AMD, 140억 달러 규모의 전력 경쟁에서 코어 사이언티픽으로부터 AI 데이터센터 용량 확보

AMD and Core Scientific (NASDAQ: CORZ) announced a 15-year infrastructure partnership on Tuesday valued at over $14 billion in base contracted revenue. The agreement grants AMD reserved access to approximately 530 megawatts of U.S. AI data center capacity across five sites, with an option to expand the footprint to 2.5 gigawatts. The deal underscores a broader shift in the AI arms race: securing power, land, and purpose-built infrastructure has become as critical as competing on silicon.

Announced alongside Core Scientific’s second-quarter 2026 earnings, the agreement marks one of the largest infrastructure commitments in AMD’s history. It follows closely on the heels of the company’s Advancing AI 2026 conference in San Francisco, where AMD unveiled the Instinct MI455X GPU, the EPYC “Venice” server processor, and Helios, its first rack-scale AI platform.

**AI Compute Has Become a Real Estate Problem**

For much of AI’s recent history, competitive advantage boiled down to a single metric: chip performance. AMD’s partnership with Core Scientific signals a shift in that dynamic. Power capacity, strategically located land, and long-term contractual access to AI-ready infrastructure have emerged as standalone strategic assets—capable of being secured years ahead of competitors.

“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.

Under the initial tranche, Core Scientific will deliver approximately 530 megawatts across five campuses. AMD holds direct triple-net leases covering roughly 380 megawatts at facilities in Pecos and Hunt County, Texas, and Muskogee, Oklahoma. An additional ~150 megawatts at sites in Auburn, Alabama, and Dalton, Georgia will serve an unnamed Neocloud customer, with AMD providing full credit support across the entire 15-year lease term. The agreement also grants AMD reservation rights for up to 1.9 gigawatts of future capacity through late 2028, bringing the total potential footprint to 2.5 gigawatts.

According to Core Scientific’s earnings call, the first AMD megawatts at the Pecos facility are scheduled to come online in the first half of 2027. Approximately 265 megawatts—roughly half of the initial commitment—is expected to be delivered by year-end 2027, with the remainder coming online in 2028.

As part of the transaction, AMD received market-priced warrants to purchase up to 30 million shares of Core Scientific common stock. Approximately 6.5 million warrants vest immediately, with the remainder tied to specific megawatt-delivery milestones.

**How Does AMD's Rack-Scale Helios Platform Compare to Nvidia's Vera Rubin?**

The partnership is structured to deploy AMD’s MI455X GPU and Helios rack-scale platform—unveiled July 22–23 at Advancing AI 2026—at a scale requiring dedicated, purpose-built infrastructure, AMD noted at the conference.

A single Helios rack integrates 72 Instinct MI455X GPUs based on AMD’s CDNA 5 architecture, paired with 18 sixth-generation EPYC “Venice” CPUs, Pensando networking, and the ROCm open-source software stack. At the rack level, the system delivers up to 2.9 exaFLOPS of FP4 compute and 1.4 exaFLOPS of FP8, supported by 31 terabytes of HBM4 memory and 1.67 petabytes per second of aggregate memory bandwidth. Individually, each MI455X accelerator features 432 gigabytes of HBM4 and delivers approximately 40 petaFLOPS of FP4 compute.

AMD claims Helios delivers 15% more FP4 compute, 50% greater HBM memory capacity, and 50% more scale-out bandwidth than Nvidia’s Vera Rubin NVL72 rack system. While accurate, the comparison omits key operational variables: Helios consumes approximately 140 kilowatts per rack, versus 190 to 230 kilowatts for Nvidia’s setup. This power-density advantage comes with a caveat. AMD’s UALink-over-Ethernet interconnect, which handles GPU-to-GPU communication within the rack, relies on third-party switching silicon from vendors including Astera Labs and Enfabrica—technology that remains limited in broad availability as of mid-2026.

