Power, Not Capital, Is the New Constraint in AI

For most of the last decade, the dominant constraint on technology was capital. If you had money, you could buy servers, rent cloud capacity, hire engineers, and ship product. Compute was expensive but fungible. You could throw dollars at almost any bottleneck and eventually get through it.

That assumption just broke.

The thing stopping the next wave of AI from being built is not money. It is not talent. It is not even chips, though chips get all the press. The binding constraint right now, the thing that is genuinely scarce, is electricity — and the physical infrastructure to deliver it reliably at scale.

Dylan Patel and the team at SemiAnalysis have been publishing detailed analysis on this for a couple of years now. Their work on the AI datacenter industry model is some of the most rigorous public analysis I’ve read on the infrastructure side of AI. The core finding: SemiAnalysis estimates a gap north of 50 gigawatts between needed AI power generation and what is actually being added to the US grid by 2028. Fifty gigawatts. That is not a rounding error. That is equivalent to roughly half of all US power consumption growth over the last decade, needed specifically for AI compute, in a four-year window.

I want to sit with that number for a moment because it changes how I think about the entire AI stack.

Why Power Became the Bottleneck

The shift happened gradually and then suddenly. In 2022 and into 2023, the bottleneck in AI buildout was chip packaging — specifically CoWoS, the advanced packaging technology that NVIDIA depends on to assemble its H100s. There were not enough chips to go around and everyone who wanted to train a frontier model was on a waitlist.

That constraint eased as TSMC and others ramped capacity. What nobody adequately planned for was what happens when you actually deploy all those chips. Each H100 draws roughly 700 watts. A rack of eight draws about 5.6 kilowatts. A datacenter with 100,000 H100s — not a hypothetical, Microsoft and Google are building at this scale now — needs somewhere around 70 megawatts just for the GPUs, before you count cooling, networking, or facility overhead. A single large AI campus today can demand 400 to 500 megawatts.

SemiAnalysis estimates that NVIDIA alone will have shipped accelerators representing the power needs of 5 million H100-equivalent units through end of 2024. And the demand crossing 10 gigawatts in early 2025 was their conservative scenario.

Here is the problem: building a datacenter takes 18 to 24 months in a fast execution. Getting a new power substation permitted, built, and connected to the grid takes five to seven years in most US jurisdictions. The project queue for new power generation has wait times that SemiAnalysis found average over five years.

You can raise $100 billion for AI infrastructure tomorrow morning. You cannot compress a five-year grid interconnection process into six months. That asymmetry is the real constraint.

The Workarounds Being Deployed Right Now

I have been watching how serious operators are solving this, because where they spend their engineering effort tells you what they believe.

The first wave of workarounds is on-site generation. AI labs are building and leasing natural gas turbines — specifically aeroderivative turbines, which can spin up in minutes rather than hours — and connecting them directly to their campuses. The economics make sense: SemiAnalysis estimates a gigawatt of AI datacenter capacity can generate $10 to 12 billion annually in revenue. Getting a 400 MW facility online even six months earlier than grid timing allows is worth billions.

Some facilities are looking at fuel cells. Some are exploring small modular nuclear reactors for the 2030s. Microsoft signed a deal to restart the Three Mile Island nuclear plant. These are not incremental plays — they are signals that the top players genuinely believe power is the constraint for the next ten years.

The second wave of workarounds is efficiency improvement. Power-constrained operations care about throughput per watt far more than raw throughput. This is pushing the industry toward more efficient inference hardware, better cooling (liquid cooling is now the norm in serious deployments, not a premium option), and workload scheduling that minimizes peak demand.

What This Means If You Are Building

I have spent time thinking about what energy-bound AI means for people building actual products, not just operating hyperscale infrastructure.

First: geography is going to matter more in AI than it has since the mainframe era. The cost of inference — and therefore the cost of your AI product’s compute — will increasingly be a function of where the datacenter sits and what its electricity cost is. Regions with cheap renewables, existing grid capacity, and favorable regulatory environments are going to become genuine competitive advantages for cloud providers. Iceland, Texas (with its deregulated grid), parts of Norway and Canada — these are not exotic choices anymore. They are strategic infrastructure decisions.

Second: the energy constraint is going to accelerate the push toward model efficiency even faster than benchmark pressure would. When you are power-capped, a model that achieves equivalent quality at half the compute is worth double — literally. This is why I expect inference efficiency to become one of the most competed-on dimensions in foundation model development over the next two to three years. It is not just about costs; it is about physical capacity to serve demand.

Third: I think this creates a genuine moat for whoever solves the power problem well. The hyperscalers with large land banks, existing utility relationships, and teams that understand grid interconnection are building a type of infrastructure advantage that is very hard to replicate quickly. A startup that builds great software on top of this infrastructure is in a different position than one trying to compete at the infrastructure layer. That boundary is getting cleaner.

The Deeper Shift

There is something philosophically interesting about this transition. For most of the software era, the key inputs to computing were human capital and financial capital. Both are relatively liquid. You can hire engineers anywhere in the world. You can raise money in weeks if your story is right.

Physical energy infrastructure does not work that way. It is slow, local, regulated, capital-intensive in a way that does not compress with urgency, and deeply tied to the physical world — land, transmission lines, fuel supply chains, water for cooling. AI has become a problem that lives in the physical world in a way that pure software never was.

I spent years as an enterprise architect at companies like HP and Dell/EMC. The conversations we had about datacenter capacity planning were serious but bounded — we were thinking about megawatts, not gigawatts, and the timelines were years, not decades. What is happening now is genuinely different in kind, not just in scale.

The bottleneck has always shaped the industry. When the bottleneck was transistors, the industry built TSMC. When the bottleneck was bandwidth, the industry built fiber. The bottleneck is now electrons — and the grid and generation infrastructure to reliably deliver them. That is what the next decade of infrastructure investment will be about, whether the people building AI products realize it or not.

---