The artificial intelligence industry’s enormous appetite for computing power is creating a new challenge for the companies building the world’s most advanced models: securing enough capacity quickly enough to keep up with demand.
Image Courtesy : skyracks.com
Anthropic and OpenAI have already committed to some of the largest computing and data-center projects in the technology industry, with plans involving hundreds of megawatts and, in some cases, gigawatts of capacity. But according to people familiar with the companies’ infrastructure discussions, both AI labs are now also exploring much smaller deployments in the range of roughly 20 to 30 megawatts.
The shift reflects a practical problem emerging alongside the massive expansion of AI. Huge data centers can provide enormous amounts of computing power, but they take years of planning, permitting, construction and power development. Smaller facilities can potentially be brought online faster, giving AI companies additional capacity while their largest infrastructure projects are still being developed.
Sources familiar with the discussions told CNBC that Anthropic has explored agreements involving 20- to 30-megawatt deployments in the United Kingdom and Nordic countries. OpenAI has also examined opportunities for deployments of a similar size in the Nordic region, while discussions involving potential U.S. capacity at that scale have also taken place. The sources requested anonymity because the commercial negotiations are private.
The strategy does not represent a retreat from massive AI infrastructure projects. Instead, it points toward a more diversified approach in which companies combine enormous long-term campuses with smaller pockets of capacity that can become operational sooner.
An OpenAI spokesperson told CNBC that the company is building a diversified computing portfolio to meet growing AI demand around the world. The company also said different workloads require different infrastructure and that it evaluates potential partnerships based on requirements such as performance, reliability, timing and cost.
That flexibility is becoming increasingly important as AI systems move from experimental research into widespread commercial use. Training increasingly sophisticated models requires enormous computing resources, but serving those models to millions of users can require a different infrastructure profile. AI agents, coding assistants, enterprise applications, reasoning systems and consumer products can all generate demand at different locations and at different times.
The result is a computing market where access to usable capacity can be just as important as the total amount of capacity a company has contracted.
OpenAI has been particularly aggressive in expanding its infrastructure footprint. In an April 2026 update, the company said its Stargate initiative had already surpassed its original commitment to secure 10 gigawatts of AI infrastructure in the United States by 2029, with more than 3 gigawatts added during the preceding 90 days. OpenAI said the expansion was being driven by accelerating demand from consumers, businesses, developers and governments.
The company's infrastructure strategy involves a broad network of partners spanning data centers, cloud providers, chip companies, energy providers, construction firms and investors. OpenAI has argued that no single company can build the infrastructure required for the emerging AI economy on its own.
That approach also helps explain why smaller facilities can be attractive. A 20- or 30-megawatt deployment is tiny compared with the enormous campuses being planned for frontier AI, but it can still represent a meaningful addition to available compute. If the facility is already powered or close to completion, it may provide a faster route to additional capacity than waiting for a multigigawatt project to be finished.
Anthropic has been pursuing an equally aggressive expansion. The company recently signed a reported $45 billion compute-capacity agreement with Nscale covering approximately 460 megawatts at a data-center development in West Virginia. The deal is expected to involve Nvidia's Vera Rubin systems, with capacity scheduled to come online in stages.
Anthropic has also reportedly entered into more than a dozen preliminary agreements with U.S. data-center developers as it explores taking greater control over the facilities that house its computing equipment. The Information reported that those letters of intent collectively represented more than 1 gigawatt of potential data-center capacity, although the agreements are nonbinding and some may not ultimately move forward.
The company is simultaneously relying heavily on traditional cloud and specialized AI infrastructure providers. Anthropic has committed to more than 10 gigawatts of server capacity through cloud arrangements, including a reported $200 billion agreement with Google over five years.
Another major Anthropic infrastructure deal illustrates how quickly the market is expanding. In September, the company reportedly agreed to a $35 billion cloud-capacity arrangement with Lambda involving 350 megawatts at a data center under development in Nueces County, Texas, near Corpus Christi. The campus is being developed by Hut 8 and is targeting an initial energization in 2027.
Taken together, these commitments demonstrate that the industry's infrastructure race is no longer simply about finding one enormous data-center site. AI companies are effectively assembling portfolios of computing resources across different providers, regions, power markets and facility sizes.
That portfolio approach can offer several advantages. A company dependent on one giant campus could face delays if construction, permitting, transmission upgrades or equipment deliveries fall behind schedule. Multiple facilities can provide more flexibility, although managing a geographically dispersed computing network also introduces additional operational complexity.
Power availability is one of the biggest constraints.
Modern AI data centers require vastly more electricity than conventional enterprise facilities, particularly when they are filled with dense racks of specialized accelerators. Securing land is therefore only one piece of the equation. Developers must also obtain access to sufficient electrical generation and transmission infrastructure, navigate permitting requirements and build cooling systems capable of handling the heat produced by high-performance computing equipment.
These challenges can make enormous campuses difficult to bring online quickly.