Buyers must also account for architectural lock-in. The MI455X’s HBM4 memory utilizes a 2,048-bit routing interposer that is physically incompatible with Nvidia’s HBM3e-based platforms. Selecting Helios precludes hardware reuse on Nvidia systems; migrating to a different vendor would require a complete infrastructure replacement.

**What $14 Billion Buys — and What It Cannot**

The headline figures obscure a critical reality: securing 530 megawatts of physical infrastructure does not guarantee equivalent commercial utilization. AMD’s commitment will only realize its value if its software ecosystem can attract and retain developers willing to prioritize AMD hardware over Nvidia for mission-critical workloads.

Nvidia controls an estimated 80–88% of the AI accelerator market by revenue as of mid-2026, a figure corroborated by Bloomberg Intelligence, TrendForce, and IDC. This dominance stems less from hardware specifications than from CUDA, Nvidia’s parallel computing platform. Over 18 years of investment has cultivated a developer base exceeding 4 million and entrenched CUDA across every major AI framework. As of mid-2026, foundational libraries such as cuDNN, TensorRT-LLM, and FlashAttention 3 lack direct ROCm equivalents.

AMD’s ROCm platform has advanced considerably. The May 2026 stable release, ROCm 7.2.4, introduced native PyTorch support. For standard inference workloads running PyTorch and vLLM, AMD’s Instinct MI355X and MI455X hardware achieves approximately 90–95% of Nvidia H100 throughput. Training performance lags further behind: independent analysis indicates AMD’s rocBLAS and MIOpen libraries trail Nvidia’s cuDNN and Tensor Core optimizations by roughly 20–30% on fine-tuning tasks.

To bridge this software divide, AMD announced a separate partnership at Advancing AI 2026 with Anthropic. The agreement covers the deployment of up to 2 gigawatts of MI450 Series GPUs in Helios systems and establishes a multiyear engineering collaboration wherein Anthropic’s Claude models will optimize workloads for AMD hardware and accelerate ROCm development. As part of the arrangement, AMD pledged a strategic equity investment of up to $5 billion in Anthropic.

The strategic trajectory is evident, though execution remains ongoing. AMD’s “OneROCm” initiative seeks to unify the software stack across its entire processor lineup, while the Anthropic partnership leverages AI itself to compress the CUDA ecosystem gap. The pace at which Core Scientific’s infrastructure capacity reaches utilization will hinge largely on the maturity of these software initiatives.

**Core Scientific's Bitcoin Exit, by the Numbers**

For Core Scientific, the AMD agreement validates a strategic pivot initiated in 2024. Once predominantly a Bitcoin mining operator, the company reported second-quarter 2026 revenue of $164.2 million, a 109% increase from $78.6 million in the same period of 2025. Colocation services fueled the expansion, generating $136.7 million in Q2 2026 compared to $10.6 million a year prior, accounting for 83% of total revenue.

By mid-July, Core Scientific was billing customers for 437 megawatts of capacity, translating to approximately $635 million in average annualized colocation GAAP revenue. When combined with existing CoreWeave contracts covering roughly 590 megawatts, the company now manages approximately 1.1 gigawatts of total leased customer power capacity, representing over $24 billion in potential contracted revenue.

“We are proud to establish a strategic relationship with AMD and support the continued deployment of its next-generation products,” said Core Scientific CEO Adam Sullivan.

The financial mechanics of this transition remain complex. Core Scientific reported a GAAP net loss of approximately $1.2 billion in the second quarter, primarily attributable to non-cash charges related to its $3.3 billion Senior Secured Notes offering and warrant accounting. During the quarter, the company also retired a $1 billion term loan and invested $233 million to acquire the Hunt County, Texas campus—all expenditures executed ahead of the AMD contract obligations. Capital expenditures surged to $797.5 million in Q2, up from $389.2 million in Q1, as construction accelerates across five expansion campuses.

**How Wall Street Is Reading the Numbers**

Markets reacted with caution rather than celebration. Core Scientific shares initially surged as much as 10% in premarket trading following the announcement, before retreating approximately 3% during intraday trading.