September 18, 2026
The Anthropic OpenAI AI data centers race is entering a new phase.
After committing to some of the world's largest AI infrastructure projects, both companies are now looking for something much smaller: data center capacity in the 20–30 megawatt range that can be brought online faster.
The shift reveals an important change in the artificial intelligence industry. The competition is no longer simply about securing the biggest possible data center. For companies building and deploying frontier AI models, getting the right computing capacity online at the right time is becoming just as important as securing massive long-term infrastructure deals.
Sources familiar with private discussions told CNBC that Anthropic has explored 20–30 MW opportunities in the U.K. and Nordic countries, while OpenAI has also investigated smaller deployments in the Nordics. Sources said discussions around similarly sized U.S. capacity have also taken place.
Anthropic OpenAI AI Data Centers Signal a New Compute Strategy
The move toward smaller facilities may look surprising considering the scale of the AI infrastructure investments announced by both companies.
But the two approaches can coexist.
Large data centers are designed to provide enormous computing capacity for the long term. Smaller facilities can offer a different advantage: speed.
If a 20–30 MW site can be secured and connected faster than a much larger campus, an AI company can deploy computing resources without waiting for a multi-year infrastructure project to reach full capacity.
For Anthropic and OpenAI, that matters because demand for AI models continues to expand across consumer products, software development and enterprise applications.
OpenAI told CNBC that it is building a diversified compute portfolio to meet growing global AI demand, with infrastructure decisions based on factors including performance, reliability, timing and cost.
Why AI Compute Is Becoming the Real Bottleneck
The AI boom depends on more than increasingly powerful models.
Every request made to an AI system requires computing resources. Training frontier models requires enormous amounts of compute, while serving those models to millions of users creates a separate and growing requirement known as AI inference.
That distinction is becoming increasingly important.
Training happens when a model is developed or significantly improved. Inference happens when the model actually responds to users and applications.
As AI adoption grows, inference capacity becomes a major infrastructure requirement.
Industry data cited by CNBC indicates that inference represented about 9% of global data-center workloads in 2025, compared with 14% for training. By 2030, inference is projected to rise to 37%, while training is projected at 13%.
That shift helps explain why smaller, distributed AI data centers are attracting attention.
Anthropic Is Building a Massive AI Infrastructure Network
Anthropic is not abandoning large-scale computing.
The company has already announced enormous infrastructure commitments to support Claude, its family of frontier AI models.
In April, Anthropic announced an agreement with Amazon that could secure up to 5 gigawatts of additional computing capacity for training and deploying Claude. The company said the agreement involved more than $100 billion in commitments to AWS technologies over the next decade.
Anthropic has also expanded its relationship with Google and Broadcom through an agreement covering multiple gigawatts of next-generation TPU capacity expected to come online from 2027.
The company said it uses a combination of AWS Trainium, Google TPUs and NVIDIA GPUs, giving it access to different types of AI computing hardware.
The new search for smaller data centers therefore appears to complement, rather than replace, those large infrastructure commitments.
OpenAI Is Also Expanding Its AI Infrastructure
OpenAI is pursuing a similarly broad strategy.
The company has been expanding its data-center footprint through Stargate and other infrastructure partnerships while developing a wider computing stack around its AI models and products.
OpenAI has described its compute strategy as an integrated system connecting data centers, chips, frontier models and AI products.
The company is therefore facing the same infrastructure question as Anthropic: how can it secure enough computing capacity to support rapidly growing AI workloads without relying exclusively on a handful of enormous facilities?
Smaller deployments could provide another layer of flexibility.
Why 20–30 MW AI Data Centers Matter
A 20–30 MW facility is tiny compared with the gigawatt-scale infrastructure projects now associated with frontier AI.
But that is precisely why the model could become useful.
Large AI campuses require substantial amounts of land, electricity, grid infrastructure, cooling systems and construction time. Smaller sites can potentially be deployed in locations where a massive AI campus would be difficult or take much longer to build.
The Wall Street Journal has also reported on a shift toward smaller AI data centers, with companies such as Crusoe investing in facilities intended to be faster and cheaper to build than some of the enormous projects currently under development in the United States. CNBC reported that Crusoe recently raised $3.9 billion at a $30.9 billion post-money valuation.
