Anthropic and OpenAI Data Center Deals Are Getting Smaller to Bring AI Capacity Online Faster

Anthropic and OpenAI are reportedly pursuing 20 to 30 megawatt sites, despite committing to AI infrastructure measured in hundreds of megawatts or gigawatts. These smaller Anthropic OpenAI data center deals would give both labs usable computing capacity before their largest campuses are ready.

The development does not mean either company has abandoned its giant infrastructure projects. Instead, it reveals a growing mismatch between the pace of AI product demand and the time required to build large data centers.

Training increasingly capable models still rewards enormous, tightly connected computing clusters. Serving millions of independent requests, known as inference, creates a different requirement. Those workloads can often be distributed across smaller regional facilities, provided the operator can manage latency, reliability, security, and cost.

That distinction turns this reported search into more than another round of capacity procurement. Anthropic and OpenAI are assembling compute portfolios with different sites for different jobs. They are also accepting greater operational complexity to secure something the largest projects cannot immediately provide: speed.

The Reported Deals Target 20 to 30 Megawatts

The immediate change is a search for smaller powered sites that can begin serving AI workloads sooner.

According to smaller-site discussions reported by CNBC, Anthropic has explored potential agreements in the United Kingdom and Nordic countries. Four people familiar with those conversations described Anthropic’s activity to the publication.

Two sources said OpenAI had also examined opportunities in the Nordics. Another source knew of discussions involving both companies and possible deployments at a similar scale in the United States.

The talks reportedly concern facilities in the 20 to 30 MW range. That remains significant industrial infrastructure, but it is much smaller than the campuses dominating recent AI announcements.

A megawatt measures power, not computing performance. The useful output of a facility depends on its chips, networking, cooling, software, and actual utilization. Still, power capacity offers a common way to compare projects before their final hardware configurations become public.

Neither company announced a completed 20 to 30 MW agreement in the report. Anthropic declined to comment, while OpenAI said it evaluates infrastructure opportunities according to performance, reliability, timing, cost, and workload requirements.

That distinction matters. The evidence supports a reported procurement strategy, not a confirmed portfolio of new sites. Locations, providers, hardware, opening dates, and contract terms remain undisclosed.

OpenAI nevertheless described its broader objective clearly. A spokesperson told CNBC that the company was building a diversified compute portfolio to meet worldwide demand. That language places smaller facilities within its existing expansion strategy rather than presenting them as replacements for Stargate.

Anthropic has followed a similarly broad approach. It uses capacity from major cloud platforms and infrastructure specialists instead of relying on one site, one provider, or one chip architecture.

The reported geographic range also makes operational sense. Nordic markets can offer cooler climates, established energy systems, and experienced data center operators. The United Kingdom places capacity near a large commercial market, while additional American sites can support domestic traffic and regulatory needs.

Those advantages do not guarantee fast deployment. Available grid power, transmission capacity, planning approval, equipment deliveries, and local support still determine whether a nominal site becomes usable.

The key phrase is usable capacity. A large planned campus has limited value to an AI product team until power, cooling, networking, and accelerator systems work together under production conditions.

Structure Research analyst Jabez Tan summarized the attraction as “speed to usable capacity.” He told CNBC that a few megawatts at an already powered location can be more practical than waiting for a much larger allocation.

Several smaller blocks can also accumulate into meaningful capacity. That approach works best when the underlying workloads do not require every accelerator to participate in one tightly synchronized computation.

This is why the scale of the reported talks points directly toward inference. It also explains why smaller projects are emerging alongside, rather than instead of, enormous training campuses.

Anthropic OpenAI Data Center Deals Now Serve Two Timelines

Large campuses address long-term model ambitions, while smaller deployments address the near-term pressure of serving products reliably.

Anthropic and OpenAI continue to pursue infrastructure at a radically larger scale. Their public plans include multiyear projects whose capacity will arrive in phases.

OpenAI said in April 2026 that it had surpassed Stargate’s original goal of securing 10 GW of American AI infrastructure by 2029. It also said more than 3 GW had been added to its plans during the preceding 90 days.

The company’s Stargate infrastructure plan describes compute as the input connecting model development, product usage, revenue, and further infrastructure investment. OpenAI acknowledges that financing structures can change, but emphasizes capacity arriving on time.

Anthropic has made comparable commitments through several partners. Its expanded Amazon compute agreement covers up to 5 GW for training and deploying Claude.

Anthropic said that agreement included capacity in Asia and Europe for international inference. It also expected nearly 1 GW of combined Trainium2 and Trainium3 capacity to be available by the end of 2026.

