Nicolas Sauvage runs TDK Ventures — the corporate venture arm of Japanese electronics giant TDK — with $500M across four funds. His investment thesis is simple and contrarian: the money in AI isn't in the applications everyone talks about. It's in the infrastructure that makes AI work.

His framing: "Identify the bottleneck four years out, then find the founders already working on it." It takes four years for the best bets to look smart.

That philosophy has led him to some of the most interesting infrastructure bets in AI.

What "Boring AI" Actually Means

"Boring" in Sauvage's vocabulary doesn't mean uninteresting — it means foundational. The technologies that solve fundamental bottlenecks rather than producing flashy demos:

Inference chips: Compute for AI model responses. This is Groq — the AI inference chip startup now valued at $6.9 billion, founded by a Google TPU engineer. While everyone was focused on training chips, Sauvage was betting on inference performance. The reasoning: as AI moves from training to deployment at scale, inference costs become the dominant cost structure.

Power infrastructure: Solid-state grid transformers and sodium-ion batteries for data centers. The AI boom is a power boom. Data centers are competing for electricity in ways they've never had to before. The companies solving power delivery and storage for AI infrastructure are in the bottleneck.

Robotics for mundane tasks: Agility Robotics (warehouse robots moving items) and ANYbotics (ruggedized robots for hazardous environments). These aren't humanoid robots doing impressive demos — they're single-task robots that do one boring thing reliably. Sauvage's bet: the AI value isn't in the robot body, it's in the AI that makes the robot useful.

The Physical AI Thesis

Sauvage is positioning around "physical AI" — AI systems that interact with the physical world. This is the convergence of software AI and hardware that isn't a phone or laptop.

His prediction that caught my attention: CPUs will see a renaissance for AI agent orchestration. The reasoning is practical: as AI agents handle increasingly complex multi-step tasks, the orchestration layer — coordinating which models run where, managing state, handling errors — requires general-purpose compute that GPUs aren't optimized for. CPUs are back.

This is a contrarian bet. The industry narrative is that GPUs are winning everything. Sauvage is arguing that the agent orchestration layer will be a different computing problem — and that the infrastructure optimized for inference chips and model training isn't the same as what's needed for agent coordination.

Why Boring Infrastructure Wins

The pattern Sauvage is betting on is the same pattern we've seen in every major technology shift:

  • The applications get the attention and the valuations
  • The infrastructure captures the durable value
  • The bottlenecks are always somewhere unexpected

In cloud computing, everyone remembers the dot-com applications. The durable winners were the infrastructure companies — AWS, Cloudflare, Datadog. In mobile, everyone talked about the apps. The durable winners were the infrastructure — Firebase, Twilio, the companies that built the platforms apps were built on.

Sauvage's "boring AI" thesis is the same pattern applied to AI: find the bottlenecks that the application companies need but don't want to build themselves, and back the founders working on those problems.

What This Means for AI Builders

If you're building AI applications, Sauvage's thesis has a practical implication: the infrastructure you're building on top of is being capitalized aggressively. TDK Ventures' willingness to invest in inference chips, power management, and robotics means the bottleneck-solving infrastructure will be funded — which means your application can rely on improving infrastructure rather than having to build it yourself.

If you're deciding where to build, the boring AI thesis suggests: look for the bottlenecks. What's the thing that every AI application needs but no one wants to build because it's not glamorous? That's where the durable infrastructure value accumulates.

The boring parts of AI are where the real money is being made — quietly, by investors who think four years ahead.

Related posts: AI Agent Infrastructure Readiness — the $1.7T gap between models and production. Venture Capital's $3.2B AI Bet — a16z and Katie Haun targeting AI infrastructure. BMW i Ventures $300M AI Fund — industrial capital betting on agentic and physical AI. ComfyUI at $500M — node-based AI workflow tools reaching $500M valuation. Samsung $1T AI Chips — HBM chip shortage driving $1T semiconductor valuations. QuTwo at $380M — compute orchestration layer routing AI across classical, quantum, and hybrid substrates.