The most important item in an artificial-intelligence data center may not be the GPU. It may be the piece of paper saying the local utility will supply several hundred megawatts, followed closely by a substation, a cooling loop and a crew that can turn all of it into a working building. Core Scientific owns and leases a collection of those ingredients across the United States. That has made a company once known almost entirely for mining Bitcoin newly relevant to the race to build AI infrastructure.
The pivot is physical, not cosmetic. Bitcoin mining taught Core Scientific to find inexpensive, dependable power; operate specialized computers around the clock; negotiate with utilities; and fix large fleets when something failed. AI training and inference require many of the same muscles, then add harder specifications: high-capacity fiber, stricter uptime, redundant electrical systems, direct liquid cooling and racks that can consume 50 to more than 200 kilowatts apiece.
Core Scientific now sells that upgraded environment as high-density colocation. A customer brings the computing equipment and reserves a large block of electrical capacity. Core Scientific designs, builds and operates the powered shell around it. The offering begins around 30 megawatts, a scale intended for hyperscale cloud providers, model builders and enterprises rather than a startup looking for a few cabinets.
A business measured in megawatts
Software companies count users and subscriptions. Core Scientific increasingly counts leased power, billable power and the months between them. By mid-July 2026, the company said 437 megawatts were billing, equal to about $635 million in average annualized colocation revenue under generally accepted accounting principles. It had leased roughly 1.1 gigawatts of customer power, representing more than $24 billion in potential contracted revenue over the life of signed agreements.
Those three numbers are not interchangeable. A signed megawatt does not produce revenue until a data hall is commissioned and service begins. Potential contract value can span 12 or 15 years, may include conditions and is not cash already in the bank. That gap is the central drama of the business: Core Scientific must finance and build vast facilities before the rent checks fully arrive.
In the second quarter of 2026, colocation produced $136.7 million of the company's $164.2 million in total revenue. Self-mining added $21.5 million and hosted mining $6.0 million. One year earlier, the mix leaned far more heavily toward Bitcoin. The transition is visible in the income statement before it is complete in the landscape.
The new industrial loop
The useful inheritance of Bitcoin
Core Scientific was founded in 2017 and began mining in North Carolina and Georgia the next year. It opened a Kentucky site in 2019, reached roughly 500 megawatts of operating capacity by 2022 and became one of North America's largest public miners. The expansion delivered scale, but it also exposed the company to Bitcoin prices, network difficulty, energy costs and debt.
In December 2022, the company entered Chapter 11. It continued operating, reorganized and emerged in January 2024 after cutting debt by approximately $400 million. The experience is not a decorative comeback chapter. It explains the current appetite for longer contracts and revenue that does not reset with the Bitcoin market every morning.
The pivot is less about leaving crypto than finding a steadier tenant for the same scarce electrical address.YesPress analysis
Core Scientific still mines Bitcoin and still hosts customer-owned mining machines, but it plans to repurpose remaining mining facilities for high-density colocation as conditions allow. Mining provides cash and lets the company use capacity during a staged conversion. It is also an awkward roommate: every megawatt retained for miners is a megawatt that cannot yet be rebuilt for a long-term AI customer.
The infrastructure is not identical. A conventional mining hall can tolerate simpler airflow and less network complexity. GPU clusters demand tightly controlled cooling, low-latency connectivity and much stronger service guarantees. The conversion therefore depends on design and construction, not merely replacing one metal box with another.
Illustrative rack power density
Kilowatts per rack · ranges described in company materials
One anchor customer, then a second
CoreWeave is the relationship that made the strategy legible. Core Scientific had hosted its GPU equipment as early as 2019. In February 2024, the companies signed for 16 megawatts at an Austin facility. Core Scientific delivered that capacity more than 30 days early, then the relationship expanded across multiple campuses into hundreds of megawatts and billions of dollars in potential revenue.
