The market is pricing AI like it runs on code alone. Kimmeridge, an energy-focused investment firm, just threw a wrench into that narrative: nearly half of U.S. data centers are facing delays. Not because of chip shortages or software bugs. Because of politics, power grids, and the physical world's refusal to move at silicon speed.
Tracing the gas leaks before the code compiles, this isn't a supply chain blip. It's a structural confession. The AI trade has hit its first true physical bottleneck, and the market hasn't fully repriced the friction.
Kimmeridge isn't a random voice in the wilderness. They deploy capital into energy infrastructure. When they flag delays, they're reading their own portfolio's pain points. Their warning centers on political backlash and regulatory hurdles stalling data center construction across the U.S. The immediate cause: communities pushing back on land use, water consumption, and rising electricity costs. The underlying cause: a fundamental mismatch between AI's exponential appetite for compute and the linear, bureaucratic pace of building the physical plants that house it.
This is not a temporary hiccup. It's the collision of two different time scales. One is measured in GPU generations and model training runs. The other is measured in transformer delivery lead times and municipal zoning board meetings.
The core issue is order flow, but not the kind I usually analyze. On-chain, we track liquidity pools and arbitrage windows. Here, the order flow is electrons and cooling water. The latency isn't measured in milliseconds; it's measured in months and years. Grid interconnection queues in the U.S. are backlogged. Key equipment like transformers has a lead time of one to two years. This isn't a software patch problem; it's a steel, copper, and concrete problem.
The market structure here is a classic supply-demand imbalance with a regulatory twist. Demand for AI compute is surging from every hyperscaler and well-funded startup. Supply is constrained by physics and local politics. The result is that already-operational data centers gain massive pricing power. New projects face cost overruns and timeline slippage. This creates a two-tier market.
I saw this pattern in DeFi during the 2020 liquidity mining craze. Projects subsidized TVL with token emissions, creating an artificial impression of demand. When the subsidies stopped, the real users vanished. The same principle applies here, but the stakes are physical. The subsidies are cheap electricity deals and tax breaks offered by desperate local governments. The real users are AI models that need reliable, massive compute. If the power doesn't flow, the model doesn't train.
The contrarian angle is uncomfortable. The market narrative is that AI is a software revolution. The reality is that it's becoming a utility business, and utilities are regulated, slow, and politically contentious. The winners won't just be the best algorithm teams. They'll be the firms that can navigate zoning laws, secure power purchase agreements, and manage community relations. This is the "boring" infrastructure work that tech founders hate. But it's the new moat.
Silence between the blocks tells the real story. The silence here is the absence of new data center announcements from the usual suspects. They're hitting the same walls. The hidden risk is that this accelerates compute centralization. The giants with locked-in resources and long-term power contracts will pull further ahead. New entrants will find it nearly impossible to secure the physical footprint needed to compete. This is the anti-competitive outcome that no algorithm can solve.
I've spent two decades watching markets misprice physical constraints. In 2022, the LUNA collapse taught me that models relying on infinite growth assumptions are fragile. This is the same lesson in a different costume. The market is pricing infinite compute growth. The physical world is offering linear, contested growth. The gap between those curves is where the risk lives.
The opportunity is in the friction. For investors, this means looking at companies solving the bottleneck: liquid cooling specialists, modular data center builders, and energy management software. The firms that can reduce the physical footprint per FLOP will be the arbitrageurs of this cycle. The takeaway is not to abandon AI exposure, but to shift it from pure-play models to the infrastructure layer that enables them. The model didn't fail, but the container holding it is cracking.
We're moving from the era of the algorithm to the era of the asset. The next bull market won't be won by the best code alone. It will be won by whoever can flip the switch on the power. Watch the grid, not just the GPU. The rug wasn't pulled; it just hasn't been built yet.