AI's real bottleneck is not the chip. It is a transformer with a five-year wait.

The most important number in artificial intelligence right now is not a parameter count or a benchmark score. It is a megawatt, and more precisely, a megawatt that has been promised but not yet delivered. Of the roughly 16 gigawatts of US data-center capacity announced for 2026 across about 140 projects, only around 5 gigawatts is actually under construction. The other 11 are stuck at the announcement stage. For 2027 the gap is wider: of some 21.5 gigawatts announced, only about 6.3 have broken ground. Add it up and there is more than 50 gigawatts of capacity that the industry has told the market it will build and has not started building. That is not a chip shortage and it is not a money shortage. It is a power shortage, and power is the part of this story that does not care how good the model is.

I have spent my career on the unglamorous end of energy, the part measured in cost per kilowatt-hour and years of lead time rather than press releases, and the AI build-out has now arrived squarely on that terrain. The chips improved on a schedule the industry controls. The grid does not run on that schedule. It runs on a slower, more physical clock, and the gap between the two is where the next few years of this story will actually be decided.

What a data center looks like to the grid

From the grid's point of view, a large AI data center is a single, enormous, almost constant load, often hundreds of megawatts in one place, drawing close to that around the clock. Building the shell and filling it with servers is the fast part. A data center building can go up in eighteen to twenty-four months. Getting the electricity to it is the slow part, and the slow part has several layers, each with its own queue.

Start with interconnection, the process of being allowed to plug into the grid at all. Nearly 2,300 gigawatts of generation and storage are currently waiting in US interconnection queues. That is more than the entire installed power capacity of the country, sitting in line. In many regions the wait now runs beyond five years. But the queue is no longer even the worst of it. The bottleneck has moved downstream, to the physical equipment that actually moves and steps the power: transmission lines, and the substations and transformers that connect them. Projects coming into service in 2025 took, on average, more than seven years from start to energized. The building was never the holdup.

Here is the number that captures the whole problem, and it is the kind of boring number I trust. A high-power transformer, the unglamorous steel-and-copper box that makes grid-scale electricity usable, had a lead time of roughly 24 to 30 months before 2020. Today that lead time stretches to about five years. You can order ten thousand of the latest accelerators and take delivery in a reasonable window. The transformer to power them is a five-year wait, and there is no version of this where the software timeline wins against the transformer timeline. The model gets better every few months. The transformer arrives when it arrives.

You can take delivery of ten thousand accelerators in a reasonable window. The transformer to power them is a five-year wait. The software timeline does not win that race. — Maya Adeyemi

Who actually pays for the electrons

Whenever a new load this size appears on a grid, the first question worth asking is the one the announcements skip: who pays for the power, and at what price. A data center does not bring its own grid. It connects to one that other people are already using, and the upgrades it requires, new lines, bigger substations, more generation, cost money that has to be recovered from someone.

Often that someone is the ordinary ratepayer. When a utility builds out its system to serve a giant new customer, the cost of that build can land in everyone's bill, and a large new buyer competing for the same finite supply of electricity can push the price up for the households already on the line. This is the part of the AI energy story that gets the least attention and deserves more, because it is a transfer. The compute is private and the profit is private, but the grid is shared, and a share of the cost of powering the boom is quietly socialized onto people who will never run a training job. Energy is never just a technical input. It is a price, and a question of who is made to pay it.

It matters for the companies too, because power is not a rounding error in this business. Energy can run from 30 to 50 percent of a data center's total operating cost. That single fact reorders the whole strategy. When power is half your running cost and five years out of your control, the cheapest, most available, most predictable electricity is not a sustainability talking point. It is the product.

There is no power that is fast, cheap, clean, and firm

Faced with the wall, the hyperscalers are doing the rational thing and going around the grid where they can, signing direct deals for their own power: renewables, nuclear, gas, and demand-response arrangements where they agree to pull back when the grid is stressed. It is worth pricing each of these honestly, because the press releases tend to feature the cleanest option and build the dirtiest one.

