Nicolas Sauvage is betting on the boring parts of AI

Nicolas Sauvage believes it takes 4 years for the most effective bets to look apparent — considering that he shared on stage final week at StrictlyVC’s San Francisco occasion, which TDK Ventures co-hosted.

It’s a concept he’s been working to show since 2019, when he based the company enterprise arm of the Japanese electronics big, which is now managing $500 million throughout 4 funds. The AI chip startup Groq, valued at $6.9 billion throughout its most up-to-date funding spherical final fall, is the highest-profile instance of this considering.

In 2020, nicely earlier than the generative AI increase made infrastructure bets look apparent, Sauvage wrote a examine into the corporate, which was based by Jonathan Ross — one of many engineers who constructed Google’s Tensor Processing Models. Groq was centered from the beginning on inference: the computational heavy lifting that occurs each time a mannequin responds to a question. Ross had designed his chip by constructing the compiler first, stripping the structure down till, as Sauvage describes it, “you may’t take away one half and have it nonetheless work.”

It may need seemed area of interest to some, however realizing what he did about his guardian firm’s constraints, Sauvage noticed asymmetry. In contrast to shopper {hardware}, which has a pure ceiling, demand for inference retains compounding with each new software and each new mannequin. Sauvage couldn’t know then that demand for inference would explode this 12 months, thanks to each AI agent that plans and acts throughout dozens of calls (the place a single question used to suffice).

However in some methods, Ross obtained fortunate, too. In any case, a Japanese electronics conglomerate finest identified for magnetic tape isn’t, on its face, the obvious investing associate. The truth is, Sauvage describes TDK Ventures’ personal existence as impossible. However after two back-to-back Stanford lectures — one making the case for company VC, one cataloguing each purpose it fails — Sauvage, who’s French and joined TDK in Silicon Valley by way of an acquisition, pitched the thought to higher-ups at TDK headquarters regardless of having no apparent standing to take action. (“I’m not Japanese. I don’t converse Japanese; I don’t reside in Tokyo,” he instructed this editor.)

After refusing to take no for a solution, he lastly obtained the inexperienced gentle in to construct a fund whose mandate was to reply one query: What’s the following large factor for TDK, and what may kill it?

The portfolio he has since assembled is dotted with applied sciences which have change into extra broadly fascinating to VCs over the past 12 months: solid-state grid transformers, sodium-ion batteries for information facilities, different battery chemistries that sidestep the geopolitical fragility of lithium and cobalt.

The self-discipline behind all of it’s the similar: establish the bottleneck 4 years out, then discover the founders already engaged on it.

The query, after all, is what’s subsequent. For his half, Sauvage is watching bodily AI intently — not all of robotics however robots with a extremely particular job to be carried out. Agility Robotics, for instance, in his portfolio, focuses on the only, mundane process of transferring issues from one place to a different in warehouses dealing with workforce shortages. One other portfolio firm, Swiss portfolio ANYbotics, builds ruggedized robots for environments too hazardous for human employees — locations the place the job definition is basically to go the place folks can’t. The through-line is readability of function. The robots Sauvage is betting on don’t attempt to do every thing; as an alternative, they do one onerous factor reliably.

Sauvage says he’s additionally watching the compute stack shift once more. GPUs dominated coaching — the huge, parallel computation of educating a mannequin. Inference chips like Groq’s are reshaping what occurs when that mannequin speaks: sooner, cheaper, at scale. Now, Sauvage argues, CPUs are due for a renaissance. They’re not essentially the most highly effective chips or the quickest. However they’re essentially the most versatile and finest suited to the branching, decision-making logic of orchestration. When an AI agent delegates a process, checks on its progress, and loops again throughout dozens of steps, one thing has to handle the entire choreography. That one thing, more and more, appears like a CPU.

After which there’s China. A current report from Eclipse — a enterprise agency he follows intently — documented what Sauvage describes as “vibe manufacturing” — the fast, AI-assisted iteration of bodily {hardware} prototyping, mirroring what vibe coding did for software program. Chinese language producers, the report discovered, are compressing the design-build-test cycle for bodily merchandise in methods Western provide chains aren’t but outfitted to match.

For Sauvage, it’s a bottleneck sign — and one he’s already transferring on with TDK Ventures’ numerous investments. One remaining unsolved drawback, he says, is dexterity. Fashions are bettering quick sufficient that bodily AI feels inevitable; what’s nonetheless lacking is the bodily fluency to match. The nations and corporations that work out methods to iterate on atoms as quick as others iterate on code could have a producing benefit. That’s the wave for which he’s positioning TDK Ventures at the moment.