Share with your CTO
AMD is making a $14 billion infrastructure land grab, signing a 15-year deal with data center operator Core Scientific for 530 megawatts of U.S. AI compute capacity across five sites, with options to expand to 2.5 gigawatts. The deal is built to house AMD’s new Helios rack platform, which packs 72 MI455X GPUs delivering 2.9 exaFLOPS of FP4 compute per rack. First capacity comes online at Pecos, Texas in early 2027. The strategic read: AMD has concluded that winning AI infrastructure share requires controlling power and land, not just shipping faster chips.
What this means for your business
The story this deal tells isn’t really about Core Scientific. It’s about what AMD now believes it takes to compete with Nvidia, and whether that belief is correct matters directly to every CTO currently weighing a multi-year AI infrastructure commitment. If you’re already deep in Nvidia’s CUDA ecosystem, this deal doesn’t move you today. If you’re still making architectural choices, AMD just removed one of the most practical objections to choosing its hardware: “I can’t find the power capacity to run it.”
AMD’s Helios rack posts genuinely strong specs on paper, 15% more FP4 compute and 50% more HBM memory than Nvidia’s Vera Rubin NVL72, at lower power draw per rack (roughly 140 kilowatts versus 190 to 230 kilowatts). But the hardware comparison flatters AMD more than the full picture warrants. ROCm, AMD’s software platform that competes with CUDA, reaches about 90 to 95% of Nvidia H100 throughput on inference workloads running PyTorch and vLLM. For training, the gap is wider, roughly 20 to 30% behind on fine-tuning, because Nvidia-specific libraries like TensorRT-LLM and FlashAttention 3 have no full ROCm equivalents yet. AMD is using its engineering partnership with Anthropic to close that gap faster, including using Claude itself to accelerate ROCm development. That’s an interesting strategy. It’s not a solved problem.
There’s a lock-in consequence that belongs in any vendor evaluation right now. The MI455X uses HBM4 memory on a 2,048-bit routing interposer that’s physically incompatible with Nvidia’s HBM3e-based systems. Deploying Helios isn’t a hedge against Nvidia; it’s a full architectural commitment. That’s not a disqualifier, but it’s a decision frame that should be explicit in procurement conversations, not buried in a footnote. The recurring mistake in enterprise infrastructure is treating hardware selection as reversible when the actual switching cost is a full system replacement.
AMD’s accumulating supply-side commitments (OpenAI at 6 gigawatts, Anthropic at 2 gigawatts, Microsoft Azure, and now Core Scientific) mean a credible alternative to Nvidia exists for the first time at genuine scale. The question a CTO needs to answer before 2027 capacity decisions close isn’t whether AMD can compete on specs. It’s whether ROCm will be mature enough for your specific training workloads by the time your hardware actually lands. If inference is your dominant use case, AMD is a real option today. If large-scale training is the core workload, the software timeline is still the variable that decides the bet.
Concept deep-dive: Triple-net lease
A triple-net lease makes the tenant responsible for property taxes, insurance, and maintenance on top of base rent, think of it as the difference between renting a furnished apartment and owning the building’s operating costs without owning the title. In this context, AMD holds direct triple-net leases on roughly 380 megawatts of Core Scientific capacity. That structure gives Core Scientific highly predictable, long-duration revenue, which is why it could justify raising $3.3 billion in secured debt to fund construction before customers were signed.
Based on reporting from AMD’s $14B Data Center Bet on Core Scientific Targets Nvidia’s AI Infrastructure Lead, originally published 2026-07-28 11:58:00.
