Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
The financing package stitched together for Meta’s humongous Hyperion data centre campus in Louisiana made Alphaville curious about just how much energy the new AI infrastructure will consume if it all comes online.
After all, massive new projects are being announced almost every week, in what even KKR’s digital infrastructure lead called a “bragawatts” phenomenon in MainFT on Monday.
The latest example is OpenAI on Thursday revealing plans for a 1+ gigawatt data centre hub in Michigan. Together with previously announced “Stargate” projects this brings the total to over 8 gigawatts — close to the 10 target it floated earlier this year. This will cost over $450bn over the next three years, according to the company that spends more on marketing and employee stock options than it makes in revenue.
So how many data centre projects have now been started or announced? Which ones will actually happen and which ones are fantasy? As Barclays noted last week, tracking “what is real vs. speculative is a full-time job”, but the bank has forced some poor sell-side plebs to at least tally all the announcements and collect some rudimentary details.
So what is the total so far? With OpenAI’s Michigan project they now total 46 gigawatts of computing power. Apologies for the virtual shouting, but this seems a bit mad.
These centres will cost $2.5tn to build, according to Barclays, to service an industry that still doesn’t turn a profit. But the maddest bit arguably is how much energy they will require once completed. Using Barclays’ 1.2 “Power Use Effectiveness” ratio, all these data centres — if they are all completed — would need 55.2 gigawatts of electricity to function at full capacity.
If we also use Barclays’ rule of thumb that 1 gigawatt can power over 800,000 American homes, it means that these data centres will consume as much energy as 44.2mn households — almost three times California’s entire housing stock.*
So where is the energy all coming from? Well, OpenAI’s Michigan Stargate is perhaps instructive.
All the electricity for Michigan’s Stargate site will come from DTE Energy, which stressed in its third-quarter earnings last week that it wouldn’t negatively affect ordinary consumers, and said that Developer Related Companies — which is actually building the campus — would cover the costs of the new power infrastructure needed:
The new electricity demand will be supported with existing capacity and new energy storage investments which will be paid for by the data center. This data center growth will create substantial affordability benefits for existing customers as DTE sells excess generation, and contract terms will also ensure that the data center absorbs all new costs required to serve them.
However, DTE increased its five-year investment plan by $6.5bn, which Barclays noted included replacing one of its coal plants with gas turbines. And this seems to be the emerging norm.
In many other cases, the data centre business plans include the installation of at least some of the energy generation. For example, Meta’s Prometheus campus includes plans for 516 megawatts from solar and gas turbines. Amazon’s Pennsylvania data centres have been promised 1.9 gigawatts from Talen Energy’s nuclear power plant.
However, in many cases the regional power grids still seem hopelessly inadequate to cope with demand possibly surging over the next few years.
Beyond the sheer scale, the nature of the AI-related power demand is also particularly problematic. As a recent Nvidia report noted (with Alphaville’s emphasis below):
Unlike a traditional data center running thousands of uncorrelated tasks, an AI factory operates as a single, synchronous system. When training a large language model (LLM), thousands of GPUs execute cycles of intense computation, followed by periods of data exchange, in near-perfect unison. This creates a facility-wide power profile characterized by massive and rapid load swings.
This volatility challenge has been documented in joint research by NVIDIA, Microsoft, and OpenAI on power stabilization for AI training data centers. The research shows how synchronized GPU workloads can cause grid-scale oscillations.
The power draw of a rack can swing from an “idle” state of around 30% to 100% utilization and back again in milliseconds. This forces engineers to oversize components for handling the peak current, not the average, driving up costs and footprint. When aggregated across an entire data hall, these volatile swings — representing hundreds of megawatts ramping up and down in seconds — pose a significant threat to the stability of the utility grid, making grid interconnection a primary bottleneck for AI scaling.
That’s presumably why the likes of the Michigan Stargate project will include lots of energy storage. But it also explains why OpenAI this summer asked the Trump administration to ensure that the US brings a massive 100 gigawatts a year online to feed the gaping AI maw.
So how likely is this to all pan out the way the AI visionaries et al envisage it?
Well, OpenAI weirdly warned that the US now faced an “electron gap” versus China — a nod to Cold War “missile gap” that famously proved to be complete and utter hogwash. Even at the time, the CIA knew it was baloney. Sure, China might have built a lot power generation lately, but that OpenAI would invoke such an well-known mirage is . . . curious.
However, once the hype dies down at least we might be left with a bigger, stronger and cheaper energy grid in its wake?
*Or 46 time jumps with the DeLorean.