Another disclosed project shows the longer development cycle. TeraWulf announced a 20-year Anthropic lease covering approximately 401 MW of critical information technology load in Hawesville, Kentucky.

According to the Kentucky lease filing, initial capacity is expected during the second half of 2027. The facility is scheduled to reach its full planned load by early 2028.

That timetable illustrates the central pressure behind smaller deals. AI companies cannot make current product reliability depend entirely on campuses scheduled for future years.

Users experience capacity constraints through slower responses, service limits, unavailable features, or inconsistent performance. Developers experience them through rate limits and uncertainty about whether a model will support production traffic.

Enterprise customers add another layer. They expect contractual reliability, stable latency, predictable regional availability, and compliance with data residency requirements. Those demands often favor multiple regional deployments over one distant computing complex.

The two timelines therefore solve different problems. Giant campuses support frontier training and exceptionally large pools of future demand. Smaller sites can absorb regional inference, specialized workloads, or overflow while those campuses are built.

This division also reduces dependence on a single construction schedule. A permitting delay or transmission problem at one large location does not stop every planned unit of capacity from arriving.

However, diversification does not eliminate infrastructure risk. It changes the shape of that risk from concentration to coordination.

A company using many facilities must maintain compatible model versions, security controls, monitoring systems, and networking policies. It must route traffic around failures without creating inconsistent user experiences.

Hardware diversity creates another challenge. A model optimized for one accelerator platform may require engineering work before it runs efficiently on another. Performance comparisons also become harder when clusters use different chips, memory systems, and network designs.

Anthropic’s agreements show why that work has become strategic. Its capacity spans Amazon chips and infrastructure available through other major cloud platforms. OpenAI likewise works with a widening group of cloud, chip, financing, and construction partners.

The result is not a simple competition over who announces the largest number. The more useful comparison concerns how much capacity becomes operational, what workloads it supports, and how efficiently each lab uses it.

That places delivery teams under pressure. They must convert a portfolio of contracts and construction projects into dependable computing resources before product demand outruns supply.

Inference Changes the Ideal Data Center Map

The smaller-site strategy works because inference can often be divided across locations more easily than frontier model training.

Training a large model requires many accelerators to exchange data rapidly while updating a shared set of parameters. Slow communication between machines can leave expensive chips waiting instead of calculating.

That requirement rewards dense clusters with high-bandwidth networking and carefully coordinated systems. Splitting one training run across remote locations can introduce latency that undermines performance.

Inference begins after training. It is the computation performed when a deployed model processes a prompt, writes code, analyzes a document, or completes another user request.

Many inference requests are independent. A Claude user in London does not usually need the same physical cluster serving a ChatGPT user in California at that moment.

Traffic can therefore be routed among several sites, assuming each location has the correct model, sufficient memory, and suitable security controls. This makes regional capacity more useful as product adoption grows.

The workload mix is also changing. CNBC cited a JLL projection that inference would overtake training as the leading AI workload during 2027.

The same forecast put inference at 9 percent of global data center workloads in 2025. It projected that share to reach 37 percent by 2030, compared with 13 percent for training.

Those figures are forecasts rather than guaranteed outcomes. They depend on sustained product adoption, viable AI economics, and enough electricity and hardware to meet demand.

Still, the direction helps explain why Anthropic and OpenAI would want smaller sites before their largest campuses open. Training remains concentrated, while everyday usage produces a broad and recurring stream of inference requests.

Agents intensify that pattern. A conventional chatbot exchange might require one response. An agent can make repeated model calls while searching documents, writing code, checking results, and revising its work.

One user action can therefore produce a chain of inference operations. Longer reasoning, multimodal inputs, and persistent tool use can increase the computation behind each visible response.

Regional sites can also improve responsiveness. Physical distance is not the only source of latency, but a closer serving location can reduce network travel and provide another route during congestion.

Data residency can influence placement as well. Some organizations need particular information processed or stored within approved jurisdictions. Regional infrastructure gives model providers more options for meeting those requirements.

The model itself does not need to be developed in every region. Providers can train a model in a large centralized cluster, distribute approved model versions, and serve them from multiple locations.

This architecture resembles content delivery only at a high level. AI inference needs substantial accelerator capacity, specialized networking, careful scheduling, and controls for sensitive input data.

Smaller does not mean easy. A 20 MW AI facility requires serious power delivery, cooling equipment, servers, network connections, and technical staff. It also has to operate at high utilization to justify the resources committed to it.

