On July 22, AMD and Anthropic announced a strategic partnership: Anthropic will deploy up to 2GW of AMD Instinct MI450-series GPUs within AMD's Helios rack-scale solutions, with the first 1GW landing in the first half of 2027. In return, AMD committed a strategic equity investment of up to $5 billion in Anthropic.
Piecing this together with events from the past two weeks, the full picture emerges:
- February: AMD × Meta partnership — Meta to deploy 6GW of AMD AI chips over five years, with AMD granting Meta equity worth roughly 10% of the company
- July 20: AMD × Microsoft expanded partnership — Azure data centers will deploy AMD Helios rack-scale systems, shipping starting in H2 2026
- July 22: AMD × Anthropic — 2GW, first 1GW in H1 2027, $5 billion equity
- At the same time: OpenAI, Oracle, and Meta are also on the Helios deployment list
What $5B in equity plus 2GW of GPU capacity means
Based on rough industry estimates, building 1GW of data center capacity costs roughly $40–60 billion, so 2GW corresponds to at least $80 billion in scale. AMD is trading $5 billion in cash for hardware orders in the $80–120 billion range plus engineering feedback from Anthropic (whose Claude models will be used to optimize AMD's ROCm software stack and engineering development). AMD can absorb this math.
Engineering details
This is not a single-point GPU deal but a complete rack-scale platform: AMD Instinct MI455X GPU + AMD EPYC "Venice" CPU + AMD Pensando networking + the ROCm software stack. Anthropic was already using MI355X, meaning this deployment is incremental rather than a "switch from NVDA" — a clearly deliberate risk-diversification strategy.
The other side of the industry
NVIDIA's Vera Rubin is advancing in the same period: Vera CPUs were delivered to OpenAI, Anthropic, and SpaceX in June, with the full platform spanning 350 manufacturing sites across 30 countries. Both GPU camps are accelerating simultaneously, with Anthropic sitting in the middle pursuing "dual-source procurement." This is the first time in AI compute history that GPUs are no longer a single source in actual engineering practice, not just on paper.
Signals worth tracking
1. AMD Helios has already eaten into OpenAI's $2 trillion data center plan (disclosed on 7/22: OpenAI plans new data centers with total spending through 2030 revised up sharply). At this scale of compute expansion, OpenAI will not let a single vendor take it all. 2. Anthropic's deep coupling with AMD means Claude models will be optimized for ROCm from the ground up — the Claude-on-AMD path should bring Anthropic's training/inference efficiency on AMD GPUs closer to NVDA levels, which will feed into Claude API pricing and SLAs. 3. The bigger picture: Anthropic's $5B equity deal with AMD belongs to the same playbook as Microsoft-OpenAI and AWS-Anthropic — large model companies and large compute providers binding to each other. This is the core narrative of the 2026 compute landscape.
Link: https://www.amd.com/zh-cn/newsroom/press-releases/amd-anthropic-strategic-partnership.html (AMD Chinese press release, with full deal figures and timeline)
Caveat: the earliest English leak of this news was a SemiAnalysis X post on 7/22, posted 9 minutes ahead of the announcement — no surprise in the industry. Capital markets showed little reaction this time: the new information was not "Anthropic + AMD" but the specific contract terms — 2GW scale + $5 billion equity.