High-voltage transmission lines — the real bottleneck for AI. The chips are ready; the grid that must carry their power is years behind. Photo: Matthew T Rader (CC BY-SA 4.0).

There is a line making the rounds among people who build artificial intelligence for a living, and it is not about chips. It goes something like this: the new bottleneck isn't compute, it's the kilowatt. Satya Nadella put it more bluntly, describing GPUs sitting in inventory that he cannot plug in because there is no power to run them.

For about fifteen years, American electricity demand was essentially flat. Efficiency gains canceled out growth, and utilities planned for a quiet future. That era ended abruptly. AI data centers have become the fastest-growing new load on the grid, and the result is the defining infrastructure problem of this decade: not whether we can make the chips, but whether we can power them.

It is worth being precise about what changed. The constraint on AI used to be the supply of advanced chips. That is still tight, but it is now downstream of something harder. You can manufacture a GPU in months and stand up a data-center shell in a year. You cannot, on that timeline, build a power plant, string a high-voltage transmission line, or get a new connection approved. The slowest link in the chain moved from the fab to the grid.

A lead comes in on Instagram. Another on Messenger. Three more on SMS.

Your team switches tabs, repeats answers, and loses context while hot leads wait hours for replies. At 2am, nobody responds at all.

That’s not a people problem. It’s a process problem.

Wati brings Instagram DM, Facebook Messenger, TikTok, WhatsApp, SMS, RCS, and web chat into one AI-powered inbox. Automations instantly respond, qualify leads, and route conversations to the right person, 24/7.

Your team stops firefighting. Your leads stop waiting. Your pipeline starts moving.

See How Wati Works

The numbers come with wide error bars, and that honesty matters. US data centers used roughly 176 terawatt-hours in 2023, about 4.4% of national electricity. Lawrence Berkeley National Laboratory projects that could reach 325 to 580 TWh by 2028. The Electric Power Research Institute models a range of 380 to 790 TWh by 2030. Note the ranges: anyone quoting a single confident number is overselling the certainty.

US data-center electricity use, drawn as the range each study actually gives. The gap between low and high cases is enormous, which is the point. Source: LBNL, EPRI. Chart: Silicon & Steel.

Globally the pattern repeats. The IEA estimates data centers used around 485 TWh in 2025, rising toward 950 TWh by 2030, roughly a doubling. China, for its part, added an estimated 550 TWh of total demand in 2024 alone, about twice the US pace, a reminder that this is not only an American story.

Translate that into share of the grid and the strategic weight becomes clear. Data centers go from about 4.4% of US electricity in 2023 to a modeled 9 to 17% by 2030. The 17% figure is an upper-bound scenario, not a base case, and should be read that way. But even the floor of that range roughly doubles today's share, in a system that was built for flat demand.

Data centers as a share of US electricity. The upper bound is a scenario; the lower bound still represents a structural break from fifteen flat years. Chart: Silicon & Steel.

Here is where the wall is most visible. To connect anything to the grid, a new power plant or a large load, you join an interconnection queue. At the end of 2024 that queue held roughly 2,290 gigawatts of proposed capacity, with a median wait of about 4.5 years. An important caveat: queue gigawatts are not built megawatts. Most projects in the queue never get completed. But the wait itself is the constraint, four years is an eternity when a model generation lasts eighteen months.

Large-load interconnection requests in Texas, before and after the AI surge. Requests overstate what gets built, but the jump is the signal. Source: ERCOT. Chart: Silicon & Steel.

The pressure shows up in the data. In Texas, large-load interconnection requests to ERCOT jumped from about 63 to 226 gigawatts in a single year. And the PJM capacity auction, the market that pays generators to be available, cleared at a record $329.17 per megawatt-day, a price that ultimately flows toward everyone's bill.

Part of the problem is density. A conventional server rack a few years ago drew perhaps 10 kilowatts. Nvidia's GB200 NVL72 rack draws roughly 120 to 132 kilowatts, and the newer GB300 pushes toward 150. A single modern AI rack can demand more power than ten older racks, which is why these facilities concentrate load in a way the grid has never had to serve before, and why cooling and power delivery have become as important as the chips themselves.

Power per rack, nameplate figures. AI accelerators concentrate electrical load to a degree the grid was never designed around. Chart: Silicon & Steel.

