The constraint of the AI era is no longer processing capacity. It is power. Grid connection, interconnection queues and the manufacture of high-voltage transformers all run on cycles measured in years, and none of them kept pace with the jump in demand. For energy companies, an infrastructure bottleneck has become the single largest value shift in a generation.
For two years, the conversation about artificial intelligence revolved around one component: the chip. Whoever had the GPUs pulled ahead; whoever didn’t waited. It made sense. That is where the boom began.
But the constraint has moved, and much of the public debate hasn’t caught up. Today, whether a multi-billion-dollar data centre comes online in 2027 or in 2031 is no longer decided by silicon. It is decided by the socket. More precisely: by the queue to connect to the grid, the transformer that won’t arrive, and the substation that doesn’t yet exist. The chip scaled. The electrical infrastructure did not.
For anyone leading an energy company, that sentence deserves a pause. The most contested resource of the digital decade is now precisely what the power sector produces, transmits and guarantees. The question this raises isn’t operational. It is strategic.
Why did power become the biggest constraint on AI?
Because chip supply scaled and the grid didn’t. Advanced semiconductor packaging capacity has doubled more than once since 2024, and GPU volumes followed. The grid has no such rhythm. Interconnection, transformers and utility capacity planning run on cycles measured in years, not in chip generations.
The Uptime Institute was blunt in its 2026 predictions: power is the constraint that defines data centre growth worldwide, and no developer is going to “outrun the power shortage”. The institute projects an additional 75 to 125 GW of global data centre power demand through 2030 (Source: Uptime Institute, Five Data Center Predictions for 2026).
The reading that matters here isn’t scarcity. It is displacement. The choke point has left one sector, semiconductors, and entered yours.
What is the grid bottleneck, in practice?
It is three failures stacked on top of each other.
First, the connection queues. Connecting a large new load to the grid requires a transmission-capacity assessment, construction and permitting, and all of it accumulates over multi-year timeframes. EPRI estimates that in some regions a data centre relying on the grid alone can take up to ten years to energise. Ten years. That is longer than the useful life of much of the hardware it would house.
Second, the high-voltage transformers. Lead times for large power transformers are now measured in three to five years, against roughly one year at the start of the decade. A unit ordered today may not arrive before 2029 or 2030 in the most constrained categories. At the root of the problem sits a material almost no one outside the sector has heard of: grain-oriented electrical steel (GOES), used in transformer cores. Its production is concentrated in a handful of countries. The United States, for instance, has a single domestic manufacturer. When the input is scarce and concentrated, money cannot buy speed.
Third, the competition for equipment. Data centres are not the only project in the queue. Grid modernisation, renewable interconnection, industrial electrification and end-of-life equipment replacement are all fighting over the same transformers, breakers and switchgear. The result already shows in the numbers: market estimates suggest that 30% to 50% of the US data centre pipeline planned for 2026 could be delayed or cancelled for want of electrical infrastructure.
Sources: EPRI, “Powering Intelligence 2026”; analysis of electrical and grid-equipment infrastructure (2026).
How much will demand actually grow?
Enough to make this irreversible for the decade.
The International Energy Agency (IEA) projects 3.6% growth in global electricity demand in 2026 and estimates that the next five years will add, on average, 50% more demand per year than the previous decade did. This isn’t a spike. It is a new baseline.
The specific engine is the data centre. Also according to the IEA, data centre electricity consumption grew 17% in 2025, and consumption at AI-focused data centres surged 50%, far above the 3% growth in global electricity demand. In the United States, EPRI estimates data centres could account for 9% to 17% of national electricity consumption by 2030.
One figure gives the scale of the capital at play: infrastructure investment by the five largest technology companies passed US$400 billion in 2025 and is set to rise a further 75% in 2026. That money is looking for power. The question is who will supply it, on what timeline, and with what guarantee.
Sources: IEA, Electricity 2026 and Energy and AI (2026); EPRI, “Powering Intelligence 2026”.
Why won’t money solve the grid problem?
Because the bottleneck is physical and industrial, not financial.
Building a transformer plant, training a specialised workforce, expanding GOES production, permitting and erecting transmission lines: each of these runs on years. None of them accelerates simply because more capital is available.
That is why the most revealing move in the market isn’t a bigger cheque. It is a behaviour. The largest data centre developers have started building their own generation on site, with gas turbines and dedicated plants, to avoid waiting in the grid queue.
It is worth pausing on that signal. No one builds a power plant next to their own data centre when what is missing is money. They do it when what is missing is the connection. “Time to power” has stopped being a project detail and become a priced variable, often the most important one in the whole calculation.
Source: EPRI (2026); Uptime Institute (2026).
What does this change for an energy company?
Here is the pivot, and it is uncomfortable on purpose.
The bottleneck the entire market treats as the power sector’s problem is, read from another angle, the largest market the sector has seen in a generation. The scarcity holding the data centre back is exactly the scarcity the energy company exists to resolve: capacity, reliability and speed of connection.
There are two ways to respond. The first is to see the bottleneck as an obstacle: manage the queue, service requests, and wait for the investment cycle to run its course. The second is to see it as a market and build revenue on top of the scarcity itself.
The difference between the two isn’t ambition. It is the installed capability to turn that reading into a new offering. And that is what the next question is about.
Frequently asked questions
Is the AI bottleneck the chip or the power? It is no longer the chip. GPU supply scaled; electrical infrastructure did not. What now sets the pace of a data centre project is access to power: connection queues, transformer availability and transmission capacity.
What is a grid connection queue? It is the process of assessing, approving and building the infrastructure needed to connect a large new load to the grid. It involves capacity studies, transmission works and permitting, and in some regions it can take years, up to a decade, where the only route is the public grid.
Why is there a high-voltage transformer shortage? Because global demand surged simultaneously (data centres, renewables, electrification, replacement) while production is slow to scale and depends on an input concentrated in a few countries, grain-oriented electrical steel. The result is lead times of three to five years.
How much will data centre electricity demand grow? According to the IEA, data centre electricity consumption grew 17% in 2025, with a 50% rise at AI-focused data centres. In the US, EPRI projects they could reach 9% to 17% of national electricity consumption by 2030.
Why are data centres building their own power generation? To bypass the connection queue. When the grid cannot energise a project in time, developers build generation on site, usually gas turbines, so they can start operating. It is the clearest sign that the scarce resource isn’t money; it is connection.
The takeaway
The AI era was told as a race for chips. It has turned into a race for power. And the power sector, treated for years as a stable, regulated utility, finds itself at the centre of the decade’s biggest value shift.
This isn’t an automatic promise. Being in the right place is not the same as capturing the value. The difference will come down to how quickly each company turns its reading of this shift into market position, before the window closes.
The Bakery helps energy companies read this movement and position themselves while the game is still open.
