AMD's Next-Gen Chips Reported to Secure Major Anthropic Order

Continuing tightness in AI compute supply is accelerating the reshaping of chip procurement landscapes for large technology companies, benefiting AMD.

According to media reports citing market sources, Anthropic plans to purchase AMD's next-generation Instinct MI450 acceleration chips for its server infrastructure construction. Analyst Ben Bajarin posted on social media stating, "It is entirely reasonable for AMD to secure an Anthropic order given the current compute shortage situation:"

"This will become another top-tier AI enterprise signing a heavyweight partnership with AMD, following OpenAI and Meta, further consolidating AMD's competitive position in the AI data center market.

The rumored order points to AMD's upcoming MI450 chip. Featuring the CDNA 5 architecture, official data shows its FP8 compute power reaches 40 PFLOP, doubling that of the previous MI350 series. Memory capacity has increased from 288GB HBM3e to 432GB HBM4, a rise of approximately 50%, while memory bandwidth has surged dramatically from 8 TB/s to 19.6 TB/s. For AI enterprises requiring large-scale deployment of inference and training tasks, this specification upgrade holds significant practical value.

Supply Tightness Drives Diversified AI Customer Strategies

Persistent global AI chip supply chain constraints form the core background of this transaction rumor. Bajarin noted, "All available and feasible compute resources are being purchased." This assessment aligns closely with current market dynamics—ranging from hyperscale cloud providers to startups, all parties are actively securing next-generation compute resources.

AMD has already positioned itself for this trend. The company previously disclosed that multiple "OpenAI-scale" customers are queuing for its existing and future Instinct series accelerators. Among them, Meta has signed a purchase commitment of up to 6 gigawatts (GW), covering multiple generations of Instinct chips. Should the Anthropic order materialize, it would mean AMD's primary AI customer base will expand further, providing positive support for its revenue prospects.

AMD has directly anchored the MI450 series specifications against NVIDIA's Vera Rubin platform. Internal comparison data indicates that MI450 leads competitors by 1.5 times in memory capacity and 1.5 times in Scale-Out Bandwidth, while matching them in FP4/FP8 floating-point compute, memory bandwidth, and inter-chip expansion bandwidth.

In terms of specific parameters, MI450 offers 40 PFLOP FP8 compute, 432GB HBM4 memory capacity, 3.6 TB/s expansion interconnect bandwidth, and 300 GB/s per-card node-to-node bandwidth. AMD positions this combination as a core component of rack-level AI ecosystems and markets it under the "Helios" brand.

Anthropic Compute Strategy: Multiple Bets, Risk Diversification

Notably, Anthropic is not betting on a single supplier for compute sources. Currently, its infrastructure integrates NVIDIA GPUs and Amazon Trainium chips, and it recently signed strategic partnerships with Google and Broadcom, expecting to receive multi-gigawatt level next-generation TPU compute starting in 2027. In an official statement, Anthropic indicated that these TPU resources will support its cutting-edge Claude models and global customer demands.

Furthermore, analysts suggest that Anthropic's collaboration with Broadcom, experienced in custom ASICs, may lay the groundwork for potential future in-house development of dedicated Claude chips. However, this direction remains speculative at present, with no specific information disclosed. This diversified compute strategy demonstrates Anthropic's strong willingness to secure high-quality suppliers, adding leverage to AMD's negotiations.

AMD Simultaneously Advances Government-Level AI Cooperation

Beyond commercial clients, AMD is recently expanding into the government sector. AMD and the French government announced deepened cooperation to support France's national AI strategy, covering initiatives to promote local AI innovation, expand coverage of open compute resources, and participate in the construction of France's first exascale supercomputer, "Alice Recoque." AMD stated it will open its high-performance computing platforms and open software ecosystem to the French AI community.

Overall, should the Anthropic transaction ultimately materialize, it will further validate AMD's commercial capabilities in the high-end AI chip market and provide strong support for market expectations ahead of formal shipments of its MI400 series.