Advanced Micro Devices disclosed a $14 billion infrastructure agreement with Core Scientific, the publicly traded bitcoin miner turned AI hosting operator, with embedded rights to scale capacity to 2.5 gigawatts. The deal marks the largest single capital commitment by a semiconductor company to secure physical compute outside hyperscaler partnerships and attaches AMD's inference roadmap directly to power-constrained, campus-scale infrastructure. Core Scientific shares rose 5% in after-hours trading.
The contract structure allows AMD to deploy MI300-series accelerators across Core Scientific's refitted colocation facilities, most of which were previously optimized for ASIC bitcoin mining before the 2022 restructuring. The 2.5 GW option — equivalent to roughly 1.7 million MI300X GPUs at full thermal load — provides AMD with a committed utility-grade power envelope that cannot be reallocated to competing workloads. Core Scientific retains operational control of the sites but cedes compute allocation authority to AMD under multi-year capacity reservations. No buildout timeline was disclosed, though the miner's existing North Carolina and Texas facilities already carry substations rated above 500 MW each.
The arrangement solves two problems. AMD gains reserved capacity for inference workloads that do not require the low-latency networking topology of training clusters, letting the company place models closer to enterprise customers without negotiating rack-by-rack with wholesale providers. Core Scientific converts stranded mining assets into long-duration contracts with investment-grade counterparties, eliminating bitcoin price exposure while preserving the power purchase agreements that define the business. The $14 billion figure represents a mix of capital expenditure, capacity reservations, and revenue commitments spread across the contract term, though the precise allocation was not broken out in the disclosure.
This changes the competitive map for AI infrastructure. Nvidia has relied on hyperscaler capex and sovereign GPU clusters to absorb Blackwell production, leaving inference deployments to third-party resellers and edge providers. AMD now controls a vertically integrated stack — silicon, firmware, and gigawatt-scale hosting — that can be sold as a bundled service to enterprises unwilling to build internal data centers or negotiate Azure/GCP rate cards. The move also signals that AI compute is bifurcating: training remains concentrated in hyperscaler hands, but inference is becoming a margin game fought over power costs, latency, and dedicated capacity.
Operators should track three items over the next 90 days: Core Scientific's earnings call for capex pacing and site activation schedules, AMD's enterprise AI bookings as a percentage of data center revenue, and any competing deals from Nvidia or Intel with wholesale colocation providers. If Core Scientific begins retrofitting additional sites beyond the initial tranche, the 2.5 GW option moves from contractual to operational.
The deal does not make AMD a hyperscaler. It makes them the only chip company with reserved power at a scale that lets them guarantee delivery timelines independent of Azure's provisioning queue.