The Architecture of Outsourced Infrastructure
When we talk about artificial intelligence infrastructure, we tend to flatten the stack into a neat single story. A giant tech firm builds a server farm, fills it with chips, and turns on the model training. But reality is far messier. The modern AI compute boom relies on a web of subcontracting, where the party that owns the land is not the party that owns the building, which is not the party that owns the networking gear, and none of them are necessarily the ultimate end user.
CoreWeave’s Marble site in North Carolina offers a textbook case study in this modular approach. Located in Marble, NC, the facility represents a fascinating hybrid of legacy infrastructure adaptation and cutting-edge GPU deployment. It is operational, highly dense, and built on foundations that originally had nothing to do with deep learning.
Understanding how Marble came to be requires looking past the glossy press releases and examining the mechanical and financial handoffs between companies like Core Scientific and CoreWeave. It is a window into how capital-intensive AI infrastructure actually gets assembled on a compressed timeline.
From Crypto Mining to High-Density GPU Hosting
The physical shell of the Marble data center was not originally designed for massive transformer models or multi-billion-parameter LLM clusters. Instead, the site started its modern life as a cryptocurrency mining facility operated by Core Scientific.
As the crypto economics shifted and the demand for high-performance computing (HPC) skyrocketed, infrastructure operators faced a choice: keep mining bitcoin on depreciating ASICs or pivot the real estate toward the voracious energy demands of generative AI. Core Scientific chose the latter, embarking on a major retrofit to convert the Marble property into a high-density GPU hosting environment.
This conversion was no minor cosmetic update. High-density AI chips require specialized cooling, heavy electrical distribution upgrades, and robust substation capacities that standard commercial offices or older server rooms simply cannot handle. Core Scientific structured the project to hand over roughly 70 megawatts of high-density infrastructure out of 100 megawatts of total owned capacity at the location. Site modifications kicked off in earnest during the second half of 2024, transforming a remote crypto-mining footprint into an industrial-grade compute hub.
By March 2026, the physical reality on the ground had begun to match the ambitions. Two of the facility's planned data halls were online, delivering about 37 billable megawatts of power and supporting roughly 15,000 Grace Blackwell GPUs. A third hall remained in active commissioning, highlighting the incremental, phased nature of modern data center buildouts where racks are energized and loaded while construction crews finish drywall down the corridor.
The Numbers: Power, Compute, and Capital Cost
By May 2026, the handoff was effectively complete. Core Scientific announced that the Marble site had been fully transferred to CoreWeave, bringing roughly 65 megawatts of billable customer power completely online.
Quantifying the compute capacity of a facility like Marble requires translating raw megawatts and silicon counts into standardized metrics. According to tracking data from Epoch AI, Marble hosts an estimated 79,000 H100-equivalents of compute capacity. That output is driven primarily by advanced Nvidia B200 hardware—specifically, an estimated 31.1 thousand Blackwell units operating across the facility's modernized data halls.
To keep these chips from melting, the facility relies on a robust thermal management system. Public documentation and satellite telemetry indicate the presence of 51 air-cooled chillers providing approximately 67.8 megawatts of cooling capacity, closely matching the site's 65-megawatt IT load.
The financial footprint is equally striking. The cumulative capital expenditure for the Marble site sits at an estimated $2.5 billion. For perspective, when the site was partially energized at 36 megawatts in early 2026, the estimated cost was roughly $1.4 billion. Scaling from that intermediate milestone up to the full 65-megawatt buildout nearly doubled the capital tally, illustrating the steep marginal costs associated with equipping every square foot of high-density floor space with top-tier Nvidia silicon and advanced cooling infrastructure.
The OpenAI Connection and Chip Ownership Realities
One of the most persistent questions surrounding major AI data centers is simple: who is actually running workloads on these chips? In the case of Marble, the ownership and operational lines are distinct, but the ultimate end-user remains a subject of intense industry speculation.
The physical chips inside Marble are owned by CoreWeave. CoreWeave operates the infrastructure as a specialized cloud provider, leasing out GPU cycles to clients who need raw compute for training and inference. However, market intelligence and industry reports frequently tie Marble's capacity to OpenAI, suggesting that a significant share of the facility's 79,000 H100-equivalents is dedicated to fueling OpenAI's frontier models.
This dynamic highlights a broader shift in the tech ecosystem. Companies like OpenAI do not necessarily need to own or operate real estate in rural North Carolina. Instead, they secure multi-billion-dollar compute arrangements with specialized infrastructure partners like CoreWeave, who in turn source their real estate from landlords like Core Scientific. It is a multi-tiered supply chain where financial risk is distributed across several balance sheets.
For CoreWeave, securing a site like Marble is essential to maintaining its competitive edge against hyperscalers like Microsoft, Amazon, and Google. By aggregating third-party real estate and outfitting it with cutting-edge Nvidia Blackwell architecture, CoreWeave positions itself as an agile landlord of the AI era—capable of spinning up massive clusters faster than a traditional cloud titan can get through municipal zoning boards.
What Marble NC Tells Us About the Modern AI Cluster
The Marble, North Carolina facility is much more than a remote collection of servers; it is a microcosm of the entire AI infrastructure boom. It demonstrates how capital gets deployed when speed is the only metric that matters.
When you strip away the marketing terms, Marble is a study in industrial repurposing. It proves that the physical constraints of the AI revolution are being solved not just by building brand-new greenfield campuses from scratch, but by cannibalizing and upgrading yesterday's crypto-mining facilities to feed tomorrow's neural networks.
As the industry moves deeper into the Blackwell era, sites like Marble will become increasingly common. They show that the future of compute will likely be fragmented—owned by specialized financial entities, operated by agile cloud providers, cooled by massive chiller arrays, and ultimately rented out to the handful of labs building artificial general intelligence. Marble NC may sit quietly in the mountains of North Carolina, but its economic and operational blueprint is reshaping how the entire world computes.
