The most profitable AI bets aren't happening in large language models or chatbots; they're happening in the unglamorous infrastructure that powers data centers and keeps AI systems running 24/7. A growing number of venture capitalists are identifying critical bottlenecks in AI infrastructure four years before they become obvious, then backing the founders already working on solutions. From specialized inference chips to truck-mounted nuclear reactors, the infrastructure layer is where the next wave of AI wealth is being created.

Why Are Investors Suddenly Obsessed With "Boring" AI Infrastructure?

Nicolas Sauvage, who manages $500 million across four funds at TDK Ventures, has built his investment thesis on a simple principle: identify the bottleneck four years out, then find the founders already working on it. His track record suggests the strategy works. In 2020, well before generative AI became mainstream, Sauvage invested in Groq, an AI chip startup founded by Jonathan Ross, one of the engineers who built Google's Tensor Processing Units (TPUs). Groq is now valued at $6.9 billion and focuses exclusively on inference, the computational work that happens every time an AI model responds to a user query.

What made Groq's bet seem niche at the time has become obvious today. Inference demand is compounding with every new AI application and every new model released. Unlike consumer hardware, which has a natural ceiling for demand, inference needs keep growing as more AI agents plan and act across dozens of steps in a single task. Sauvage couldn't have predicted that AI agents would explode in 2026, but he recognized the asymmetry in the market early enough to position TDK Ventures ahead of the curve.

What Infrastructure Bottlenecks Are Investors Targeting Right Now?

The infrastructure challenges facing AI data centers extend far beyond chips. Power, cooling, and physical space are becoming critical constraints. Sauvage's portfolio reflects this reality, with investments spanning multiple layers of the AI infrastructure stack:

• Inference Acceleration: Groq's specialized chips that process AI model responses faster and cheaper than general-purpose GPUs, enabling AI systems to respond at scale without massive power consumption.
• Energy Storage and Grid Management: Sodium-ion batteries designed specifically for data centers and solid-state grid transformers that improve power distribution efficiency without relying on scarce lithium and cobalt resources.
• Physical Automation: Agility Robotics and ANYbotics, which build specialized robots for warehouse logistics and hazardous environments, addressing workforce shortages in the physical infrastructure that supports data centers.

The through-line connecting these investments is clarity of purpose. Rather than building general-purpose solutions, Sauvage backs companies solving one hard problem reliably. Agility Robotics doesn't try to build a humanoid robot that does everything; it focuses on moving packages in warehouses. ANYbotics doesn't attempt to replace human workers; it deploys ruggedized robots in environments too dangerous for people.

How Is Nuclear Power Reshaping AI Data Center Economics?

Perhaps the most striking infrastructure innovation emerging is mobile nuclear power. A Chinese research team led by Professor Wu Yican at the Institute of Nuclear Energy Safety Technology has developed a prototype of a 10-megawatt truck-mounted nuclear reactor designed to serve as a mobile power source for AI data centers, remote communities, and emergency sites.

The unit is framed as the world's first vehicle-mounted nuclear power bank at this scale and is still in the engineering-test and safety-evaluation phase. One 10-megawatt module is roughly sufficient to supply a medium-sized AI data center, which requires continuous, high-density electricity. The reactor can run for decades without refueling, essentially functioning as a mobile microreactor that eliminates what the team calls "battery anxiety" in off-grid or mission-critical applications.

"One 10-MW module is roughly sufficient to supply a medium-sized artificial intelligence data center, which need continuous, high-density electricity," noted the research team at the Institute of Nuclear Energy Safety Technology.

Professor Wu Yican, Institute of Nuclear Energy Safety Technology

The potential applications extend beyond traditional data centers. The reactor could be trucked into areas with weak or nonexistent grid infrastructure, providing baseload-grade power for small towns or islands. It could serve as a rapid-deployable energy source for disaster-relief zones, mines, construction camps, or military outposts. The team has also flagged uses for maritime propulsion and as a compact power source for orbital or deep-space missions.

What's the Next Bottleneck Investors Are Watching?

Sauvage argues that the compute stack is shifting again. GPUs dominated the training phase, the massive parallel computation of teaching a model. Inference chips like Groq's are reshaping what happens when that model speaks: faster, cheaper, at scale. Now, CPUs are due for a renaissance. They're not the most powerful chips or the fastest, but they're the most flexible and best suited to the branching, decision-making logic of orchestration. When an AI agent delegates a task, checks on its progress, and loops back across dozens of steps, something has to manage the whole choreography. That something, increasingly, looks like a CPU.

Another emerging bottleneck is physical manufacturing speed. A recent report from Eclipse, a venture firm Sauvage follows closely, documented what he describes as "vibe manufacturing," the rapid, AI-assisted iteration of physical hardware prototyping, mirroring what vibe coding did for software. Chinese manufacturers are compressing the design-build-test cycle for physical products in ways Western supply chains aren't yet equipped to match. For Sauvage, it's a bottleneck signal, and one he's already moving on with TDK Ventures' various investments.

One remaining unsolved problem is dexterity. Models are improving fast enough that physical AI feels inevitable; what's still missing is the physical fluency to match. The countries and companies that figure out how to iterate on atoms as fast as others iterate on code will have a manufacturing advantage. That's the wave for which Sauvage is positioning TDK Ventures today.

How to Identify AI Infrastructure Bottlenecks Before They Become Obvious

• Look Four Years Ahead: Sauvage's core thesis is that it takes four years for the best infrastructure bets to look obvious. Investors and founders should identify constraints that will become critical in the medium term, not the immediate future.
• Follow the Asymmetry: Seek opportunities where demand is compounding but supply is constrained. Unlike consumer hardware with natural ceilings, infrastructure bottlenecks often have unlimited upside as new applications emerge.
• Bet on Clarity of Purpose: Back founders solving one hard problem reliably rather than those attempting to build general-purpose solutions. Specialized infrastructure often outperforms generalized alternatives.
• Monitor Geopolitical Supply Chains: Watch for emerging manufacturing advantages in other regions. Chinese "vibe manufacturing" is compressing design cycles in ways that signal where Western companies need to innovate.

The unsexy infrastructure layer of AI is where the real economic value is being created. While everyone watches the latest large language model benchmarks, the investors and founders who will build generational wealth are solving the power, cooling, compute, and manufacturing problems that make those models possible at scale. The lesson is clear: in AI, the infrastructure is often more valuable than the application.