Nicolas Sauvage is betting on the boring parts of AI

Nicolas Sauvage believes it takes 4 years for the champion bets to look evident — reasoning that helium shared connected signifier past week astatine StrictlyVC’s San Francisco event, which TDK Ventures co-hosted.

It’s a mentation he’s been moving to beryllium since 2019, erstwhile helium founded the firm task limb of the Japanese electronics giant, which is present managing $500 cardinal crossed 4 funds. The AI spot startup Groq, valued astatine $6.9 billion during its astir caller backing circular past fall, is the highest-profile illustration of this thinking.

In 2020, good earlier the generative AI roar made infrastructure bets look obvious, Sauvage wrote a cheque into the company, which was founded by Jonathan Ross — 1 of the engineers who built Google’s Tensor Processing Units. Groq was focused from the commencement connected inference: the computational dense lifting that happens each clip a exemplary responds to a query. Ross had designed his spot by gathering the compiler first, stripping the architecture down until, arsenic Sauvage describes it, “you can’t region 1 portion and person it inactive work.”

It mightiness person looked niche to some, but knowing what helium did astir his genitor company’s constraints, Sauvage saw asymmetry. Unlike user hardware, which has a earthy ceiling, request for inference keeps compounding with each caller exertion and each caller model. Sauvage couldn’t cognize past that request for inference would detonate this year, acknowledgment to each AI cause that plans and acts crossed dozens of calls (where a azygous query utilized to suffice).

But successful immoderate ways, Ross was besides making a bet, too. After all, a Japanese electronics conglomerate champion known for magnetic portion is not, connected its face, the astir evident investing partner. In fact, Sauvage describes TDK Ventures’ ain beingness arsenic precise unlikely. But aft 2 back-to-back Stanford lectures — 1 making the lawsuit for firm VC, 1 cataloguing each crushed it fails — Sauvage, who is French and joined TDK successful Silicon Valley done an acquisition, pitched the thought to Tokyo office contempt having nary evident lasting to bash so. (“I’m not Japanese. I don’t talk Japanese; I don’t unrecorded successful Tokyo.”)

After refusing to instrumentality nary for an answer, helium yet received the greenish airy successful 2019 to physique a money whose mandate was to reply 1 question: What’s the adjacent large happening for TDK, and what mightiness termination it?

Image Credits:Slava Blazer for TechCrunch/StrictlyVC /

The portfolio helium has since assembled is dotted with technologies that person go much wide absorbing to VCs implicit the past year: solid-state grid transformers, sodium-ion batteries for information centers, alternate artillery chemistries that sidestep the geopolitical fragility of lithium and cobalt.

The subject down each of it is the same: place the bottleneck 4 years out, past find the founders already moving connected it.

The question, of course, is what’s next. For his part, Sauvage is watching carnal AI intimately — not each of robotics but robots with a highly circumstantial occupation to beryllium done. Agility Robotics, for example, successful his portfolio, focuses connected the single, mundane task of moving things from 1 spot to different successful warehouses facing workforce shortages. Another portfolio company, Swiss portfolio ANYbotics, builds ruggedized robots for environments excessively hazardous for quality workers — places wherever the occupation explanation is fundamentally to spell wherever radical can’t. The through-line is clarity of purpose. The robots Sauvage is betting connected don’t effort to bash everything; instead, they bash 1 hard happening reliably.

Sauvage says he’s besides watching the compute stack displacement again. GPUs dominated grooming — the massive, parallel computation of teaching a model. Inference chips similar Groq’s are reshaping what happens erstwhile that exemplary speaks: faster, cheaper, astatine scale. Now, Sauvage argues, CPUs are owed for a renaissance. They’re not the astir almighty chips oregon the fastest. But they’re the astir flexible and champion suited to the branching, decision-making logic of orchestration. When an AI cause delegates a task, checks connected its progress, and loops backmost crossed dozens of steps, thing has to negociate the full choreography. That something, increasingly, looks similar a CPU.

And past there’s China. A caller study from Eclipse — a task steadfast helium follows intimately — documented what Sauvage describes arsenic “vibe manufacturing” — the rapid, AI-assisted iteration of carnal hardware prototyping, mirroring what vibe coding did for software. Chinese manufacturers, the study found, are compressing the design-build-test rhythm for carnal products successful ways Western proviso chains aren’t yet equipped to match.

For Sauvage, it’s a bottleneck awesome — and 1 he’s already moving connected with TDK Ventures’ assorted investments. One remaining unsolved problem, helium says, is dexterity. Models are improving accelerated capable that carnal AI feels inevitable; what’s inactive missing is the carnal fluency to match. The countries and companies that fig retired however to iterate connected atoms arsenic accelerated arsenic others iterate connected codification volition person a manufacturing advantage.