AI will drive more than 50% of global data center capacity and more than 70% of revenue opportunity, according to Omdia’s Research Director for Digital Infrastructure Vlad Galabov, who said massive productivity gains across industries driven by AI will fuel this growth. Speaking during Data Center World 2025’s analyst day, Galabov made a number of other predictions about the industry:
- NVIDIA and hyperscalers’ 1 MW-per-rack ambitions probably won’t materialize for another couple of years until engineering innovation catches up to power and cooling demands.
- By 2030, over 35 GW of data center power is expected to be self-generated, making off-grid and behind-the-meter solutions no longer optional for those looking to build new data centers, as many utilities struggle to deliver the necessary power.
- Data center annual capital expenditure (CAPEX) investments are expected to reach $1 trillion globally by 2030, up from less than $500 billion at the end of 2024.
- The strongest area for CAPEX is physical infrastructure, such as power and cooling, where spending is increasing at a rate of 18% per year.
“As compute densities and rack densities climb, the investment in physical infrastructure accelerates,” Galabov said. “We expect a consolidation of server count where a small number of scaled-up systems are preferred to a scaled-out server strategy. The cost per byte/compute cycle is also decreasing.”
Data center power capacity explodes
Galabov highlighted the explosion AI has caused in data center power needs. When the AI wave began in late 2023, the installed capacity of power in data centers worldwide was less than 150 GW. But with 120 kW rack designs on the immediate horizon, and 600 kW racks only about two years away, he forecasts nearly 400 GW of cumulative data center capacity by 2030. With new data center capacity additions approaching 50 GW per year by the end of the decade, it won’t be long before half a terawatt becomes the norm.
But not everyone will survive the wild west of the AI and DC market. Many startup DC campus developments and neoclouds will fail to build a long-term business model, as some lack the expertise and business savvy to survive. Don’t focus on a single provider, Galabov cautioned, as some are likely to fail.
More Data Center World 2025 coverage: NVIDIA’s Vision For AI Factories
AI drives liquid cooling innovation
Omdia’s Principal Analyst Shen Wang laid out the cooling repercussions of the AI wave. Air cooling hit its limit around 2022, he said. The consensus is that it can deliver up to 80 Watts per cm2, with a few suppliers claiming they can take air cooling higher.
Beyond that range, single-phase direct-to-chip (DtC) cooling — in which water or a fluid is taken to cold plates that sit directly on top of computer chips to remove heat — is needed. Single-phase DtC can go as high as 140 W/cm2.
“Single-phase DtC is the best way to cool chips right now,” Wang said. “By 2026, the threshold for single-phase DtC will be exceeded by the latest racks.” That’s when two-phase liquid cooling should begin to see a ramp-up in adoption rates. Two-phase cooling runs fluids at higher temperatures to the chip, causing them to turn to vapor as part of the cooling process, thereby increasing cooling efficiency.
“Advanced chips in the 600 watt and above range are seeing the heaviest adoption of liquid cooling,” Wang said. “By 2028, 80% of chips in that category will utilize liquid cooling, up from 50% today.”