Renewables are the cheapest electricity per kilowatt-hour that has ever existed, but they are intermittent, and a data center needs power at three in the morning when the sun is down and the wind is still. To make solar and wind firm enough for a constant load, you have to pair them with storage, and storage adds cost and is itself supply-constrained. Nuclear is firm and low-carbon, which is exactly what this load wants, but new nuclear is slow and expensive to build, and the small modular reactors everyone points to are mostly still promises rather than priced, operating plants. Gas is the awkward truth in the middle: it is dispatchable, it is relatively fast to build, and it is therefore what a lot of this capacity will quietly run on in the near term, whatever the long-term nuclear commitments say. And demand-response, agreeing to use less when the grid is tight, is the cheapest option of all, but it only stretches so far before it starts to mean the expensive computer sits idle.

The honest summary is that there is no source that is fast, cheap, clean, and firm at the same time. Pick any three and you give up the fourth. The build-out is making that trade-off constantly, and the near-term answer, the one the timelines force, leans on gas more than the announcements admit. That is not cynicism. It is just what the clock and the cost curve allow.

Why the money is flowing to France

If you want the clearest proof that compute now follows power rather than the other way round, look at France. The French government says it has lined up on the order of €110 billion in AI and data-center investment. SoftBank alone has committed up to €75 billion to build five gigawatts of capacity, with a first phase of around €45 billion delivering more than three gigawatts in the north of the country. Brookfield has pledged €20 billion. None of this is happening because France has the best engineers or the cheapest land. It is happening because France has the one thing the build-out is actually short of: firm, cheap, low-carbon power, already built.

France draws roughly 70 percent of its electricity from nuclear reactors, is the largest net electricity exporter in the world, and has industrial power prices well under half of Britain's. There is a long-term pricing arrangement that puts a ceiling near €70 per megawatt-hour on a large block of that nuclear output from 2026. For a business where energy is 30 to 50 percent of operating cost and price volatility is the enemy of a ten-year capital plan, a predictable €70 ceiling is worth more than any tax break. France built that nuclear fleet decades ago for reasons that had nothing to do with AI, and it now turns out to be holding the asset the entire industry wants. The lesson is not really about France. It is that the places that built firm, cheap power years ago, for their own reasons, are the places the compute will go, and the places that did not cannot conjure it on a capex timeline.

Efficiency stops being a virtue and becomes the product

The same logic explains a deal that might otherwise look like ordinary corporate maneuvering. Foxconn and Intel announced this week that they will jointly build full data-center systems, not just chips: server racks built around Intel processors and accelerators, but also the high-speed interconnects, the cooling, and what they describe as energy-efficiency solutions for large-scale AI. Notice where the emphasis sits. The interesting parts of that list are the boring parts. Cooling and efficiency are not features you brag about when power is cheap and plentiful. They become the headline when power is the binding constraint.

Price it out and the reason is obvious. A more efficient rack is worth exactly the power it saves, multiplied by the price of that power, multiplied by the years it runs. When power is half your operating cost and effectively rationed by a five-year transformer queue, every watt you do not waste is a watt you do not have to wait years to procure. Efficiency, in other words, is the one lever on this problem that a company can actually pull on its own schedule, without waiting in anyone's queue. That is why the system-level, performance-per-watt approach is suddenly where the industry's attention is going. It is the only part of the power problem that is not someone else's timeline.

The number that decides it

Most of the coverage of this moment centers on the US and Europe, but the underlying logic is global and it is unforgiving. Compute lands where there is firm, cheap power with headroom to spare, and it does not land where the grid is already strained. In much of the world, including the grids I have spent the most time on, the megawatt an AI data center wants is the same megawatt that would otherwise go to electrifying a town. Where power is abundant that trade is easy. Where it is scarce it is sharp, and the data center, with its deep pockets and its willingness to pay, tends to win the bidding against uses that cannot.

So the number to watch is not the capex headline, which is easy to announce and cheap to revise. It is the interconnection queue, the transformer lead time, and the price of firm power in the places the build-out wants to go. Those three are slow, physical, and stubborn, and they are now the rate limit on the whole enterprise. The models have been improving faster than the grid can be rebuilt, and the grid does not grade on a curve. The bottleneck stopped being the chip a while ago. It is the substation, and the substation keeps its own time.