The closeness created opportunity and concentration. CoreWeave proposed acquiring Core Scientific in 2024 and was rebuffed. In July 2025, the companies announced an all-stock transaction valued at roughly $9 billion when signed. Core Scientific shareholders rejected it that October, and the merger agreement was terminated. The landlord remained independent; its largest commercial relationship remained intact.
AMD gives the company a second strategic axis. Announced in July 2026, the partnership includes agreements for approximately 530 megawatts across five U.S. sites beginning in 2027, with more than $14 billion of potential base contracted revenue. AMD may expand the relationship to 2.5 gigawatts. The two companies also plan to collaborate on physical designs around Instinct GPUs, EPYC processors and ROCm software.
That arrangement is meaningfully different from renting generic floor space. A chipmaker is reserving infrastructure for deployments by its ecosystem, while the operator designs around the hardware roadmap. It also reduces, though does not erase, the risk of depending too heavily on one AI cloud customer.
Why this company, and why now?
Core Scientific competes with familiar data-center names such as Equinix, Digital Realty, NTT, Switch and CyrusOne. It also meets former peers from crypto, including IREN, TeraWulf and Applied Digital, in the hunt for AI tenants. Cloud platforms can build for themselves, while specialist developers can assemble new campuses from scratch.
Its distinction is a portfolio assembled before generative AI turned grid access into a boardroom subject. At the end of 2025, Core Scientific owned or leased ten data centers across seven states with about 1.4 gigawatts of gross utility power. Existing sites can offer shorter routes to service than greenfield projects waiting years for interconnection, although conversions and expansions still carry permitting, equipment and construction risk.
The company also has the operating scar tissue of machines that never sleep. Its pitch combines power-market specialists, engineers, security teams and a 24-hour network operations center. The public values it lists - Team First, Extreme Ownership, Innovate and Simplify, and Transparency - sound like standard corporate nouns. In this setting they double as a checklist for a business where one missed commissioning date can move revenue across quarters.
Core Scientific is expanding aggressively. In 2026 it described plans that could take both its Muskogee, Oklahoma, and Pecos, Texas, campuses toward 1.5 gigawatts of gross power each. It announced up to $1 billion of strategic financing with Morgan Stanley, expanded a J.P. Morgan facility to $1 billion and priced $3.3 billion of senior secured notes. The scale clarifies both the ambition and the risk: this second act needs concrete and copper long before it earns a margin.
Beyond the AI factory
The customer map stretches beyond model training. Core Scientific markets dense infrastructure for cloud services, financial analysis, medical research, manufacturing simulation, media rendering, government workloads and oil-and-gas modeling. In March 2026 it became the official data-center partner to Cadillac Formula 1 Team, advising on infrastructure for simulation, analytics, digital manufacturing and race operations at the team's new U.S. headquarters.
That deal is small beside a gigawatt agreement, but revealing. Core Scientific's expertise can be sold as architectural guidance, not only as capacity. An F1 team and an AI cloud have different outputs; both need dense, dependable computation and both care what happens when the lights flicker.
The company's place in the market is therefore between the grid and the cloud. It does not design the leading accelerators or sell an AI model. It makes the industrial conditions under which those products can work. If software is weightless in the popular imagination, Core Scientific is a reminder that modern software has become very heavy indeed.
The product is not the room. It is the promise that a frightening amount of electricity will become useful computation on a particular date.The infrastructure thesis
Investors will watch the conversion rate from leased to billable megawatts, the cost of each build, customer concentration and the declining mining contribution. Customers will care about delivery, uptime, cooling and transparent economics. Communities will care about grid effects, water, noise, jobs and whether large loads improve or strain local systems. Core Scientific has to satisfy all three groups while constructing several campuses at once.
That is what makes the story more interesting than a fashionable pivot. Core Scientific did not discover AI and paste it onto a mining presentation. It found that a difficult capability built for one volatile market - controlling power-dense computing at industrial scale - had become valuable to another. The next test is less poetic: build the halls, energize the racks and start the meter.