For AI companies, this creates a potential middle ground between cloud-scale infrastructure and waiting years for a new mega-campus.
AI Inference Could Change Where Data Centers Are Built
The growth of AI inference could ultimately change the geography of artificial intelligence infrastructure.
Training a frontier model can be concentrated in enormous computing clusters. Inference, however, happens continuously as users interact with AI applications.
That creates an incentive to place computing capacity closer to users, networks and available power.
For OpenAI, that could mean expanding the geographic diversity of infrastructure supporting ChatGPT and other AI products.
For Anthropic, it could mean adding more distributed capacity to support Claude AI as enterprise and consumer usage increases.
Anthropic has already announced major infrastructure expansions. Its Google and Broadcom agreement is intended to support future Claude models, while its AWS relationship provides another large-scale computing platform.
The Nvidia Connection
The race for AI data centers is also a race for AI chips.
NVIDIA remains one of the central suppliers of the GPUs used to train and run advanced AI models, while hyperscalers and AI companies are increasingly developing or accessing alternative accelerators.
Anthropic's infrastructure strategy illustrates this diversification particularly clearly: the company says it uses NVIDIA GPUs alongside Google's TPUs and Amazon's Trainium processors.
That means the AI infrastructure race is evolving across several layers at once:
AI models → AI chips → data centers → electricity → networking → cloud capacity → AI applications.
A constraint at any one of those layers can slow the deployment of new AI services.
Europe Is Becoming Part of the AI Compute Race
The fact that Anthropic and OpenAI are exploring smaller facilities in the U.K. and Nordic countries is also significant.
Europe has increasingly become a strategic market for AI infrastructure, but companies face different constraints around energy, planning, regulation and data centers.
Smaller facilities may provide AI companies with more flexibility when large-scale projects encounter lengthy development timelines.
The Nordic region in particular has attracted data-center investment because of its access to renewable electricity and cooler climate, factors that can be important for energy-intensive computing.
Anthropic's recent agreement for a 2.16-gigawatt data-center campus in Australia, for example, demonstrates how AI infrastructure decisions are increasingly being shaped by electricity availability, environmental considerations and local regulation as well as computing demand.ذ
Is the AI Data Center Boom Slowing Down?
The search for smaller facilities should not automatically be interpreted as a slowdown in AI investment.
The evidence points to something more nuanced.
Anthropic and OpenAI continue to pursue very large computing commitments while simultaneously exploring smaller deployments.
That suggests the issue may be how quickly and flexibly computing capacity can be deployed, rather than simply how much capacity AI companies want.
For the AI industry, speed is becoming an infrastructure advantage.
A smaller data center that can start running workloads significantly earlier may be more useful in the short term than a much larger facility that is still waiting for construction, grid connections or other approvals.
What the Anthropic and OpenAI Data Center Shift Means for AI
The latest development highlights a fundamental change in the economics of artificial intelligence.
The next stage of the AI race will not be determined only by who develops the smartest model.
It will also depend on who can secure enough AI compute, reliable electricity, advanced chips and data-center capacity to actually deliver those models to users.
Anthropic and OpenAI are now pursuing both sides of that equation: enormous long-term infrastructure commitments and smaller capacity deals that could help them deploy AI workloads more quickly.
That makes AI data centers one of the most important stories to watch as the industry moves deeper into the second half of 2026.
Why are Anthropic and OpenAI looking for smaller AI data centers?
Anthropic and OpenAI are exploring smaller 20–30 MW AI data-center deployments because they can potentially provide computing capacity faster than massive data-center projects. The smaller facilities would complement their existing gigawatt-scale AI infrastructure commitments.
Anthropic OpenAI AI data centers, Anthropic AI, OpenAI AI infrastructure, AI data centers 2026, AI compute, AI infrastructure 2026, OpenAI data centers, Anthropic data centers, Claude AI, OpenAI ChatGPT, AI inference, AI training, AI chips, NVIDIA AI, AI computing capacity, AI data center Europe, Nordic AI data centers, UK AI data centers, frontier AI models, artificial intelligence infrastructure