Fragmentation can lower utilization if demand varies sharply by region. One location might face a traffic spike while another has idle accelerators that cannot accept the same requests because of policy or network constraints.

Operators can mitigate that problem with workload scheduling and geographic traffic management. Yet every additional rule reduces the apparent simplicity of distributing inference.

Model updates create similar complications. Teams must roll out new weights and serving software without producing incompatible behavior across regions. They also need rapid rollback procedures when a release affects quality or reliability.

Observability becomes essential. Providers need to know whether a slow response comes from the model, the serving software, the network, or the underlying hardware.

For enterprise buyers, this infrastructure design can influence service quality more directly than another distant capacity announcement. A nearby, operational inference site can matter more than a larger project arriving years later.

For developers, the most useful signals will be API reliability, latency consistency, rate-limit changes, and regional availability. Those measurements reveal whether reported capacity is actually improving the products built on top of it.

The Real Contest Is Speed Versus Concentration

The main tradeoff is no longer small data centers against large ones, but faster deployment against the efficiency of concentrated infrastructure.

Large campuses offer economies that are difficult to reproduce across scattered sites. A single development can centralize power procurement, cooling, physical security, network design, and maintenance.

Concentrated infrastructure also suits the demanding communication patterns of frontier training. Engineers can design the entire facility around a defined hardware architecture and optimize it as one system.

The disadvantage is execution time. Large projects require suitable land, substantial utility commitments, transmission upgrades, permits, construction labor, cooling systems, and enormous quantities of electrical equipment.

Any delayed component can postpone the full campus. Transformers, substations, turbines, chips, and networking equipment follow different supply chains, creating several possible bottlenecks.

Community resistance adds another variable. Residents and public officials increasingly question how large data centers affect electricity bills, water consumption, noise, land use, and local tax arrangements.

OpenAI has tried to address those concerns through community commitments and public descriptions of its cooling systems. However, each project still faces conditions that vary by location.

Smaller existing sites can bypass part of that process when power and basic infrastructure are already available. They can also use modular construction that standardizes more of the facility before installation.

Crusoe, which developed infrastructure used by OpenAI in Abilene, is pursuing that idea. The company is investing in smaller facilities that can be manufactured in factories and transported to locations with available power.

The modular data centers are intended to deploy faster than projects relying solely on large conventional builds. The strategy provides an industry comparison beyond Anthropic and OpenAI.

This does not make small facilities universally superior. The approach transfers work from construction scheduling into fleet management.

An AI lab must coordinate capacity controlled by multiple operators and utilities. It may also negotiate different service obligations, hardware refresh cycles, and security processes at every location.

Costs can become harder to compare. A facility available immediately may carry different commercial terms from capacity reserved years in advance.

The labs must also decide which workloads deserve scarce near-term capacity. Consumer chat, coding agents, enterprise APIs, model evaluations, and research jobs can compete for the same accelerators.

Allocating work poorly can erase the benefit of faster deployment. A newly available cluster does not help much if software incompatibility or insufficient networking prevents efficient use.

This is why the word “portfolio” matters. The labs are not simply collecting megawatts. They are matching workload characteristics against delivery dates, locations, chip types, and reliability requirements.

Anthropic’s position illustrates the balance. Its planned capacity includes very large cloud commitments, a major dedicated site in Kentucky, and reportedly smaller regional discussions.

OpenAI’s position is broader in stated scale, but structurally similar. Stargate provides the long-term framework, while a diversified set of partners and locations can address nearer-term requirements.

The competition between the companies therefore turns partly on execution. Announced capacity can support a strategic narrative, but operational capacity supports actual products.

A lab that deploys smaller clusters quickly can improve availability while rivals wait for larger sites. A lab that fragments too aggressively can create an expensive and difficult infrastructure estate.

The winning design will probably remain mixed. Large campuses will handle workloads that benefit from extreme concentration, while smaller clusters will serve demand that values location and timing.

That conclusion is less dramatic than a complete infrastructure reversal. It is also more consequential because it changes how the industry measures progress.

The largest number in a press release tells only part of the story. Delivery dates, workload placement, utilization, latency, and reliability determine whether capacity becomes a competitive advantage.

Smaller Sites Do Not Remove the Power Constraint

A smaller request can be easier to place, but it still competes for electricity, equipment, and community acceptance.

The reported 20 to 30 MW range is modest only when compared with a gigawatt campus. Against most commercial buildings, it represents a substantial and continuous electrical load.

Several such facilities can also recreate the aggregate demand of a much larger project. Distribution changes geography and scheduling, but it does not make the underlying energy requirement disappear.

