Vlad Galabov, Omdia’s research director for digital infrastructure, spoke at Data Center World 2025’s analyst day. He highlighted how AI will account for over 50% of global data center capacity and more than 70% of revenue opportunities. The surge in productivity thanks to AI across various sectors is driving this change.
Galabov made several key predictions during his talk:
– The push for 1 MW-per-rack by NVIDIA and hyperscalers likely won’t happen for another couple of years, as we wait for engineering advancements to meet the rising power and cooling demands.
– By 2030, more than 35 GW of data center power is set to be self-generated, making off-grid and behind-the-meter solutions essential for anyone looking to create new data centers. Many utilities can’t keep up with demand.
– Annual capital expenditures (CAPEX) for data centers are expected to hit $1 trillion globally by 2030, up from less than $500 billion at the end of 2024.
– The biggest focus for investments is in physical infrastructure, including power and cooling, with spending growing at 18% annually.
Galabov pointed out that as compute and rack densities increase, so does the need for investing in physical infrastructure. He anticipates a trend toward fewer, larger servers rather than many smaller ones. The cost per byte and compute cycle is also on the decline.
He also addressed the rapid increase in power needs due to AI. When AI gained traction in late 2023, global data centers had an installed power capacity of less than 150 GW. With upcoming designs for 120 kW racks, and 600 kW racks expected in about two years, he estimates data center capacity will approach 400 GW by 2030. New data center additions could reach 50 GW annually by the end of the decade, moving closer to half a terawatt.
However, he warned that not every new venture in the AI and data center space will succeed. Many startups may struggle to establish viable long-term business models, lacking the necessary expertise to thrive. His advice? Don’t place too much reliance on a single provider; some won’t make it.
On the cooling front, Omdia’s Principal Analyst Shen Wang explained how AI is influencing cooling technologies. He noted that air cooling reached its limit around 2022, capable of handling about 80 Watts per cm². To go beyond that, we need single-phase direct-to-chip (DtC) cooling, where fluid goes directly to cold plates on chips for heat removal—this can handle up to 140 W/cm².
Wang emphasized that single-phase DtC is currently the best cooling solution, and by 2026, the latest racks will exceed its performance threshold. Two-phase liquid cooling, which uses vaporization for greater efficiency, is set to gain traction soon after.
“Advanced chips above 600 watts are increasingly adopting liquid cooling,” Wang noted. By 2028, he predicts that 80% of these chips will use liquid cooling, up from 50% today.
