Advanced Micro Devices announced Tuesday that it has secured more than 529 megawatts of U.S. artificial intelligence data center capacity from Core Scientific (NASDAQ: CORZ) under a set of 15-year leases that could generate more than $14 billion in base contracted revenue for the operator — the single largest infrastructure commitment AMD has made and one of the largest AI colocation deals announced in 2026.

What the dollar figure obscures is the strategic logic. AMD is not buying servers or renting cloud capacity — it is locking up the physical layer of AI compute: land that already has power, substations already connected to the grid, and fiber already in the ground. That is exactly what NVIDIA does not own and has never needed to own, because NVIDIA's real competitive moat is CUDA, a proprietary software ecosystem with an 18-year head start and roughly 85 percent of the AI GPU market. AMD's Instinct GPUs and ROCm software stack are genuinely competitive on inference workloads now — AMD's MI355X came within single-digit percentage points of NVIDIA's B200 at MLPerf Inference benchmarks in April 2026 — but ROCm still cannot match CUDA's ecosystem depth on training. Locking customers into AMD-powered facilities is AMD's substitute for that missing ecosystem gravity.

The announcement landed on the same day Core Scientific reported Q2 2026 earnings showing total revenue of $164.2 million — more than double the $78.6 million posted in Q2 2025, and well ahead of analyst estimates of approximately $135 million. The company that filed for Chapter 11 bankruptcy in December 2022 now earns roughly 83 cents of every revenue dollar from leasing data center space to AI and high-performance computing customers.

Deal Structure: Five Sites, One Option That Could Triple the Partnership

The initial agreements cover 529 megawatts of critical IT capacity across five sites in four U.S. states, with lease terms running up to 30 years under certain extension options. AMD directly leased 377 megawatts at Core Scientific facilities in Pecos and Hunt County, Texas, and Muskogee, Oklahoma. An entity identified by TipRanks as Neocloud — described in Core Scientific's official press release as "an unnamed cloud provider" operating under arrangements supported by AMD — took on the remaining 152 megawatts at sites in Auburn, Alabama, and Dalton, Georgia.

AMD also received warrants to purchase up to 30 million Core Scientific shares at $23.47 per share — a stake that would give AMD meaningful skin in CORZ's success. Roughly 6.5 million of those warrants vested immediately when the initial leases were signed; the remainder vest as additional capacity comes online.

The option clause may matter more than the initial deal. AMD secured the right to reserve up to an additional 1,925 megawatts of capacity through December 28, 2028 — a window that, if exercised in full, would expand the partnership to approximately 2.5 gigawatts and potentially push contracted revenue well beyond the $14 billion base figure.

Annual escalators on the AMD contracts are set at 2.5 percent per year, meaning the real economic value of the leases compounds over the 15-year term, according to Core Scientific's Q2 2026 earnings call.

Customer deployments under the AMD partnership are expected to begin in 2027. AMD and Core Scientific will collaborate on physical data-center design and on deploying AMD's Instinct GPU accelerators, EPYC server processors, and the ROCm software stack across all five sites.

"AI deployments are accelerating rapidly, and bringing that compute online requires trusted infrastructure partners with the scale and power to support the next era of AI," said Mathew Hein, AMD's senior vice president and chief strategy officer of corporate development, in the joint announcement. "Core Scientific's extensive portfolio of AI-ready data centers expands access to the infrastructure our customers need to deploy AMD AI solutions at scale."

Why Power Megawatts Now Outrank Chip Specifications

The scale of a modern AI data center is difficult to visualize without anchoring it to something concrete. A single H100 GPU draws 700 watts under full compute load. An eight-GPU server node consumes 10 to 12 kilowatts. A rack of these nodes draws 80 to 140 kilowatts. A 10,000-GPU training cluster consumes 10 to 15 megawatts — enough to power a small town. At 529 megawatts, the initial AMD/Core Scientific commitment could, when fully built out, support the equivalent of 35,000 to 53,000 such GPU clusters, depending on configuration and workload type.

