Anthropic and OpenAI seek smaller data centers while gigawatt sites wait

Why it matters

Anthropic and OpenAI are turning compute procurement into portfolio management. Smaller powered sites can put inference capacity online sooner, tying product growth to leaders who can secure electricity, chips and delivery dates.

Anthropic, co-founded by CEO Dario Amodei, and OpenAI, co-founded by CEO Sam Altman, are exploring data center deployments of roughly 20-30 megawatts in a bid to bring AI capacity online faster, CNBC reported on September 18th.

Anthropic has sounded out potential arrangements in the U.K. and Nordic countries, according to four people familiar with the discussions. Two sources told CNBC that OpenAI had explored opportunities in the Nordics, and one source described conversations involving both labs about possible U.S. deployments at the same scale.

The discussions concern potential capacity rather than completed transactions. For Anthropic, compute is explicitly a co-founder-level function. Anthropic's leadership page says co-founder and chief compute officer Tom Brown runs the technical organization responsible for securing, scaling and using its compute resources.

That operating responsibility sits alongside the research and safety agenda Amodei carried from OpenAI into Anthropic. OpenAI said Amodei had worked on GPT-2 and GPT-3 as its vice president of research before leaving in late 2020 to pursue a new project with several colleagues. Anthropic emerged in 2021 with Amodei as CEO, his sister Daniela Amodei as president and Brown among its co-founders.

Speed is becoming a compute product

Lawrence Berkeley National Laboratory reported in June that rapid data center load growth had created bottlenecks slowing large-load grid connections. RAND has also warned that permitting and regulatory delays could prevent some new grid infrastructure from being completed by 2030. Existing powered sites can therefore offer a shorter route to usable capacity than projects that require new grid infrastructure.

Jabez Tan, head of research at Structure Research, described the advantage to CNBC as "speed to usable capacity." He said a few megawatts at an existing powered site can be more practical than waiting for a much larger block, especially when workloads can be divided across locations.

The technical reason is inference, the work performed when a deployed model answers requests. Training a frontier model requires large numbers of chips to communicate closely inside a concentrated cluster. Many inference requests can be routed independently across smaller clusters, giving Anthropic and OpenAI more freedom to place production capacity near customers or available power.

JLL's 2026 Global Data Center Outlook projected inference's share of global data center workloads to rise from 9% in 2025 to 37% in 2030. The property services group expects inference to overtake training as the dominant AI workload in 2027. That forecast depends on applications reaching enough adoption to produce sustained demand, but it explains why regional clusters are becoming strategically useful before the biggest campuses open.

Infrastructure geography can affect product performance. Google Cloud says AI inference requires low-latency, resilient networks, while its deployment guidance recommends multiple regions to improve reliability and accelerator availability. Nscale wrote in an August analysis that a 20 MW cluster containing 10,000 GPUs can cost nearly $2 billion to stand up, citing Nvidia executive Rod Evans. Small is relative when the unit of account is an AI factory.

The gigawatt projects are still coming

The smaller search complements a much larger buildout. CNBC reported in August that Anthropic had agreed to rent roughly 460 MW from Nscale at a West Virginia development under an arrangement valued at about $45 billion.

Anthropic's own May 28th funding announcement said it had signed agreements with Amazon for up to 5 GW of new capacity and with Google and Broadcom for another 5 GW of next-generation TPU capacity. Anthropic also said it had secured access to GPU capacity at SpaceX's Colossus facilities. Those Anthropic-supplied figures describe commitments and access, rather than power already serving Claude.

OpenAI is pursuing the same portfolio approach at a larger stated scale. OpenAI said in March that durable compute access was a central reason for its $122 billion financing. Its Stargate program has spread planned capacity across multiple U.S. states, while CNBC reported separate commitments covering 3 GW in Georgia and 8 GW in Ohio.

A 20-30 MW deployment may cover only a fraction of a lab's long-term requirement, but a collection of those blocks can add meaningful capacity while multigigawatt campuses remain under development.

Smaller deployments also reduce exposure to one construction schedule, power market or infrastructure provider. They add operational work. Anthropic and OpenAI must distribute models, route traffic, manage data requirements and maintain performance across a more fragmented fleet. Anthropic's decision to put a co-founder in charge of securing, scaling and using compute reflects how closely infrastructure planning now tracks Anthropic's product demands.

Anthropic's product expansion raises the near-term bill

Anthropic has been pushing Claude beyond a single chat interface into coding, research, sales and other workplace tasks. RuntimeWire reported this week that Anthropic brought 37 Salesforce pipeline skills into Claude, allowing the assistant to read and update CRM records under customers' existing permissions.

RuntimeWire also reported that Anthropic says Claude now writes 80% of its merged internal code. Anthropic said its test volume grew tenfold and continuous-integration jobs increased 25-fold in six months, forcing a rebuild of its test-selection system. Those are self-reported internal metrics, but they illustrate the infrastructure pattern: useful agents create repeated, production-grade workloads rather than occasional chatbot requests.

The reported smaller-site strategy is therefore tied to product delivery as much as data center procurement. Brown leads Anthropic's technical organization responsible for securing, scaling and effectively using Anthropic's compute resources, while Dario and Daniela Amodei continue expanding the product across enterprise tasks. The labs still want gigawatts. Their immediate customers operate on a much shorter clock.