Anthropic and OpenAI are shopping for 20MW sites at big-campus prices per megawatt

Nscale wrote in an August analysis that a 20MW cluster holding 10,000 GPUs can cost nearly $2 billion to stand up, citing Nvidia executive Rod Evans ^[6]. That works out to about $100 million a megawatt ^[7]. The other end of the same portfolio is roughly 460MW rented from Nscale at a West Virginia development, an arrangement CNBC valued in August at about $45 billion ^[5], or about $98 million a megawatt ^[8]. The small site and the big campus land within a few percent of each other per megawatt.

What the 20-30MW option buys is an earlier date at the same unit price ^[1][7][8].

Nscale's number is a claim about Nscale's build. It assumes 10,000 GPUs behind 20MW, which is 2 kW of site power per GPU ^[9], and $200,000 per GPU all in ^[10]. A shell that can only carry half that density puts 5,000 GPUs behind the same connection ^[28]. Before the figure transfers to a site someone offers you in the Nordics, somebody has to confirm the hall takes the density.

The workload is what makes a small hall viable at all. Training a frontier model needs large numbers of chips communicating closely inside one concentrated cluster, while many inference requests can be routed independently across smaller clusters ^[11]. Google Cloud says AI inference requires low-latency, resilient networks, and its deployment guidance recommends multiple regions to improve reliability and accelerator availability ^[12].

Lawrence Berkeley National Laboratory reported in June that rapid data center load growth had created bottlenecks slowing large-load grid connections ^[15]. RAND has warned that permitting and regulatory delays could prevent some new grid infrastructure from being completed by 2030 ^[16]. Jabez Tan, head of research at Structure Research, described the advantage to CNBC as "speed to usable capacity" ^[13]. 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 ^[14].

Compute procurement at Anthropic sits at founder level. Its leadership page says co-founder and chief compute officer Tom Brown runs the technical organization responsible for securing, scaling and using its compute resources ^[24].

In the same reporting, the gigawatt plans still stand. Anthropic's May 28 funding announcement said it had signed agreements with Amazon for up to 5GW of new capacity and with Google and Broadcom for another 5GW of next-generation TPU capacity ^[20]. OpenAI said in March that durable compute access was a central reason for its $122 billion financing, and its Stargate program has spread planned capacity across multiple US states ^[22][23]. A 20-30MW site is 4 to 7 percent of the West Virginia block alone ^[26]. Anthropic's figures count commitments and access; none of that power is serving Claude yet ^[21].

JLL's 2026 Global Data Center Outlook projects inference's share of global data center workloads rising from 9% in 2025 to 37% in 2030, and expects inference to overtake training as the dominant AI workload in 2027 ^[17]. That is roughly a 4.1x increase in share over five years ^[19]. JLL's forecast holds only if applications reach enough adoption to produce sustained demand ^[18].