That kind of demand has made power availability — not chip production — the binding constraint on AI infrastructure deployment. Jefferies, which raised its Core Scientific price target to $37 from $24 following the deal, described power as the "binding constraint" for AI data centers and said the company's economics on its CoreWeave leases were the best among its peer group. The race to lock up sites with power already contracted, grid connections already established, and fiber already in the ground is happening faster than new greenfield data centers can be permitted and built.

"As AI demand continues to accelerate, access to power, land and data center infrastructure has become critical to bringing new AI compute online," AMD's press release stated.

Core Scientific holds a specific structural advantage in that race. The company built its original bitcoin mining facilities to exactly the specifications AI data centers now require: high-voltage power delivery, robust cooling infrastructure, reliable fiber connectivity, and large land footprints in locations with favorable electricity economics. Bitcoin mining ASICs — the chips that perform the SHA-256 computations that secure the blockchain — cannot be repurposed for GPU workloads. Almost all of the compute-facing hardware must be stripped out and rebuilt. But the electrical and physical plant underneath it transfers directly, and that is worth roughly $7 to $9 million per megawatt in avoided greenfield construction costs.

VanEck Research put the economics starkly: retrofit conversions of existing mining sites push capital expenditure per megawatt down to roughly $3 million to $4 million, compared with $10 million to $12 million for greenfield builds — and Core Scientific's own guidance put its build-out cost at $11 million to $12 million per megawatt for new capacity. The retrofit economics explain why unlevered EBITDA yields on converted mining sites can reach 28 to 32 percent, compared with 12 to 15 percent for pure greenfield AI data centers.

How Bitcoin Mining Infrastructure Converts to AI Compute

Not everything transfers cleanly. The most important structural difference between bitcoin mining infrastructure and AI data centers is power density. Traditional mining facilities ran at roughly 5 to 20 megawatts per site. A single modern AI data center building can require 50 to 100 megawatts. Core Scientific's sites are already engineered for the higher power loads — one of the key reasons the company could move so quickly from its first AI deal (the initial CoreWeave contract in February 2024) to operating 437 megawatts of billable AI colocation capacity by mid-July 2026.

The AMD Instinct GPU line that will run in these facilities represents AMD's most serious challenge yet to NVIDIA's data-center dominance. The Instinct MI300X carries 192 gigabytes of HBM3 memory on a single GPU — more than any single NVIDIA GPU in current production — which matters practically because large AI models that require two NVIDIA H100s to run in FP16 precision can fit on a single MI300X, simplifying the serving architecture and removing NVLink interconnect complexity. The Instinct MI355X, AMD's current flagship, posted its strongest-ever result at the MLPerf Inference 6.0 benchmarks published in April 2026, landing within single-digit percentage points of NVIDIA's B200 for server inference workloads.

The performance convergence is meaningful for this deal specifically. Core Scientific's facilities will run AMD customer workloads — and for the inference-heavy deployment patterns typical of cloud providers and enterprise AI customers, AMD's price-performance gap relative to NVIDIA is at its narrowest. AMD hardware currently undercuts NVIDIA's pricing by 15 to 40 percent per performance tier; for inference workloads, the gap narrows further to roughly 25 percent cheaper per token, according to competitive GPU computing analysis.

The ROCm software stack that ties the AMD Instinct hardware together is open-source — built on AMD's Heterogeneous Interface for Portability (HIP) API — making it architecturally different from NVIDIA's proprietary CUDA in one critical respect: code written for ROCm can, in principle, run on other platforms. PyTorch, JAX, vLLM, and llama.cpp all support ROCm officially as of ROCm 7.x, and seven of the ten largest AI model builders now run production workloads on AMD Instinct hardware.

Core Scientific's Books Show a Company Already Through Its Transformation

The earnings data released Tuesday alongside the AMD announcement illustrate just how completely Core Scientific has reoriented its business. Colocation services — leasing data center infrastructure to AI and HPC customers — generated $136.7 million in Q2 2026, up from just $10.6 million in Q2 2025, an increase of more than 1,190 percent in a single year. Self-mining revenue, once the company's core business, fell approximately 66 percent year-over-year to $21.5 million in the quarter, according to Blockspace's Q2 earnings coverage.

