The AI Data Center Race Is Changing: Why OpenAI and Anthropic Want Smaller 20–30 MW Sites

2 minutes ago

9 min read

The artificial intelligence infrastructure race is entering a new phase. OpenAI and Anthropic have committed to enormous amounts of computing capacity, with projects measured in gigawatts and investments extending across multiple cloud and data center partners. Yet reports that both companies are exploring smaller data center deployments of roughly 20 to 30 megawatts reveal an important shift in how AI computing may be built and delivered.

The move does not signal the end of giant AI campuses. Instead, it highlights a fundamental distinction between AI model training and AI inference, the process through which trained models generate responses for users. Training increasingly capable models requires enormous, tightly interconnected clusters. Inference can often be distributed across multiple locations, creating a new infrastructure model in which large training facilities coexist with smaller regional computing sites.

For the AI industry, this distinction could become strategically important as demand moves from experimentation toward continuous commercial usage.

Why OpenAI and Anthropic Need More Than Gigawatt-Scale Projects

The scale of modern AI infrastructure is difficult to compare with conventional cloud expansion. Training advanced models requires thousands of accelerators operating together, supported by high-speed networking, sophisticated cooling systems, large electrical connections and substantial data center capacity.

OpenAI has pursued infrastructure ambitions measured in multiple gigawatts, while Anthropic has entered major agreements designed to secure substantial long-term computing capacity. These commitments reflect expectations that AI usage will continue expanding across consumer applications, enterprise software, coding, research, automation and other workloads.

But a future capacity commitment does not necessarily solve an immediate computing requirement.

A large data center campus can take years to progress through site selection, power procurement, transmission development, permitting, construction, equipment deployment and operational commissioning. The larger the project, the greater the number of dependencies that must align before computing resources become available.

Smaller facilities can address a different problem: time-to-compute.

A 20 MW or 30 MW facility may represent only a small fraction of a multigigawatt infrastructure portfolio, but it can become useful much earlier if an existing powered site is available and suitable for AI workloads.

That makes smaller facilities strategically valuable even for companies simultaneously pursuing some of the world's largest computing projects.

The Technical Difference Between Training and Inference

The economics of distributed AI infrastructure become clearer when training and inference are separated.

AI training

Training involves adjusting the parameters of an AI model through enormous amounts of computation. Advanced models require large groups of accelerators to communicate rapidly with one another.

This creates demanding requirements for:

High-bandwidth accelerator interconnects
Low-latency networking
Large-scale power delivery
Advanced cooling
High-density computing environments
Massive data storage and movement
Reliable infrastructure operating continuously

Because the workload is highly interconnected, concentrating computing resources in a large facility can provide substantial technical advantages.

AI inference

Inference occurs after a model has been trained. Every time a user asks a question, generates code, creates an image or interacts with an AI application, computing resources are consumed.

Unlike some training workloads, many inference requests are relatively independent. A large population of users can therefore be served by multiple geographically distributed clusters.

This creates an infrastructure architecture that resembles a network rather than a single computational megastructure.

Infrastructure requirement	Model training	AI inference
Primary objective	Build or improve models	Serve model requests
Compute pattern	Highly coordinated	Often distributable
Geographic flexibility	More limited	Generally greater
Latency sensitivity	Important, but workload-dependent	Often critical for interactive applications
Facility strategy	Large concentrated clusters	Large clusters plus distributed sites
Key constraint	Compute density and interconnect	Capacity, power, latency and availability

The distinction helps explain why smaller data centers can become increasingly important even while frontier AI companies continue building enormous campuses.

Inference Could Become the Next Major Data Center Growth Engine

For years, discussions about AI infrastructure centered primarily on training. That made sense when the industry was focused on developing increasingly capable foundation models.

The economics change once those models become widely used.

A model that serves millions of users generates recurring computational demand. Every interaction consumes inference capacity. Coding assistants, enterprise copilots, autonomous software agents, search systems, image generators and AI-powered applications can all generate sustained workloads after the underlying model has already been trained.

This means infrastructure demand is increasingly connected to AI utilization, not simply model development.

