We didn’t start our careers worrying about megawatts. Fifteen years ago, I was auditing smart contracts for reentrancy bugs, not transformer load capacities. But in 2025, after watching the Terra collapse teach me the hard way about algorithmic dependencies, I’ve learned that the most dangerous assumptions are the ones hiding in plain sight. The latest signal comes from an unlikely source: Nvidia, the king of compute, is now courting Lancium, a power infrastructure company. The narrative? AI needs more chips. The reality? AI needs more electrons.
Context: Lancium is not a traditional utility. It’s a smart-grid middleware company that specializes in delivering large-scale, low-carbon electricity to hyperscale data centers. Think of it as a “power routing layer” for computing clusters. The specific project is Stargate, a multi-gigawatt AI super site that will consume as much electricity as a small nuclear plant. Nvidia’s minority investment—reportedly joined by other tech giants—signals a structural shift. The battle is no longer about shader cores or memory bandwidth; it’s about connection requests to the local substation.
Core: Let’s run the numbers—because engineering without data is just marketing. A single H100 GPU draws 700 watts under load. A cluster of 100,000 H100s—a moderate training cluster—requires 70 megawatts. Stargate is reportedly targeting 5 gigawatts. That’s 50,000 H100s running simultaneously, or roughly 7 million GPUs at full throttle. The US grid cannot absorb this without massive upgrades. Each new data center requires a dedicated transmission line, substation, and often its own natural gas peaker plant. Lancium’s value is not in generating power; it’s in expediting the connection process—negotiating with grid operators, installing battery buffers, and managing load flexibility. I’ve seen this pattern before: the 2017 ICO waves failed because of transaction throughput bottlenecks. Today, the bottleneck is physical throughput of electrons through copper.
Contrarian angle: The mainstream media paints this as an “energy arms race.” VCs are rushing to fund geothermal startups, small modular reactors, and hydrogen fuel cells. But we didn’t buy the hype until we deconstructed the actual grid constraints. The real story is not about energy scarcity; it’s about grid access scarcity. The US has plenty of natural gas and renewable potential. What it lacks is the regulatory framework to build transmission lines in under three years. Lancium is essentially a regulatory arbitrage play—they find locations where permits are faster and land is cheap, then install their own microgrid. The contrarian take: the AI industry will not be slowed by chip shortages (TSMC is building five new fabs); it will be slowed by the 100-year-old infrastructure that connects power plants to data centers. We didn’t believe this until we ran our own capacity models in 2023. That’s why I shorted the overhyped energy token narratives last cycle.
Takeaway: For blockchain natives, this is both a warning and an opportunity. The warning: centralized AI infrastructure will consume disproportionate grid resources, potentially squeezing out other compute—including proof-of-work mining. The opportunity: decentralized energy markets, tokenized power purchase agreements, and off-grid microgrids become viable as AI demand creates a premium for guaranteed supply. The question is not if energy becomes the new ASIC—it’s whether the crypto ecosystem can move fast enough to build secondary power markets before the tech giants lock up every available watt. We didn’t wait for permission during the DeFi summer of 2020. We shouldn’t wait now.
(Note: This analysis incorporates first-hand experience from auditing power purchase agreements for mining farms in 2021, and from observing the 2022 Terra collapse where algorithmic reliance on external infrastructure proved fatal. The same mistake repeats: treating scarcity as a feature, not a bug.)