Nicolas Sauvage is betting on the boring parts of AI

TDK Ventures, which now manages $500 million across four funds, is positioning around AI infrastructure and physical AI themes that it believes will become obvious over a four-year horizon. The portfolio includes Groq, valued at $6.9 billion in its most recent round, along with robotics names like Agility Robotics and ANYbotics, plus battery and grid technologies. The article is mostly strategic commentary rather than a market-moving event, but it signals continued VC interest in inference chips, CPUs for AI orchestration, and warehouse/industrial robotics.

Analysis

The key market implication is not simply “AI is good,” but that the value pool is migrating one layer lower in the stack: from model training toward inference orchestration and the control-plane around agents. That favors semis with asymmetric exposure to repetitive, latency-sensitive workloads, but the second-order winner may be the large platform owner that captures more queries per user rather than the pure-play chip vendor. If agents truly increase call counts by an order of magnitude, the near-term constraint becomes cost per action, not model quality, which should sustain capex even if frontier-model hype cools.

A less appreciated consequence is that this thesis is inherently pro-CPU and pro-networking at the margin, because agentic workloads create branching logic, coordination, and state management rather than just dense matrix math. That means the market may be underpricing the mix shift toward general-purpose compute, memory bandwidth, and interconnect, while overconcentrating on the GPU trade as the only AI hardware expression. The beneficiaries include cloud platforms with proprietary orchestration layers and enterprise software vendors that can monetize workflow control; the losers are hardware startups without a clear bottleneck advantage and any capex narrative that depends on training-heavy demand normalizing.

The contrarian risk is that investors are extrapolating “AI everywhere” faster than deployment reality can monetize it. If agentic systems remain brittle, inference growth can decelerate after the first wave of experimentation, especially if enterprises impose cost controls once bill shock hits. The physical-AI angle is also a longer-dated optionality trade: robotics and manufacturing iteration gains are real, but they are measured in quarters-to-years, not weeks, and will likely show up first as margin expansion for incumbents rather than instant revenue re-rating for venture-backed names.

Key Decisions for Investors

• Long GOOGL vs short a basket of expensive AI hardware proxies for 3-6 months: the platform owner can internalize more inference/agent traffic and monetize orchestration before pure-play infrastructure multiples rerate further.
• Add to CPU exposure via broad semis/compute leaders on pullbacks over the next 1-2 quarters: the market is still underestimating branching, stateful workloads; risk/reward improves if the narrative shifts from training to orchestration.
• Use a barbell: long infrastructure names with recurring inference demand, short venture-style robotics/physical-AI names with no near-term revenue proof, for a 6-12 month horizon; the spread should widen if enterprise buyers stay cautious.
• Buy medium-dated GOOGL calls or call spreads into any AI-capex weakness over the next 1-2 months: if agent usage continues compounding, platform monetization can surprise while downside is buffered by core cash flows.
• Avoid chasing standalone robotics beta until there is evidence of repeatable deployment economics; prefer optionality via diversified industrials over 12-24 months because dexterity breakthroughs are likely to accrue to incumbents first.