For the AI industry, bigger has become almost synonymous with better. OpenAI and Anthropic have spent the past year lining up enormous amounts of data center capacity, with projects measured in hundreds of megawatts and even gigawatts.

Now, however, the two AI labs appear to be making room for something considerably smaller: data center deals in the range of just 20 to 30 megawatts.

That sounds tiny compared with the massive AI infrastructure projects grabbing headlines, but that’s exactly the point. According to people familiar with the discussions cited by CNBC, Anthropic and OpenAI are exploring smaller deployments in the U.K., Nordic countries and potentially the U.S. as they look for ways to get usable computing capacity online faster.

In other words, the AI data center race isn’t necessarily about finding the biggest possible building anymore. Sometimes it’s about finding one that can actually start doing something before everyone gets tired of waiting.

Sometimes smaller is faster

Anthropic has reportedly sounded out potential 20-to-30-megawatt agreements across the U.K. and the Nordics, according to four people familiar with the conversations. Two sources said OpenAI has explored similar opportunities in the Nordics, while another source was familiar with discussions involving both companies about U.S. capacity at roughly the same scale.

These are reported negotiations, not announced contracts. CNBC‘s sources did not identify specific counterparties, prices or delivery dates, and neither company has publicly confirmed a particular 20-to-30-megawatt deal.

OpenAI did confirm to CNBC that it is building what it described as a diversified compute portfolio. The company said different workloads require different infrastructure and that it evaluates potential opportunities based on requirements including performance, reliability, timing and cost. Anthropic did not comment on the reported discussions.

The attraction of smaller capacity is fairly straightforward: speed.

Jabez Tan, head of research at Structure Research, told CNBC that smaller deals can provide “speed to usable capacity.” Rather than waiting for an enormous new campus to secure power, complete construction and become operational, an AI company can potentially take capacity at an existing powered site and put it to work sooner.

And if the workload can be distributed across multiple locations, several smaller deployments can eventually add up to a substantial amount of computing capacity.

That’s particularly useful for AI inference, which is the computing required to actually serve AI models to users after those models have been trained.

The mega-projects aren’t going away

This isn’t OpenAI or Anthropic suddenly deciding that giant data centers were a terrible idea.

Both companies are still pursuing infrastructure deals on a completely different scale.

Anthropic, for example, has reportedly agreed to a roughly $45 billion arrangement with Nscale that involves about 460 megawatts of computing capacity at a data center development in West Virginia. OpenAI, meanwhile, has continued expanding its Stargate infrastructure plans, saying it surpassed its original 10-gigawatt target in April and later committed to another 3 gigawatts in Georgia and 8 gigawatts in Ohio.

So the strategy is becoming less “build one gigantic AI computer” and more “get compute wherever you can reasonably get it.”

There is a practical reason for that.

Large data centers take enormous amounts of land, electricity, transmission capacity, construction work and permitting. In the U.S., some projects have also faced opposition from local communities. European markets have their own constraints, including limited land and available power.

Waiting years for a massive campus isn’t particularly helpful when customers are already asking AI systems to do more today.

Inference changes the equation

The shift toward smaller facilities also makes more sense as AI workloads evolve.

Training a frontier AI model can require huge numbers of chips operating together, making large centralized computing facilities extremely valuable. Serving that trained model to users is different. Inference workloads can often be handled by smaller clusters that don’t necessarily need to sit inside one enormous campus.

That makes a distributed infrastructure strategy more practical.

Research cited by CNBC indicates that inference is expected to become an increasingly large share of data center capacity. JLL’s projections put inference at 9% of global data center workload capacity in 2025, compared with 14% for training, with inference expected to overtake training by 2027 and reach 37% by 2030.

That changes what “good” AI infrastructure looks like.

For training, having thousands of GPUs tightly connected in one location can be critical. For inference, having computing capacity closer to users and available in multiple locations can be useful for latency, resilience and simply getting enough capacity online.

And suddenly, that 20-megawatt data center doesn’t look quite so small.

AI’s infrastructure problem is becoming a race against time

The interesting part of OpenAI and Anthropic’s reported strategy is that it highlights a problem beneath the AI industry’s enormous infrastructure spending: securing enough money isn’t necessarily the same thing as securing usable computing capacity.

An AI lab can announce a multibillion-dollar infrastructure agreement, but the chips still need somewhere to run. The facility needs electricity. The electricity needs transmission infrastructure. The building needs to be constructed. Permits need to be approved. And, increasingly, companies need to deal with communities that may not be thrilled about having a giant power-hungry data center built nearby.

A smaller existing or faster-to-deploy facility can sidestep some of those bottlenecks.

So yes, OpenAI and Anthropic are reportedly looking at smaller data center deals. But this isn’t really a retreat from the giant AI infrastructure race.

It’s arguably the opposite.

They’re looking for more ways to get computing power into service faster — because when AI demand is growing this quickly, having 20 megawatts available now can be more useful than having 500 megawatts that is still waiting for a construction crew, a power connection and a very large stack of paperwork.