For most of the AI boom, the bottleneck was always the chip. Could Nvidia make enough of them? Could you train something bigger without the whole run falling over? Could anyone actually get their hands on the memory these models need?
The issue at hand is still ongoing, but now there’s another concern that’s come up – and it’s not nearly as thrilling: the electrical grid.
The growth of AI data centres from megawatt-scale to gigawatt-scale has hit a roadblock that has nothing to do with software. The issue is that you can’t simply connect a power plant to a server rack and expect it to work. Electricity needs to travel long distances, be rerouted, and pass through multiple voltage transformations before it can be used by a chip. This process requires a lot of heavy equipment, which is currently in short supply.
The problem is that this equipment, such as transformers and transmission lines, is not available in the quantities needed to support the growing demand for power in AI data centres.
As a result, the industry is facing a major challenge in terms of providing the necessary infrastructure to support the growth of these data centres.
The equipment needed to transmit and transform electricity is not only heavy, but also complex and expensive, making it difficult to scale up to meet the increasing demand. Furthermore, the process of transmitting electricity over long distances and transforming it to the right voltage is not straightforward, and requires careful planning and execution to ensure that the power is delivered safely and efficiently.
Overall, the industry is facing a significant challenge in terms of providing the necessary power infrastructure to support the growth of AI data centres, and it will require significant investment and innovation to overcome this hurdle
What Does 10 Gigawatts Actually Look Like?
Ten gigawatts is a genuinely hard number to picture, so here’s a comparison: it’s roughly what it would take to power eight million homes at once, all the time.
This isn’t hypothetical. In March 2026, the US Department of Energy announced a partnership with SB Energy, SoftBank, and AEP Ohio to build a 10-gigawatt data-centre campus alongside 10 GW of brand-new power generation to feed it. Separately, Nvidia and OpenAI have announced their own plans for at least 10 GW of Nvidia-powered AI clusters.
For context, a fairly ordinary data centre today draws somewhere in the tens of megawatts — think the electricity footprint of a small town. A big hyperscale cloud campus might draw a few hundred megawatts, closer to a medium-sized city. A 10-gigawatt AI cluster is a different order of magnitude entirely: 10,000 megawatts, or roughly the demand of an entire regional grid.
Here’s the part that surprises people: having a 10 GW power station nearby doesn’t mean that power is ready to use. Electricity leaves a power station at hundreds of thousands of volts, because high voltage is the efficient way to move electricity long distances without losing most of it as heat along the way. AI chips, on the other hand, run on roughly 1 to 12 volts of direct current. Getting from one to the other means climbing down a very tall staircase, and every step of that staircase is its own bottleneck.
From Power Station to Chip: The Step-Down Chain
Power can’t be wired straight from the grid into a server rack. It goes through several stages first:
- High-voltage transmission lines carry power in at anywhere from 115,000 to 765,000 volts, because higher voltage means less energy wasted over long distances.
- Substation transformers — huge outdoor units — step that down to a more manageable “medium” voltage, typically around 34,500 volts, for distribution across the site.
- On-site transformers drop it further, down to something like 480 volts of AC power.
- Switchgear and UPS systems (uninterruptible power supplies — essentially industrial-scale surge protectors and backup batteries) smooth out spikes and outages.
- Server power supplies do the final conversion, turning AC into the roughly 1-volt DC trickle that a processor actually runs on.
At a normal scale, none of this is remarkable — it’s just how electrical infrastructure has always worked. The trouble starts when you try to do it at gigawatt scale. You’re not just running thicker cable; you need acres of industrial-grade equipment built to handle a continuous, enormous electrical load without failing.
Why Transformers, of All Things, Are Now Strategic Assets
A transformer’s job is to raise or lower electrical voltage using magnetic fields, without changing the frequency of the current. It sounds mundane, but nothing on the grid works without one.
Here’s the problem: you can’t just order a giant power transformer the way you’d order server racks. These are multi-tonne, custom-built machines, engineered for the specific electrical characteristics of one site, and made from a fairly narrow supply of specialised electrical steel, huge quantities of copper, and specialised insulating fluid — built by a shrinking pool of skilled industrial workers.