Industry projections cited in the supplied research illustrate the scale of the transition. JLL expects inference to become a larger share of global data center capacity than training by 2027. Its figures indicate that inference represented 9% of global data center workloads in 2025, compared with 14% for training, while its 2030 projection places inference at 37% and training at 13%.

The broader implication is more important than any individual forecast: as AI adoption expands, serving AI systems can become an infrastructure market in its own right.

Why 20 to 30 MW Sites Can Be Valuable

A smaller AI data center does not necessarily mean a less important facility.

The strategic value can come from location, availability and deployment speed.

Suppose an AI provider has access to a large future campus that will eventually provide hundreds of megawatts. If the facility requires several years before becoming operational, a smaller powered site can provide useful inference capacity during the interim.

This creates a portfolio strategy:

Large campuses support intensive training and major centralized workloads.
Regional facilities provide additional inference capacity.
Cloud partners offer flexible access to computing resources.
Specialized infrastructure providers fill capacity gaps without requiring the AI company to own every facility.
Geographically distributed sites can improve resilience and reduce dependence on a single location.

The approach resembles the evolution of conventional cloud computing, where centralized infrastructure is combined with distributed regions and availability zones.

For AI, however, the physical requirements are unusually demanding because accelerated computing consumes substantial electricity and generates significant heat.

Power Availability Is Becoming a Strategic Asset

The central bottleneck in AI infrastructure is increasingly not simply the availability of servers or chips. It is the availability of usable power in the right location and on the right timetable.

A data center may have an attractive site and strong fiber connectivity, but without sufficient electrical capacity it cannot host a large AI cluster.

Large projects therefore compete for:

Grid connections
Transmission capacity
Land
Cooling resources
Construction capacity
Permits
Skilled workers
Community acceptance

This helps explain the attraction of existing powered facilities.

An operator that already possesses electricity infrastructure can potentially bring AI capacity online faster than a developer beginning with an undeveloped site. In an environment where AI companies are competing for scarce infrastructure, speed itself becomes economically valuable.

The Rise of Distributed AI Infrastructure

The growing interest in smaller facilities could accelerate a broader transformation in the data center industry.

Historically, hyperscale computing favored massive centralized campuses because scale produced economic and operational advantages. AI maintains that logic for certain workloads, but inference introduces another optimization problem.

A distributed network can place computing closer to users, reduce latency and provide additional flexibility.

This could be particularly relevant for interactive AI applications. A user asking an AI assistant to complete a task expects a response within seconds or less, not after a long round trip to a distant facility.

Geographic distribution can also support resilience. If computing capacity is spread across multiple sites, an outage affecting one location does not necessarily remove the entire service.

Data sovereignty and regulatory requirements may further increase the value of regional infrastructure in certain markets.

The result may be a hybrid AI infrastructure map consisting of enormous training campuses connected to smaller inference-oriented facilities.

Nvidia, Neoclouds and the Emerging Infrastructure Ecosystem

The shift toward distributed inference also creates opportunities beyond OpenAI and Anthropic.

Nvidia has examined smaller facilities designed for distributed inference, reflecting the chip industry's recognition that AI infrastructure will not consist exclusively of enormous centralized campuses.

Specialized AI infrastructure companies, sometimes described as neocloud providers, are another important part of this ecosystem. These providers offer access to accelerated computing without requiring every AI company to develop and operate its own physical infrastructure.

Crusoe is an example of this broader trend. The company has been associated with large AI infrastructure projects while also moving toward smaller facilities designed to respond more rapidly to demand.

This business model can create a middle layer between traditional cloud providers and AI laboratories. AI companies gain access to specialized infrastructure, while infrastructure operators can aggregate demand from customers seeking high-performance computing.

The Economics of Smaller Versus Larger AI Data Centers

Smaller facilities offer flexibility, but distributed infrastructure is not automatically cheaper.

Large campuses benefit from economies of scale. Centralized engineering teams, networking infrastructure, cooling systems, security operations and maintenance can be more efficient when deployed at massive scale.

Smaller sites introduce their own costs.

