We have written here extensively about the revenue side of the AI infrastructure buildout. My last piece, AI’s $600B question, focused on the implied revenue expectations for AI, and questioned the time horizon in which we’ll be able to meet those lofty goals.

This piece turns to the cost side of the equation. In particular, we will focus on the data center buildout, the rise of the “AI factory,” and its implications for energy, construction and the industrial supply chain. We believe that 2025 will be the “Year of the Data Center” and that we are on the cusp of transitioning from a hype cycle into an industrial-driven build cycle.

A few predictions, and then let’s get into the details:

1. AI will catalyze an energy transformation. New solar construction, battery innovation, a resurgence in nuclear energy—these will be long-term effects of the AI wave
2. Some hyperscalers will find that they are not nimble enough to address rapidly changing data center requirements—new industrial AI players will emerge to fill this gap
3. Starting in the next 6 months, there will be a lot of headlines about delays in data center builds due to issues with liquid cooling, cluster size and power access
4. The industrial capacity needed to build new AI data centers will serve as an economic stimulus and create jobs in the real economy: Steel, energy, trucking and construction
5. When new data center capacity comes online, the cost of training and inference delivered by AWS, Azure and GCP will go down, to the benefit of startups

In contrast to the big technical unknowns about the future of AI, there is a glimmer of clarity here: we can begin to plan for a 2-3 year period of industrial scaling. We are ready to move beyond abstract commitments (see below) to shovels hitting the ground.

Here is a summary of the new data center projects that have been announced in the last year— which we believe construction is now set to accelerate:

• Amazon: In the first half of 2024, AWS announced $50B in new data center projects, including 216 new buildings. Overall, Amazon has committed $100-150B over the next 15 years. Recent commitments include: An $11B campus in Indiana, $10B across two campuses in Mississippi, $5.3B in new data centers across Saudi Arabia, a new nuclear-powered data center near Salem, Pennsylvania, a new planned data center near Round Rock, Texas and a $15B commitment in Japan. Germany, Taiwan and Singapore are also being considered for new projects. 
• Microsoft: Microsoft now has 5GW of energy capacity and is reportedly doubling new data center construction in 2024. Recent announcements include: $3.3B in Mount Pleasant, Wisconsin, $1B in Northwest Indiana, $1B in Floyd County, Georgia, $4.3B in France, $3.5B in Germany, $3.2B in the UK, $3.2B in Sweden, $2.1B in Spain, $2.2B in Malaysia, $1.7B in Indonesia, $1B in Kenya and a new data center in Mexico. There have been reports of a $100B Stargate data center, although this is not confirmed.
• Google: Google is the smallest of the three cloud providers by a wide margin. The GCP pitch has long been that it’s better for AI companies. Now, that is being put to the test. Google is building a new $2B data center in Indiana, a $1B data center in Kansas City, Missouri, a $1.1B data center in Finland and a $576M data center in Cedar Rapids, Iowa. Google has the added challenge of scaling its own TPU clusters at some sites. 
• Meta: Meta does not operate a cloud business, but has nonetheless been scaling its data center capacity to support Llama and other internal AI initiatives. Meta recently announced that it has accumulated 350k H100 GPUs, as part of a total fleet of 600k H100 equivalents. The company also announced two 24k GPU clusters dedicated for Llama 3 training. Meta has four new data centers in the works, including in Kuna, Idaho, Temple, Texas, Davenport, Iowa and Cheyenne, Wyoming. 

The sheer industrial scale of what’s happening is tremendous. These announcements represent an industrial scale-up that happens once or twice in each generation.

It’s important to understand how unique the upcoming challenges will be, and how pressing. There is a long backlog of energy projects that need to be connected to the grid. Even with faster interconnection, we will need more power to support all these new data centers. Generation capacity will need to be added, largely in the form of solar and wind power, and we will need to be creative in leveraging existing energy resources. These power constraints are especially profound in “prime” data center markets like Virginia, Nevada and California. As a result, a lot of the newer developments are happening in “secondary” markets like Wyoming, Indiana, Iowa and Illinois. The need for more power and a better functioning grid was clear before AI. Now, it is becoming urgent.

Other technology issues abound: Next-gen Nvidia chips will require liquid cooling, and there is now a shortage in the liquid cooling supply chain. There is a two-year wait for diesel generators. Cluster sizes are reaching unprecedented territory: Elon Musk has announced a 300k GPU cluster. Models are getting so big, they may eventually need to be trained on distributed clusters across multiple data centers. Lithium-ion batteries have become a staple of new data center buildouts; new approaches are being contemplated to further lower cost and increase capacity.

Hyperscalers are known for their operational rigor in building data centers, but this new wave of construction will test even their best teams. Expect to see some distribution in outcomes across the hyperscalers—there will be winners and losers here. New industrial AI players will have an opportunity to fill any operational gaps. Existing market participants like Equinix, Digital Realty and CyrusOne are experiencing a “demand shock”—either they will step up and benefit from this, or they will lose share to new entrants. 

Expect to see a lot of headlines in 2025 about data center construction delays. Also expect to see some big, unexpected successes. When we exit the realm of bits and enter the realm of atoms, a new skillset takes hold and operational rigor becomes paramount. As we’ve seen with SpaceX and Tesla, the types of companies that thrive in these messy, fast-moving environments are not always the existing incumbents. Strong leadership and agility pay their biggest dividends during moments of complex change. 

The AI industrial phase should have a real economic stimulus effect, especially in areas of the economy that badly need it. Beneficiaries will include component makers in the industrial supply chain, energy companies building new generation assets like solar and wind, nuclear reactor operators and many others. For example, we are visiting a battery factory this week in West Virginia, which is being built on the site of an old coal plant. In the near-term, more job gains are likely to accrue to construction and industrial labor than to the small group of researchers architecting models in Silicon Valley.

How is all of this growth being financed? For the most part, Big Tech companies are deploying capital off their balance sheets. With the Magnificent 7 now representing approximately 30% of the S&P 500, the scale and speed at which they can deploy capital is awesome. Financial firms are participating as well, generating additional leverage. Many private equity firms are happy to provide upfront capital for construction and GPU purchases, in return for an IOU from Microsoft and a reasonable yield.

When all is said and done, we’re going to have a lot of AI factories. Whether there will be enough demand to fill them, we still don’t know yet. At the very least, training and inference costs should continue their decline, a boon to startups. 

If you are building in this space, we’d love to hear from you. Please reach out at dcahn@sequoiacap.com. We’re especially interested to hear from founders building at the intersection of energy, industrials and AI.