Compute & Market Power
AMD's $14B Power Deal Is a Distribution Strategy
AMD secured 530MW of AI capacity worth $14B over 15 years, turning energized data-center space into a channel for its chips.
AMD has secured roughly 530 megawatts of U.S. AI data-center capacity under a 15-year agreement carrying more than $14 billion in base contracted revenue. The decision turns power and liquid-cooled floor space into a distribution channel for Instinct accelerators: AMD is no longer asking customers to find somewhere to install its chips after the sale.
The product now includes somewhere to plug it in
The Core Scientific and AMD agreement covers five campuses, with deployments beginning in 2027 and an option to expand from more than 500MW to 2.5 gigawatts. The facilities will host AMD Instinct accelerators, EPYC CPUs, and ROCm software for cloud providers, model builders, and enterprises. AMD also receives market-priced warrants in Core Scientific, tying some of its upside to the landlord it is helping convert.
Normalize the base term and the scale becomes easier to compare. More than $14 billion divided by 530MW and 15 years is about $1.76 million per contracted MW-year, or roughly $147,000 per MW-month. That is revenue to Core Scientific, not AMD’s all-in compute cost, and it excludes servers, networking, and electricity. Still, the figure reveals the premium attached to deployment-ready capacity before one accelerator starts earning tokens.
The more revealing estimate comes from the build side. Data Center Knowledge reports delivered AI capacity costs $11 million to $12 million per megawatt, including construction, electrical systems, cooling, commissioning, and utility integration. Apply that range to 530MW and the initial deployment implies $5.83 billion to $6.36 billion of infrastructure. Base contract revenue is therefore about 2.2× to 2.4× the estimated delivery cost over 15 years, before operating expense and financing. The spread is not pure profit; it is the price of schedule, execution, and power certainty.
Core Scientific gains diversification and a second life for former mining sites. Its leased AI portfolio rises to about 1.1GW and more than $24 billion of potential contracted revenue. AMD gains something strategically scarcer than silicon: a credible answer when customers ask where hundreds of megawatts of MI-series systems can run. That follows the logic in AMD’s separate $5 billion Anthropic commitment, which paired a model-lab relationship with a future accelerator channel. The company is assembling demand and deployment around its roadmap rather than waiting for ROCm adoption alone to pull hardware through.
This is not AMD becoming a cloud provider. It is AMD underwriting a path for other providers to deploy. Analyst Matt Kimball captured the distinction: the company is not merely selling accelerators but selling a deployment path. When power, substations, cooling equipment, and interconnection queues bind supply, chip specifications cease to be a complete product.
The agreement also clarifies why Nvidia’s reported Texas lease with Hut 8 matters beyond its headline. Both chipmakers are reaching beyond the package and into energized real estate. Nvidia reportedly takes the more vertically circular role as tenant, architect, and supplier; AMD uses a partner to make capacity available to its ecosystem. Reuters reports Nvidia’s maximum lease value depends on renewals, while AMD’s 2.5GW is also expansion optionality. The competitive question is shifting from whose chip benchmarks best to whose customers can deploy first.
Buy the path, but price the wait
Who should change course? Cloud providers and large model builders planning 2027–2028 clusters should add the Core Scientific footprint to AMD evaluations, especially if Nvidia availability or pricing makes an alternative valuable. Enterprises buying tens of racks rather than hundreds of megawatts should not contract around a press release; they should ask which downstream providers will actually expose this capacity, under what service-level agreement, and with which MI generation.
The cost is not just migration to ROCm. Teams must validate kernels, collective communication, inference runtimes, observability, and model-serving behavior on the hardware that will ship in 2027. A nominally cheaper accelerator can become expensive when engineering time, lower utilization, or missing libraries stretch deployment. The lesson from Intel’s recent supply ceiling is that demand signals do not automatically become deliverable systems. Power reservations and chip roadmaps must arrive together.
The expansion number also deserves discipline. AMD has the opportunity to scale to 2.5GW, but the initial commitment is roughly 21.2% of that ceiling: 530MW divided by 2,500MW. Operators should model 530MW as the evidence and 2.5GW as optionality. The same skepticism applies to the $14 billion base revenue, which accrues across 15 years rather than landing as a single booking. Core Scientific’s release confirms both the 2027 start and the warrant structure, reinforcing that this is a staged infrastructure partnership rather than instant capacity.
Execution is the strongest counterpoint. Capacity starts in 2027 and rolls through 2028 across Pecos and four other campuses. Construction delays, transformer shortages, financing costs, utility constraints, or an AMD product slip could erode the advantage. Core Scientific’s mining heritage supplies powered land and operating experience, but high-density AI halls demand different networking, cooling, uptime, and customer-support capabilities.
The warrant structure adds another tension. Equity alignment can encourage both sides to deliver, yet it also makes the chip vendor partly invested in the infrastructure customer’s market value. Builders should insist on transparent pricing and portability so financial alignment does not become a reason to lock workloads to one facility or stack.
What evidence would change the verdict? First, signed downstream customers naming capacity, dates, and accelerator generations. Second, measured ROCm utilization and reliability on production clusters, not lab benchmarks. Third, energization milestones at the five campuses. If those arrive, AMD’s infrastructure channel becomes a durable counterweight to Nvidia. If they do not, the deal is a large option on power wrapped around an uncertain software conversion.
For now, the operator decision is conditional but real. Teams that want a second source should begin porting and benchmarking before the halls open, while preserving a model-serving abstraction that can move between vendors. AMD has bought the runway. Its customers still have to prove they can land on it.