Core Scientific was billing customers for 437 megawatts of capacity as of mid-July, a figure the company linked to roughly $635 million in annualized colocation revenue. The AMD deal, once construction is complete and billing begins in 2027, will bring the total leased portfolio to approximately 1.1 gigawatts, representing more than $24 billion in potential contracted revenue across all active deals.

Capital expenditures reached $797.5 million in Q2, more than double the $389.2 million spent in Q1 and more than six times the $121.3 million invested in Q2 2025. The company posted a net loss of $1.155 billion for the quarter, or $3.32 per share — driven primarily by non-cash write-downs and depreciation charges tied to the accelerated wind-down of mining assets and ongoing construction of new AI capacity, not by operational losses in the colocation business itself. Core Scientific ended the quarter with approximately $1.8 billion in liquidity.

Adam Sullivan, Core Scientific's chief executive, pointed to the company's execution record as the foundation for securing a chipmaker of AMD's scale as a direct tenant. "Our proven execution capabilities and ability to deliver high-density infrastructure at scale position us to support AMD's technology roadmap and grow our relationship meaningfully over time," Sullivan said in the joint press release.

Mining, Measured in Its Final Months

The AMD deal effectively closes the book on Core Scientific as a bitcoin mining company. The company ended June with nearly 30 percent fewer miners online than it had at the close of Q1 2026, and it is now running self-mining operations at only two of its sites. Management has indicated that mining activity will continue declining through the rest of 2026, operating primarily to offset contractual power costs during the wind-down period. Hosted mining — in which Core Scientific ran mining equipment on behalf of third-party clients — is expected to end entirely by year-end.

Core Scientific also terminated its agreement to purchase bitcoin-mining chips from Block (NYSE: SQ), recording a $41.9 million charge as a result. The canceled agreement, originally signed in 2024, had covered 3-nanometer chips representing approximately 15 exahashes per second of mining capacity.

Even its bitcoin treasury is being managed toward an exit. The company held 848 bitcoin worth approximately $49.7 million at the end of June and had sold 2,385 bitcoin for $208.2 million in Q1 2026 to fund capital expenditures.

Core Scientific's pivot is not unique in the industry. IREN, TeraWulf, Bitfarms, CleanSpark, and Hive Digital have all moved or are moving significant portions of their power infrastructure into AI and high-performance computing. By the end of 2026, sector analysts at CoinShares expect AI and HPC to represent approximately 70 percent of revenue for miners that have executed major colocation contracts — a figure that is already effectively true for Core Scientific, where AI colocation now generates more than 83 cents of every revenue dollar.

But the scale and speed of Core Scientific's transition stand out. Its existing 590-megawatt contract with CoreWeave, signed initially in 2024 and expanded multiple times since, had already projected $10.2 billion in revenue over 12 years before the AMD deal added another $14 billion on top. No other former bitcoin miner has accumulated a contracted revenue pipeline of comparable size.

Does AMD's Infrastructure Bet Change the AI GPU Competitive Dynamics?

AMD holds roughly 5 to 7 percent of AI GPU market share by revenue, compared with NVIDIA's approximately 85 percent. The gap is real and the causes are well-documented: NVIDIA's CUDA platform has roughly 5.9 million registered developers, 18 years of accumulated libraries, and deep integration with every major machine learning framework. ROCm, despite dramatic improvements in its 7.x versions, still lags CUDA on training workloads by 20 to 30 percent and requires more engineering expertise to deploy, according to competitive GPU stack comparisons.

The Core Scientific deal does not solve the software gap. What it does is create a physical layer where AMD's infrastructure is the default — where customers who want to run large-scale AI inference workloads on AMD hardware in 2027 and beyond will find dedicated, purpose-built facilities already designed for AMD's specific power and cooling requirements. NVIDIA's path to customer stickiness runs through the CUDA developer ecosystem. AMD's path, as this deal illustrates, runs through the building itself.

"CORZ is the first mover among bitcoin miners in what we expect will be a fast-growing, high-demand environment for HPC data center capacity," one analyst noted in 2024 commentary quoted in SEC filings. The AMD partnership extends that first-mover position into a direct relationship with a major chipmaker — something none of Core Scientific's peers in the crypto-to-AI transition have yet secured.