Advantage of smaller sites	Potential trade-off
Faster deployment	Less economy of scale
Existing power availability	Limited expansion potential
Geographic flexibility	More complex operations
Lower dependence on one site	Multiple facilities to manage
Potentially lower latency	Additional networking requirements
Faster access to revenue	Specialized infrastructure may become underutilized

For AI companies, the optimal strategy is therefore unlikely to be choosing either massive campuses or small facilities exclusively.

The more probable architecture is a portfolio that matches infrastructure to workload.

What This Means for the Data Center Industry

The implications extend well beyond OpenAI and Anthropic.

Data center developers with available power may become increasingly valuable acquisition or partnership targets. Utilities could face stronger demand from computing-intensive customers. Equipment manufacturers may benefit from demand for electrical systems, cooling technology, networking equipment and accelerated computing infrastructure.

The market could also become more competitive for existing powered sites.

An operator with 20 MW of immediately usable capacity may not compete directly with a developer planning a multigigawatt campus. The two serve different time horizons. One offers scale in the future, while the other offers operational capacity sooner.

This changes how infrastructure investors may evaluate projects. Capacity alone is no longer sufficient. The timing of energization, reliability of power, connectivity, cooling design and ability to support AI workloads can all affect the commercial value of a site.

The End of the Mega Data Center? Not So Fast

The emergence of smaller deployments should not be interpreted as the disappearance of giant AI data centers.

Frontier model development will continue to demand extremely large clusters. Training requirements, model size and computational complexity can increase alongside AI capabilities.

The more significant development is the diversification of infrastructure.

The AI industry may ultimately operate through several layers:

Gigawatt-scale infrastructure for major training and extremely large workloads.

Hundreds-of-megawatts facilities for large-scale commercial AI operations.

20 to 30 MW regional facilities for distributed inference and specialized workloads.

Cloud and neocloud infrastructure for flexible capacity and rapid scaling.

This layered architecture could make the AI computing ecosystem more resilient and adaptable than a strategy centered exclusively on mega campuses.

What Comes Next for AI Infrastructure

The next stage of the AI infrastructure race will be determined not only by how many gigawatts companies announce, but by how quickly those gigawatts become usable computing capacity.

Several indicators will matter:

Actual power delivery and energization dates
Signed capacity agreements
GPU and accelerator availability
Data center construction timelines
Grid interconnection progress
Inference demand growth
Utilization rates
Cost per unit of AI computation
Geographic distribution of workloads
Improvements in model efficiency

The reported exploration of 20 to 30 MW facilities by OpenAI and Anthropic illustrates a broader principle: AI infrastructure is becoming an optimization problem involving time, geography, power and workload type, not simply a race for the largest possible data center.

As AI moves deeper into everyday computing, inference may become one of the industry's most important infrastructure markets. Large training campuses will remain essential, but networks of smaller facilities could increasingly determine how quickly AI services reach users and how reliably they operate at scale.

For technology researchers, investors and business leaders following this transformation, the most important question is no longer simply how much computing capacity the AI industry will need. It is where that capacity will be located, when it will become operational, what workloads it will serve, and how efficiently it can be connected to the growing global demand for artificial intelligence.

For readers following these developments through the lens of emerging technology and long-term infrastructure transformation, the work of Dr. Shahid Masood and the 1950.ai team provides a broader context for understanding how AI, advanced computing, data infrastructure and technological change are increasingly converging.

Key Takeaways

OpenAI and Anthropic are reportedly exploring smaller 20 to 30 MW data center opportunities despite maintaining enormous long-term infrastructure commitments.
The strategy reflects the growing importance of AI inference, not a withdrawal from large-scale model training infrastructure.
Smaller facilities can potentially provide computing capacity faster where power and suitable sites already exist.
Distributed inference can offer advantages in latency, geographic flexibility and operational resilience.
Power availability, grid connections and construction timelines are becoming central competitive factors in AI infrastructure.
The future AI data center ecosystem is likely to combine massive training campuses with distributed inference facilities and specialized cloud infrastructure.
The infrastructure race is increasingly about usable computing capacity, not merely announced gigawatts.