It may have seemed niche to some, but Sauvage, knowing what he was doing against the constraints of his parent company, saw an opportunity. Unlike consumer hardware, which has a natural upper limit, the demands on inference continue to grow in complexity as new applications and new models emerge. At the time, Sauvage couldn’t have predicted that the demand for inference would explode this year, thanks to all the AI ​​agents planning and executing dozens of calls (previously one query was enough).

But in a way, Ross also got lucky. After all, the Japanese electronics conglomerate best known for magnetic tape isn’t the most natural investment partner on the surface. In fact, Sauvage says the existence of TDK Ventures itself is highly unlikely. But after giving two lectures in a row at Stanford University, once making the case for corporate VC and once listing all the reasons for its failure, Mr. Sauvage, a Frenchman who had joined TDK in Silicon Valley through an acquisition, pitched the idea to upper management at TDK headquarters, despite having no clear position. (“I’m not Japanese. I don’t speak Japanese. I don’t even live in Tokyo,” he told this editor.)

After refusing to take no for an answer, he finally received the green light to establish the fund. Its mission is to answer one question: What’s the next big thing for TDK, and what could kill it?

The portfolio he has since assembled is dotted with technologies that became of broader interest to VCs last year, including solid-state grid transformers, sodium-ion batteries for data centers, and alternative battery chemistries that circumvent the geopolitical vulnerabilities of lithium and cobalt.

The discipline behind everything is the same. Identify the bottleneck after 4 years and find founders who are already working on it.

The question, of course, is what happens next. Sauvage is looking closely at physical AI. It’s not all about robotics, but robots that have very specific jobs to perform. For example, Agility Robotics in his portfolio focuses on the single, mundane task of moving things from one place to another in warehouses facing labor shortages. Another portfolio company, Swiss portfolio ANYbotics, builds rugged robots for environments that are too dangerous for human workers, places where the definition of work inherently prevents people from going. Consistency is clarity of purpose. The robots Sauvage is betting won’t try to do everything. Instead, they make sure to do one difficult thing.

Sauvage said he’s also watching for another shift in the computing stack. Training (massive parallel computation to teach models) has been dominated by GPUs. Inference chips like Groq are reshaping what happens when their models speak faster, cheaper, and at scale. Sauvage argues that CPUs are undergoing a renaissance. These are not the most powerful or fastest chips. However, they are the most flexible and best suited for branching decision logic in orchestrations. As an AI agent delegates tasks, checks its progress, and loops back through dozens of steps, something must manage the entire choreography. That something looks more and more like a CPU.

And then there’s China. A recent report from Eclipse, a venture he closely follows, documents what Sauvage describes as “vibe manufacturing,” the use of AI to rapidly iterate on prototyping physical hardware, mirroring what vibe coding has done for software. According to the report, Chinese manufacturers are shortening the design, manufacturing and testing cycles for physical products in ways that Western supply chains are not yet equipped to match.

This is a signal for Sauvage, who is working on various investments for TDK Ventures. One issue that remains unresolved, he says, is dexterity. Models are improving at such a rate that physics AI feels inevitable. What is still missing is a comparable physical fluency. Countries and companies that figure out how to iterate atoms as fast as other countries and companies iterate code will have a manufacturing advantage. That’s the wave in which he positions TDK Ventures today.