Nscale and Microsoft Plan a West Virginia AI Campus With 430,000 NVIDIA GPUs

Nscale has announced a letter of intent with Microsoft for 1.35 GW of AI capacity at its new Monarch campus in West Virginia. The project was unveiled on March 16, 2026, at GTC. It is built around about 430,000 NVIDIA Vera Rubin GPUs, with first deliveries scheduled for late 2027.

Scale of the deal

This is not simply a chip purchase but a long-term arrangement in which Nscale will build and operate infrastructure for Microsoft workloads. The company describes a multi-year agreement for compute services and a long-term data center lease, with deployment rolling out in several phases.

Formally, this is still a letter of intent, not a final contract. Even so, the stated parameters make it one of the largest AI infrastructure announcements of 2026. In an AI context, 1.35 GW of compute capacity is comparable not to a single data center but to an entire industrial site.

The project centers on the NVIDIA Vera Rubin NVL72 architecture and the NVIDIA Vera Rubin DSX AI Factory reference design. Nscale calls it the first large commercial deployment of this configuration. If the project reaches its stated size, Microsoft gets a dedicated site for training and running inference on large models, while Nscale becomes one of the key operators of the new wave of GPU factories.

Site and power

The campus is being built in Mason County, West Virginia, at the Monarch Compute Campus. Nscale simultaneously announced the acquisition of AIPCorp, which owns the site. The site covers up to 2,250 acres, and its main distinction is that the project is described as the first state-certified AI microgrid in the US. In other words, it is a data center designed from the start around its own power supply, rather than around waiting several years for a connection to the external grid.

• about 430,000 NVIDIA Vera Rubin GPUs
• 1.35 GW of capacity for Microsoft's needs
• 2 GW of on-site generation by the first half of 2028
• potential to scale beyond 8 GW at a single site

Power comes through a partnership with Caterpillar: Nscale plans to deploy G3500-series gas generators and reach 2 GW of on-site generation in the first half of 2028. The campus is planned to run independently of the local grid, so it does not have to wait for scarce grid connections or shift load onto existing consumers. The design still allows for a future grid connection and even for feeding power back. Nscale also talks about carbon capture plans and says the site will use less water without affecting municipal water supply, though such claims have yet to be proven in practice.

Why it matters for Nscale

For Nscale, this is not an isolated deal but part of a much larger bet on vertical integration. A week before the announcement, the company closed a $2 billion Series C round at a $14.6 billion valuation, and on March 16, 2026, it announced not only the Microsoft letter but also the AIPCorp acquisition. At the same time, Nscale is setting up a division, Nscale Energy & Power, in Houston. The logic is simple: in AI infrastructure, ordering GPUs is no longer enough. You need to control land, power, construction, commissioning and operational management of the whole stack.

"Our approach is to combine our own data centers, leasing and strategic partnerships,"

Microsoft said.

The two companies already have a shared history: Microsoft was already a customer of Nscale at the Narvik data center in Norway. The new US campus takes the partnership to another level, closer to Ashburn in Virginia and to Chicago, major hubs of cloud and network infrastructure. For Microsoft, it is a way to add AI capacity quickly through a mix of its own facilities, leases and outside partners. For Nscale, it is a chance to gain a foothold in the world's largest AI market and show it can sell not just GPU time but megawatts.

What it means

The Nscale-Microsoft deal shows the AI race shifting increasingly from models and chips toward land, electricity and the time it takes to bring facilities online. The winner is no longer the one who simply found more GPUs, but the one who assembled the full package in advance: site, power, network, financing and a customer.

It is worth remembering that this is still a letter of intent. If deliveries really start in late 2027 and power reaches the first 2 GW in the first half of 2028, Monarch could become one of the anchor AI sites in the US.