If you ever fly to Washington DC, look out of the window as you land at Dulles Airport – and you might snatch a glimpse of the single biggest story in economics right now.
There below you, you will see scattered around the fields and woods of the local area a set of vast warehouses that might to the untrained eye look like supermarkets or distribution centres. But no: these are in fact data centres – the biggest concentration of data centres anywhere in the world.
For this area surrounding Dulles Airport has more of these buildings, housing computer servers that do the calculations to train and run artificial intelligence (AI), than anywhere else. And since AI accounts for the vast majority of economic growth in the US so far this year, that makes this place an enormous deal.
Down at ground level you can see the hallmarks as you drive around what is known as “data centre alley”. There are enormous power lines everywhere – a reminder that running these plants is an incredibly energy-intensive task.
This tiny area alone, Loudoun County, consumes roughly 4.9 gigawatts of power – more than the entire consumption of Denmark. That number has already tripled in the past six years, and is due to be catapulted ever higher in the coming years.
Inside ‘data centre alley’
We know as much because we have gained rare access into the heart of “data centre alley”, into two sites run by Digital Realty, one of the biggest datacentre companies in the world. It runs servers that power nearly all the major AI and cloud services in the world. If you send a request to one of those models or search engines there’s a good chance you’ve unknowingly used their machines yourself.
Inside a site run by Digital Realty
Their Digital Dulles site, under construction right now, is due to consume up to a gigawatt in power all told, with six substations to help provide that power. Indeed, it consumes about the same amount of power as a large nuclear power plant.
Walking through the site, a series of large warehouses, some already equipped with rows and rows of backup generators, there to ensure the silicon chips whirring away inside never lose power, is a striking experience – a reminder of the physical underpinnings of the AI age. For all that this technology feels weightless, it has enormous physical demands. It entails the construction of these massive concrete buildings, each of which needs enormous amounts of power and water to keep the servers cool.
Sky’s Ed Conway at the data centre
Read more from Ed Conway:
Blow to UK’s bid to become minerals superpower
There’s more to your AirPods than meets the eye
We were given access inside one of the company’s existing server centres – behind multiple security cordons into rooms only accessible with fingerprint identification. And there we saw the infrastructure necessary to keep those AI chips running. We saw an Nvidia DGX H100 running away, in a server rack capable of sucking in more power than a small village. We saw the cooling pipes running in and out of the building, as well as the ones which feed coolant into the GPUs (graphic processing units) themselves.
Such things underline that to the extent that AI has brainpower, it is provided not out of thin air, but via very physical amenities and infrastructure. And the availability of that infrastructure is one of the main limiting factors for this economic boom in the coming years.
According to economist Jason Furman, once you subtract AI and related technologies, the US economy barely grew at all in the first half of this year. So much is riding on this. But there are some who question whether the US is going to be able to construct power plants quickly enough to fuel this boom.
Please use Chrome browser for a more accessible video player
Is Trump’s AI plan a ‘tech bro’ manifesto?
For years, American power consumption remained more or less flat. That has changed rapidly in the past couple of years. Now, AI companies have made grand promises about future computing power, but that depends on being able to plug those chips into the grid.
Last week the International Monetary Fund’s chief economist, Pierre-Olivier Gourinchas, warned AI could indeed be a financial bubble.
He said: “There are echoes in the current tech investment surge of the dot-com boom of the late 1990s. It was the internet then… it is AI now. We’re seeing surging valuations, booming investment and strong consumption on the back of solid capital gains. The risk is that with stronger investment and consumption, a tighter monetary policy will be needed to contain price pressures. This is what happened in the late 1990s.”
‘The terrifying thing is…’
For those inside the AI world, this also feels like uncharted territory.
Helen Toner, executive director of Georgetown’s Center for Security and Emerging Technology, and formerly on the OpenAI board, said: “The terrifying thing is: no one knows how much further AI is going to go, and no one really knows how much economic growth is going to come out of it.
“The trends have certainly been that the AI systems we are developing get more and more sophisticated over time, and I don’t see signs of that stopping. I think they’ll keep getting more advanced. But the question of how much productivity growth will that create? How will that compare to the absolutely gobsmacking investments that are being made today?”
Whether it’s a new industrial revolution or a bubble – or both – there’s no denying AI is a massive economic story with massive implications.
For energy. For materials. For jobs. We just don’t know how massive yet.