Existing powered sites are especially valuable because new grid connections can take years. That scarcity can increase competition among AI labs, cloud providers, industrial users, and other data center customers.

The best locations need more than electricity. They require dependable fiber connectivity, suitable cooling, physical security, local technical expertise, and access to the required accelerator hardware.

Nordic markets offer several attractive characteristics, but they are not unlimited reservoirs of capacity. Grid constraints, industrial policy, weather conditions, and local regulation vary across countries and regions.

Cross-border operations create legal questions as well. Providers must manage data protection, customer contracts, export rules, and security requirements for each jurisdiction.

The environmental picture also resists simple claims. A region with lower-carbon electricity can still face transmission limits or competing demands from households and industry.

Water use depends on the cooling design and climate. Closed-loop systems can reduce ongoing consumption, but they do not remove the broader resource footprint of construction and electricity generation.

The reported discussions also leave a commercial uncertainty. Neither lab has disclosed which providers would own the facilities, supply the chips, or bear construction and utilization risk.

Long-term commitments protect operators from uncertain demand. They can also leave buyers exposed if models become more efficient or hardware improves faster than expected.

Efficiency gains will not automatically reduce total power demand. Lower computing cost can encourage more usage, more complex requests, and new products that were previously uneconomic.

The opposite risk also exists. If enterprise adoption grows more slowly than expected, some reserved capacity could remain underused.

Inference forecasts therefore deserve caution. Projected workload shares describe one plausible industry path, not a verified schedule for demand.

Reliability poses another test. Multiple sites can protect against one location failing, but only if applications can move between them without violating customer requirements.

Regional capacity must also receive model updates securely. A wider physical footprint creates more operational boundaries and more systems that need monitoring.

Cybersecurity responsibilities can become divided among the AI lab, the cloud provider, the data center operator, and connectivity vendors. Clear ownership matters when an incident requires rapid action.

Hardware availability remains a separate bottleneck. A powered building without suitable accelerators is not usable AI capacity, just as delivered chips without power cannot support a production service.

These constraints explain why OpenAI listed performance, reliability, timing, and cost together. Optimizing one dimension can worsen another.

A readily available site might use less familiar hardware. A highly efficient cluster might be located too far from target users. A preferred region might lack power during the required period.

Anthropic and OpenAI have enough demand to pursue several options simultaneously. Smaller AI companies generally have less bargaining power and fewer ways to absorb deployment delays.

That imbalance can strengthen large labs even when both face infrastructure constraints. They can reserve capacity across providers, while smaller customers depend on what remains available through cloud services.

The strategy could therefore deepen infrastructure concentration at the buyer level. More sites do not necessarily mean a more competitive market if the same few AI companies contract most suitable capacity.

Regulators and enterprise customers should watch that distinction. Geographic distribution and market diversity are not the same thing.

Three Signals Will Show Whether the Strategy Works

The next evidence should come from operational performance, disclosed delivery milestones, and continued investment in distributed inference.

The first signal is whether Anthropic or OpenAI confirms smaller deployments and identifies when they will become operational. Signed capacity, installed hardware, and production traffic represent three different milestones.

A public agreement without an opening schedule would add little certainty. A site serving customer requests would support the argument that smaller projects can bridge the gap before giant campuses arrive.

The second signal is product performance. Developers should watch for more stable API latency, expanded regional availability, fewer capacity-related limits, and improved reliability during peak demand.

Those outcomes would show that additional infrastructure is reaching products rather than remaining a planning figure. Persistent service constraints would suggest that capacity is arriving too slowly or demand is consuming it immediately.

The third signal is whether infrastructure providers continue building standardized smaller facilities. Crusoe’s modular effort offers an early comparison, while other operators can validate or weaken the model through their own deployments.

A growing supply of ready-to-use regional clusters would make portfolio procurement easier. Limited follow-through would indicate that power, equipment, and economics still favor fewer large developments.

The largest campus announcements remain important. Frontier training needs concentrated computing resources, and both companies expect AI demand to keep rising.

Yet the smaller-site discussions reveal a more immediate contest. Anthropic and OpenAI need infrastructure that works on a product schedule, not only a construction schedule.

For developers and enterprise buyers, the practical question is straightforward: do Claude and OpenAI services become more available, predictable, and geographically flexible?

Track those outcomes alongside megawatt announcements. The Anthropic OpenAI data center deals that matter most will be the ones that turn power, chips, and buildings into reliable AI capacity before demand moves again.