AI’s “bragawatts” era: can power grids keep up with hyperscale data centres?

AI infrastructure has entered what KKR’s digital infrastructure team recently called the “bragawatts” phase: hyperscalers and AI labs are no longer showing off model sizes and parameter counts so much as gigawatts of planned data centre capacity. Underneath the marketing, the numbers now attached to these projects raise a blunt question: even if only a fraction of them get built, can regional power grids and generation fleets actually support the load without breaking something important in the process?

From Meta’s Hyperion to OpenAI’s Stargate: the new AI build-out

The trigger for the latest round of number crunching was the financing package for Meta’s Hyperion data centre campus in Louisiana. The facility is designed as a flagship AI hub, with a power envelope measured in hundreds of megawatts. That alone would be notable, but it is only one of a growing list of hyperscale “AI factory” projects announced over the last 18 months.

OpenAI’s recent Michigan “Stargate” hub is a good example of how far expectations have shifted. The company has floated plans for a data centre cluster of more than 1 gigawatt in the state, on top of earlier Stargate announcements elsewhere. Taken together, those projects alone push past 8 gigawatts of potential AI data centre capacity, edging towards the 10 gigawatt target OpenAI talked about earlier in the year.

Barclays has tried to impose some order on this chaos by tracking new announcements. With the Michigan hub included, the bank tallies roughly 46 gigawatts of planned AI data centre capacity. Even if you assume a certain amount of headline inflation and eventual project cancellation, that is a large number.

The capex and power math: trillions of dollars, tens of gigawatts

On the capital side, Barclays estimates that building out the announced fleet of AI data centres would cost around $2.5tn. That is already a striking figure for an industry that, at the consolidated level, is still not consistently profitable. The energy side of the ledger is arguably more startling.

Barclays uses a simple but standard assumption for modern hyperscale design: a Power Use Effectiveness (PUE) of 1.2. That means for every 1 watt that reaches IT equipment, another 0.2 watts goes into cooling, power distribution losses, and ancillary services. Apply that ratio to 46 gigawatts of IT load and you arrive at roughly 55.2 gigawatts of total electrical demand if everything is built and operated at full tilt.

Translate that into everyday terms and the scale becomes more tangible. Using the bank’s rule of thumb that 1 gigawatt can supply electricity for over 800,000 US homes, 55.2 gigawatts would cover about 44.2mn households. That is on the order of three times the housing stock of California. Even if actual utilisation averages below nameplate capacity, the incremental demand is material relative to existing grid and generation footprints.

Where the power is supposed to come from

There are essentially three ways to feed a hyperscale AI campus: draw more from the existing grid, pay to expand and reinforce that grid, or integrate dedicated generation on or near the site. In practice, the current wave of projects combines all three.

Case study: OpenAI’s Michigan Stargate hub

For the Michigan Stargate site, the contracted supplier is DTE Energy. The utility has emphasised in earnings commentary that the new demand will not be allowed to raise costs for ordinary residential and small business customers. Instead, the data centre developer — in this case Related Companies — is expected to fund the incremental infrastructure needed to support the additional load.

DTE has also quietly raised its five year capital investment plan by several billion dollars. Part of that capex includes replacing one of its coal plants with gas turbines. That is not unique to Michigan. Across multiple AI data centre projects, the pattern repeats: the data centre pays for interconnection and local upgrades, and new generation comes from a mix of renewables, gas, and in some cases nuclear.

Embedded generation and bespoke deals

Meta’s Prometheus campus, another marquee AI build, has plans that include hundreds of megawatts of on site or dedicated generation from solar and gas turbines. In Pennsylvania, Amazon has lined up up to 1.9 gigawatts of nuclear power from Talen Energy to support its own data centre footprint.

On paper, this looks like an efficient matching of large, predictable industrial loads with large, capital intensive generation assets. In practice, most of these projects still depend on regional grid infrastructure that was not designed for swarms of new 500 megawatt class loads appearing within a few years of each other. Long transmission lead times, permitting bottlenecks, and local resistance to new lines can easily become the limiting factors.

