The Cost Side of AI: Data Centers and the Rise of the “AI Factory”
PLUS: Rippling-acquired founder building something new
Behind The Scenes
This section focuses on the costs associated with the AI boom, particularly the expansion of data centers, the emergence of the "AI factory," and its effects on energy, construction, and the industrial supply chain. We anticipate that 2025 will be the “Year of the Data Center,” marking a shift from hype to an industrial-driven build cycle.
Key Predictions:
AI will drive an energy transformation, spurring advances in solar construction, battery technology, and even a nuclear resurgence.
Some hyperscalers may struggle to adapt to rapidly changing data center needs, allowing new industrial AI players to step in.
Starting in the next six months, expect headlines about data center construction delays due to issues with liquid cooling, cluster size, and power access.
The industrial capacity required for new AI data centers will stimulate the economy, creating jobs in sectors like steel, energy, trucking, and construction.
As new data center capacity becomes available, the cost of training and inference through AWS, Azure, and GCP will decrease, benefiting startups.
A Look at Data Center Expansion
We are entering a 2-3 year period of industrial scaling, moving from abstract commitments to tangible construction. Here's a summary of recent data center projects set to accelerate:
Amazon: AWS announced $50B in new data center projects, with plans to spend $100-150B over the next 15 years. Major commitments include campuses in Indiana, Mississippi, Saudi Arabia, Pennsylvania, Texas, Japan, and potential sites in Germany, Taiwan, and Singapore.
Microsoft: With 5GW of energy capacity, Microsoft is doubling data center construction in 2024, including projects in Wisconsin, Indiana, Georgia, France, Germany, the UK, Sweden, Spain, Malaysia, Indonesia, Kenya, and Mexico.
Google: Google, the smallest of the three cloud providers, is testing its AI capabilities with new data centers in Indiana, Missouri, Finland, and Iowa, while scaling its TPU clusters.
Meta: Although not a cloud provider, Meta is expanding its data center capacity to support AI initiatives, including Llama 3 training. New data centers are planned in Idaho, Texas, Iowa, and Wyoming.
Upcoming Challenges and Opportunities
The scale of these projects is massive, and the challenges are significant. Energy constraints are becoming urgent, especially in prime markets like Virginia, Nevada, and California, pushing developments into secondary markets. Next-gen Nvidia chips require liquid cooling, and there are shortages in that supply chain. Diesel generators are backlogged for two years, and cluster sizes are reaching unprecedented levels, possibly requiring training across multiple data centers.
Hyperscalers are known for their operational excellence, but this new wave of construction will test them. Expect some hyperscalers to outperform, while others may falter, opening the door for new industrial AI players. Companies like Equinix, Digital Realty, and CyrusOne face a “demand shock” and must either step up or risk losing market share.
For full version of this article, you can read them here.
Geeks of the Week
Startup Name: Lightpage
Geography: US
One-liner: They are building an experience that’s a mix between a freeform notebook, a daily journal, and a wise friend.
Founder(s) Background: Employee #4 at Airtable and employee #1 at Watershed (Series C, $1.8bn valuation), Senior Engineering Manager (AI) at Airtable and Staff Software Engineer at Meta.
Startup Name: Aomni
Geography: US
One-liner: Aomni’s AI platform makes strategic B2B sales effortless by equipping reps with tailored account knowledge on demand.
Founder(s) Background: Co-Founder / CTO of Amity (Series C).