Gigawatt-scale AI campuses get all the headlines, but the real infrastructure scramble is happening at a much smaller scale. OpenAI and Anthropic are quietly hunting for localized data center deals ranging from 20 to 30 megawatts, shifting away from massive multi-hundred-megawatt hubs.
If you've been tracking the relentless energy crunch in tech, this pivot makes complete sense. Building a giga-scale facility takes years of zoning battles, grid connection approvals, and staggering capital expenditures. When you need usable compute capacity right now to serve millions of active users, waiting around for a massive campus to finish construction is a losing game. You might also find this similar story interesting: Why Smart Glasses Are Becoming A Nightmarish Privacy Problem.
Why Smaller Compute Deals Make Sense Now
For months, the tech press obsessed over multi-billion-dollar deals spanning entire regions. Announcements about massive gigawatt projects sounded impressive on earnings calls, but moving from a press release to an active, powered rack of GPUs is painfully slow.
Smaller footprints, particularly in the 20 to 30 megawatt range, offer a massive tactical advantage: speed. Existing facilities or brownfield sites with pre-existing power infrastructure can be brought online in a fraction of the time. Anthropic has explored these kinds of localized arrangements in the United Kingdom and the Nordic countries, while OpenAI has targeted similar international pockets alongside potential U.S. sites. As reported in detailed coverage by The Verge, the implications are widespread.
Let's look at the actual workload breakdown to see why this works. Training frontier models requires massive, centralized clusters of tens of thousands of accelerators tightly packed together with ultra-low latency interconnects. But running those models day-to-day—inference—is fundamentally different. Inference workloads can be distributed, modularized, and deployed closer to regional user bases.
The Shifting Balance Between Training and Inference
By 2027, industry projections show that inference workloads will surpass training in total data center capacity consumption. You don't need a massive nuclear-backed campus in rural America just to run standard user queries when you can slice smaller pods into regional European or domestic data hubs.
This isn't a replacement for massive infrastructure, but rather a diversification strategy. OpenAI acknowledged this dynamic, noting that they are actively building a diversified compute portfolio to handle global demand. When you evaluate partners based on timing, cost, reliability, and local grid constraints, smaller blocks of power become infinitely more attractive than waiting for a massive greenfield grid connection that might take five years to clear regulatory hurdles.
What This Means for the Rest of the Tech Ecosystem
The race for 20-to-30-megawatt parcels changes the competitive dynamics for regional providers. Smaller data center operators who previously struggled to compete with hyperscalers for massive enterprise contracts now find themselves holding the exact asset these AI giants desperately need.
We're moving past the era where bigger was automatically considered better. Agility, speed to power, and regional distribution are winning out over sheer architectural vanity. Expect more localized deals to drop as both companies look to squeeze every drop of efficiency out of an energy-constrained grid.