Why AI centres are harder to integrate than classic data centres

It is tempting to treat AI data centres as just bigger versions of the cloud and web hosting facilities we already have. Nvidia and its partners have been stressing that this is the wrong mental model. In a recent report and associated research with Microsoft and OpenAI, they describe the AI “factory” as a synchronous, grid scale load rather than a collection of independent servers idling at slightly different times.

When training a large model, thousands of GPUs run intensive compute cycles in near lockstep, punctuated by barrier synchronisation and data exchange. The power profile of a single rack can swing from around 30 per cent load to near 100 per cent and back within milliseconds. Components have to be sized for the peaks, not the average. When that pattern is aggregated across an entire hall, you end up with a facility capable of ramping hundreds of megawatts up and down on very short time scales.

That volatility is not something legacy grids were built to handle gracefully. In joint work on power stabilisation for AI training centres, Nvidia, Microsoft, and OpenAI have shown that synchronous GPU workloads can introduce oscillations and instability at the grid level if they are not carefully buffered and coordinated. Grid operators are used to large industrial loads, but those loads typically move more slowly. An aluminium smelter does not usually swing its consumption by a factor of three in a few milliseconds.

Storage, buffering, and the “electron gap” narrative

One response is to add local energy storage at the data centre itself. Batteries or other storage technologies can absorb some of the high frequency fluctuations, presenting a smoother net profile to the grid. That is why many of the new AI projects, including the Michigan hub, are being designed with significant storage on site.

However, storage is a buffer, not a source. It can flatten peaks and provide short term resilience, but the average energy still has to come from somewhere. That is why OpenAI and others have started talking about very large increases in total generation capacity. OpenAI has reportedly urged US policymakers to aim for on the order of 100 gigawatts of additional power per year to support AI and associated electrification trends.

In that context, the company’s framing of an emerging “electron gap” between the US and China is notable. The phrase echoes the Cold War “missile gap” narrative, which later turned out to be substantially overstated. China has indeed added large amounts of new generation in recent years, but invoking historically loaded analogies can be a way to push for accelerated investment on compressed timelines rather than a neutral assessment of underlying demand and supply.

Which projects are real and which are theatre?

Even the banks trying to quantify the AI power build-out are careful to distinguish between announced capacity and capacity that will actually be built and energised. As Barclays has pointed out, keeping track of which projects are grounded in concrete grid interconnection agreements and shovel ready construction, and which are essentially branding exercises with a land option, is a difficult job.

There are several hard constraints that will shape what gets delivered:

Interconnection queues – in many regions, projects of any kind already wait years for approval to connect new large loads or generation to the grid.
Permitting and local politics – new lines, substations, and gas or nuclear plants inevitably encounter resistance and delay, even when they primarily serve industrial loads.
Financing conditions – the interest rate environment, counterparty risk, and perceptions of AI demand durability will all influence whether the more speculative data centre announcements secure funding.

It is therefore unlikely that all 46 gigawatts of announced AI data centre capacity will arrive on schedule, if at all. But the system does not need the full list to materialise to feel the impact. Even a partial build-out will reallocate a meaningful amount of capital and grid planning attention toward AI driven loads.

The possible upside: infrastructure that outlasts the hype

Viewed charitably, the current wave of AI data centre projects could be seen as a way to pull forward investment in generation, storage, and grid reinforcement that would have been needed eventually for electrification anyway. Large, creditworthy tech counterparties are often easier to finance against than fragmented residential demand. If the build-out is handled carefully, the long term result could be a more robust and cheaper grid for everyone, even if some of the more ambitious AI usage projections do not materialise.

The less charitable reading is that the sector is committing to multitrillion dollar infrastructure plans and tens of gigawatts of incremental demand based on business models that are still unproven at scale. In that scenario, regulators and utilities will be left to untangle the consequences if data centre operators or AI tenants change course midway.

Either way, the physics and economics will eventually impose themselves. GPUs can be fabbed faster than transmission lines can be permitted, and AI roadmaps can be redrawn more readily than nuclear plants can be financed. The long term shape of the AI industry will therefore be constrained as much by transformers, substations, and gas turbines as by model architectures and training tricks. The bragawatts era is ultimately a story about power engineering as much as it is about machine learning.