Faced with this, the industry has gone looking for clean, firm power wherever it exists, and rediscovered nuclear. The signature deal: the planned restart of a reactor at Three Mile Island, about 835 megawatts, under a twenty-year power-purchase agreement with Microsoft, targeted to come online around 2027. Across roughly thirteen announced nuclear arrangements, restarts, small modular reactor plans, and PPAs, more than 9.7 gigawatts has been linked to AI demand. A technology written off a decade ago is suddenly the buildout's favorite power source.

Nuclear is slow to add, so the near-term bridge is natural gas, and the bottleneck simply moves again. GE Vernova, a leading turbine maker, is reported to have a gas-turbine backlog of roughly 80 gigawatts stretching into 2029. You cannot buy your way to the front of that line either. The wall is not a single barrier; it is a series of them, chips, then connections, then generation, then the machines that make generation.

For the investor
The cleanest exposure to AI may not be chips at all, but power: independent power producers, nuclear operators, turbine and grid-equipment makers, transformers, and cooling. These are slower, more regulated, more cyclical businesses than software, and demand forecasts are wide, so size to the range, not the headline. But the bottleneck is real and physical.

For the data-center & infrastructure leader
Treat power procurement as the long pole, not an afterthought. Secure interconnection and generation years ahead, consider on-site and behind-the-meter options, and assume four-plus-year grid timelines. The site with power beats the site with a better tax deal.

For the policymaker
Interconnection reform and transmission permitting are now AI policy, whether or not anyone labels them that. The choke point is process as much as physics. And watch the distributional question: if data-center load lifts capacity prices, ordinary ratepayers can end up subsidizing the buildout unless rates are designed to prevent it.

1. LBNL 2024 US Data Center Energy Usage Report, the careful base case [20 min]

2. EPRI, Powering Intelligence, data-center load scenarios [20 min]

3. IEA, Energy and AI, the global picture [25 min]

4. Constellation, the Three Mile Island restart and Microsoft PPA [8 min]

The listing is coming. But the investors who'll profit most aren't waiting — they're already in the three public companies with direct SpaceX revenue exposure. Here's what they're buying.

Get the Analysis

TWh / GW: a terawatt-hour is a unit of energy used over time; a gigawatt is a unit of power, the rate at which energy is delivered. Grids are sized in GW, bills in TWh.

Interconnection queue: the waiting list to connect a new power plant or large load to the grid; now a multi-year bottleneck.

PPA (power-purchase agreement): a long-term contract to buy electricity at a set price, the instrument behind most big AI power deals.

Capacity auction: a market, such as PJM's, that pays generators to be available when needed; record prices signal scarcity.

Behind-the-meter: generation built on-site, bypassing the public grid and its queue, increasingly attractive to data centers.

Baseload: steady, around-the-clock power; nuclear and gas provide it, which is why AI is drawn to them over intermittent sources alone.

SMR (small modular reactor): a smaller, factory-built nuclear design promised for the 2030s; much hoped-for, not yet at scale.

Sourced facts. The 2023 baseline (~176 TWh, ~4.4%), the LBNL 2028 range (325 to 580 TWh), the EPRI 2030 range (380 to 790 TWh, ~9 to 17%), the IEA global figures (~485 TWh in 2025 to ~950 by 2030), the ~2,290 GW interconnection queue and ~4.5-year median wait, ERCOT's 63-to-226 GW jump, the PJM $329.17/MW-day clearing price, rack-power figures (GB200 ~120 to 132 kW, GB300 ~150 kW), the Three Mile Island restart (~835 MW, 20-year Microsoft PPA, ~2027), the ~9.7 GW across nuclear deals, and GE Vernova's ~80 GW turbine backlog come from LBNL, EPRI, IEA, grid-operator and market data, and company statements.

What is estimate, and what is interpretation. Every forward demand figure is a wide range, not a point, and I have shown the ranges deliberately; the 17%-of-grid figure is an upper-bound scenario. Interconnection-queue gigawatts overstate what will actually be built. Calling power, rather than chips, the binding constraint is my framing, the counter-case, that interconnection reform and efficiency could ease it faster than expected, is real and worth holding alongside it.

Silicon & Steel Intelligence Desk, Supply Chain Strategy & Semiconductor Analysis. Nothing here is investment advice. Please check out more offerings on https://siliconandsteel.co/. Corrections & coffee: [email protected]