How Does This Deal Affect AI Chip Buyers?

For enterprises and cloud providers evaluating AMD Instinct hardware for AI deployments, the Core Scientific deal signals that AMD is committing the infrastructure layer — not just the silicon — to support large-scale customer commitments. Customers who want predictable access to 100-plus megawatt AMD compute blocks in 2027 now have a named operator and named sites to evaluate. That specificity reduces the planning risk that has made some enterprise buyers reluctant to commit to AMD's ecosystem when NVIDIA's H100 availability was more predictable.

The deal also raises the competitive stakes for NVIDIA's infrastructure strategy. NVIDIA does not own or lease data centers — its power comes from the CUDA software ecosystem and from its relationships with hyperscalers and cloud providers who build their own facilities. AMD is now betting that a chipmaker can compete differently: by controlling the physical destination of its compute deployments the way a railroad controls routes.

Frequently Asked Questions

What is Core Scientific and why did AMD partner with it instead of a traditional data center operator?

Core Scientific (NASDAQ: CORZ) is a former bitcoin mining company that has repositioned itself as a provider of high-density AI data center infrastructure. It operated some of the largest bitcoin mining facilities in North America before filing for Chapter 11 bankruptcy in December 2022 and emerging as a reorganized company in January 2024. What made it attractive to AMD is the infrastructure it built for mining — power substations, grid connections, fiber, and large land footprints across multiple U.S. states — all of which are the same inputs AI data centers require. Building equivalent capacity from scratch ("greenfield") costs roughly $10 million to $12 million per megawatt; Core Scientific's existing site infrastructure can be converted at $3 million to $4 million per megawatt, giving both parties a significant economics advantage over building new, as documented in VanEck Research's bitcoin miners AI infrastructure valuation framework.

How does AMD's ROCm software stack compare to NVIDIA's CUDA, and why does it matter for this deal?

CUDA is NVIDIA's proprietary GPU computing platform — closed-source, with roughly 5.9 million registered developers and 18 years of accumulated libraries. ROCm is AMD's open-source alternative, built on the HIP (Heterogeneous Interface for Portability) API. For training large AI models, ROCm still lags CUDA by roughly 20 to 30 percent in performance. For inference workloads — running already-trained models to generate outputs — the gap has narrowed to single digits for AMD's Instinct MI355X at the April 2026 MLPerf Inference 6.0 benchmarks. The Core Scientific/AMD deal is structured around inference and cloud-provider deployments, where AMD's price-performance ratio is most competitive. AMD is betting that building infrastructure customers can only access by using AMD hardware creates a stickiness equivalent to CUDA's software ecosystem stickiness.

What does this deal mean for CORZ investors, given the company is posting $1B+ quarterly losses?

Core Scientific's net loss of $1.155 billion in Q2 2026 is almost entirely non-cash: write-downs of mining assets and depreciation on new construction. The company's operating colocation business generated $136.7 million in revenue in Q2 — more than twelve times the $10.6 million it earned from AI colocation a year earlier. With $1.8 billion in liquidity and $24 billion in total contracted revenue across its deals with CoreWeave and AMD (plus the AMD option for up to 2.5 gigawatts more), the cash-generation picture from 2027 onward looks structurally different from the headline net loss figure. The primary risk is execution: $797.5 million in capital expenditures in a single quarter reflects a construction pace that requires flawless delivery of long-lead electrical equipment and construction milestones to convert into the contracted revenue the deals promise.

What is the $14 billion figure — is it guaranteed revenue or projected?

The $14 billion represents "potential base contracted revenue" — the total value of lease payments AMD and Neocloud are obligated to make over the 15-year primary terms of the agreements, assuming all contracted capacity is delivered and all leases run to their full term. It is not revenue Core Scientific will recognize immediately; it flows in over 15 years as capacity is built, commissioned, and billed. The 2.5 percent annual escalators mean the later years of the contracts are worth more in nominal terms than the early years. AMD's option to expand to 2.5 gigawatts total — if exercised — would add another tier of contracted revenue on top of the $14 billion base figure, as confirmed by Core Scientific's Q2 2026 earnings call.