Demand for data centres has pushed lead times for these transformers out past 160 weeks — over three years — forcing utilities and tech companies to order them years before they’ve even broken ground. That creates a strange kind of bottleneck: a company can have billions in the bank, land secured, and warehouses stacked with GPUs, and still have a dark, silent data centre — because the one transformer it needs to actually connect to the grid won’t arrive for another three years.
It’s a Local Problem, Not Just a Global One
Discussions about AI’s energy appetite tend to focus on whether the world can generate enough electricity overall. The International Energy Agency estimated that data centres used around 415 terawatt-hours globally in 2024 — about 1.5% of total electricity use — and expects that to rise to roughly 945 terawatt-hours by 2030. On a global scale, that’s manageable.
Locally, it’s a different story. AI campuses cluster huge demand into small geographic areas, and a country can have plenty of spare power on paper while a specific county simply doesn’t have the substations or high-voltage lines to deliver it to one new 500-megawatt building. The IEA estimates that around a fifth of planned data-centre projects are already facing delays because of local grid constraints, with wait times for key components roughly doubling. It’s why developers now hunt for land with existing high-capacity grid access first, and worry about the price of the land itself second.
Then the Problem Gets Worse Inside the Building
Even once power reaches the site, there’s a second squeeze happening inside the server room itself. A rack of servers that used to draw a few kilowatts now routinely draws over 100 kilowatts, and the newest facilities are being designed around single racks pulling a full megawatt. Power density in AI server racks has roughly increased elevenfold since 2020.
This runs into a basic law of electrical engineering: power equals voltage multiplied by current. If you want more power but keep voltage low, current has to rise sharply to compensate — and higher current means thicker, heavier copper cabling, plus more energy wasted as heat (a phenomenon called resistive, or “I-squared-R,” loss, which is exactly why a phone charger gets warm). At megawatt-per-rack scale, you simply can’t keep making the copper thicker. The only real fix is to raise the voltage instead.
That’s why Nvidia and others are now designing systems around 800 volts of direct current, delivered much further into the building than the traditional 480-volt AC setup. Skipping several conversion steps means less power lost as heat, thinner and lighter cabling, and racks that can safely handle a full megawatt each.
Big Tech Is Turning Into Heavy Industry
The first wave of the AI boom made stars out of chip designers and memory manufacturers. The current wave is doing the same for an unglamorous set of heavy-industrial suppliers: transformer manufacturers, switchgear and circuit-breaker builders, utility-scale generator makers, high-voltage cable producers, and industrial cooling companies. Firms from Siemens to Generac have been ramping up capital spending just to keep pace with data-centre orders.
AI, in other words, has stopped being purely a software industry. It’s now a hybrid of advanced computing, heavy electrical engineering, and civil infrastructure — and the second two move a lot slower than the first.
Software Is the One Lever Left
Given how physically constrained everything else has become, software efficiency has quietly turned into an energy strategy in its own right. If researchers can improve algorithms, or use techniques like quantisation (reducing the mathematical precision a model uses, which cuts computing cost with only a small accuracy trade-off) to get twice as much useful output from the same megawatt, that’s effectively a second data centre — without waiting years for a substation.
It’s changing how the industry even measures itself. The old scoreboard was simply “how many GPUs do you have.” The new one is closer to “how much useful output can you get per megawatt available” — because for the first time, megawatts, not chips, are the scarcer resource.
Silicon Valley Speed Meets Grid-Building Reality
That’s really the whole story in one line: software moves at the speed of a deploy button, and the grid moves at the speed of steel and copper. A model can be updated across a million machines overnight. A transformer takes years to forge, wind, ship, and install. Code can be rewritten in an afternoon. A new transmission line needs years of permitting, land rights, and physical construction before a single watt flows through it.
The companies that come out ahead in this next phase won’t just be the ones with the cleverest models — they’ll be the ones that ordered their transformers years ago, redesigned their internal wiring around higher voltages, and locked in a physical connection to the grid while they still could. At ten-gigawatt scale, the real question isn’t whether we can write smart enough code. It’s whether we can build enough heavy machinery to get the power there in the first place.