Smaller deployments may provide a way to work around some of those bottlenecks. A site with an existing power connection, available building space and established infrastructure may be able to host AI computing equipment much sooner than a greenfield facility requiring a completely new electrical and construction ecosystem.
That does not mean smaller data centers are simple or inexpensive. Twenty or thirty megawatts is still an enormous amount of power compared with an ordinary commercial facility. A site operating at that scale requires substantial electrical infrastructure, cooling, networking, security and specialized computing equipment.
The distinction is relative. In the world of frontier AI infrastructure, a 30-megawatt deployment is small compared with a 500-megawatt or multigigawatt campus, but it can still represent a major investment.
The move toward smaller facilities also reflects the changing economics of AI computing. Companies need infrastructure for different purposes, and not every workload requires access to the largest possible cluster.
Training a frontier model may require an enormous synchronized computing environment. Other workloads, including inference, application hosting, enterprise services and certain development tasks, may be distributed across smaller facilities. Being able to match the size and location of infrastructure to the workload could therefore become increasingly important.
For OpenAI and Anthropic, the urgency is particularly strong because demand for their products has expanded beyond individual consumers. Businesses are increasingly incorporating AI into software development, customer service, research, productivity tools and internal operations.
Anthropic has experienced particularly strong demand for its Claude models in software development and enterprise applications, while OpenAI operates a broad ecosystem spanning ChatGPT, developer APIs and enterprise products. As usage expands, infrastructure must grow alongside it.
This creates something of a feedback loop. More computing capacity allows AI companies to serve more users and deploy more capable systems. Greater adoption generates additional demand, which in turn requires still more infrastructure.
OpenAI described this dynamic directly in its infrastructure strategy, arguing that additional compute enables better models, greater usage and increased investment in infrastructure.
The companies are consequently competing on two infrastructure timelines at once. The first involves enormous projects designed to support AI demand years into the future. The second involves finding immediately usable capacity that can help meet demand while those larger projects are still under construction.
That second category could become increasingly competitive.
AI companies are not the only organizations searching for available data-center capacity. Cloud providers, neocloud companies, hyperscalers, enterprises and other AI developers are all competing for electricity, land, advanced chips and suitable facilities. Data-center developers can therefore find themselves negotiating with multiple potential customers for scarce power-connected sites.
The involvement of chip companies and financial institutions adds another dimension. Nvidia, Google, Broadcom and other technology companies have increasingly become connected to the financing and infrastructure arrangements supporting the AI buildout. Some deals involve chip commitments, financial guarantees or other mechanisms designed to make large data-center projects easier to finance.
For AI labs, those arrangements can help secure capacity without requiring them to finance and construct every component of the infrastructure themselves. But they also demonstrate how closely intertwined the AI model business has become with the data-center, semiconductor and energy industries.
The infrastructure race could ultimately reshape where data centers are built. Regions with abundant power, available land and favorable development conditions are increasingly attractive to AI companies and their infrastructure partners. Texas, the Midwest, the Southeast, the Nordic countries and other areas with access to substantial power resources have emerged as important locations for new projects.
The Nordic region is particularly interesting for smaller deployments because it has a long history of data-center development and access to substantial renewable and low-carbon electricity resources in several markets. For AI companies, geographic diversification can also help reduce dependence on a single power grid or infrastructure market.
At the same time, the expansion of AI infrastructure is creating new debates over electricity consumption, water use, construction, local economic benefits and the impact of large computing facilities on communities.
OpenAI has emphasized that its infrastructure projects should provide local economic benefits and has highlighted workforce development, community investment and infrastructure planning as part of its approach. Its Abilene, Texas, Stargate project has been presented by the company as an example of how large-scale AI infrastructure can be developed alongside local communities.
The broader trend suggests that the future AI data center may not be defined by a single standard size. Instead, the industry could develop a layered infrastructure model involving gigantic training campuses, medium-sized regional facilities and smaller deployments positioned close to users or specialized workloads.
For the AI companies themselves, that could provide greater flexibility. Large facilities can support the enormous computing requirements associated with training and future frontier models, while smaller installations can help deliver services, experiment with new hardware configurations and bring capacity online more quickly.
The shift also underscores how the AI boom is increasingly becoming an infrastructure story. The most advanced models may attract the headlines, but behind every AI chatbot, coding assistant and autonomous agent is a physical network of processors, data centers, electrical systems, cooling equipment and fiber connections.
As demand continues to grow, access to that physical infrastructure could become one of the industry's most important competitive constraints.
Anthropic and OpenAI's interest in smaller data-center deals does not signal that the era of massive AI campuses is ending. Instead, it shows how quickly the companies' computing requirements are evolving. The largest projects remain essential for long-term expansion, but smaller facilities can offer something equally valuable in the short term: usable computing capacity that can come online before the next generation of giant AI campuses is ready.
The result could be an increasingly fragmented but flexible global AI infrastructure network, with companies securing computing power wherever and whenever they can. In a market where demand is expanding faster than traditional data-center construction timelines, even a comparatively small facility can become an important piece of the race to keep AI systems running.