AMD and Anthropic Sign Hundreds of Billions in AI Chip Deal, Challenging Nvidia's Dominance

AMD and Anthropic recently signed an AI server cooperation agreement worth hundreds of billions of dollars. Anthropic will purchase up to 2 gigawatts of AMD's next-generation Instinct MI450 chips, with deliveries starting in the first half of 2027. AMD plans to invest up to $5 billion in Anthropic after meeting specific deployment milestones. This is AMD's first direct investment in an AI company.

This deal marks Anthropic's systematic effort to build a compute supply chain independent of Nvidia. To date, Anthropic has secured compute resources across four supply chains: up to 5 gigawatts from Amazon AWS (based on Trainium chips), up to 5 gigawatts of TPU capacity from Google, 220,000 Nvidia GPUs (about 300 megawatts) at SpaceX's Colossus 1 facility, plus 2 gigawatts of AMD MI450, totaling over 12 gigawatts of committed AI compute capacity. For reference, OpenAI, SoftBank, and Oracle's 'Stargate' project plans multiple sites totaling about 7 to 9 gigawatts, with the first site running only about 0.3 gigawatts.

Anthropic CFO Krishna Rao previously stated that expanding compute is the only way to keep up with exponential demand. According to FutureSearch's calculations based on Anthropic's Series H announcement in May, its ARR rose from $9 billion in December 2025 to $47 billion by mid-May 2026, with daily incremental annualized revenue of about $96 million, and Q2 GAAP revenue is expected to exceed $10.9 billion.

The MI450 uses TSMC's 2nm process, achieving 40 PFLOPS at FP4 precision and 20 PFLOPS at FP8, double the current flagship MI355X. Memory is upgraded to HBM4, with 432GB per chip and 19.6 TB/s bandwidth. AMD also plans the Helios rack solution: 72 GPUs integrated in one rack, totaling 31TB of HBM4 memory and 260 TB/s of vertical scaling bandwidth. At the Advancing AI 2026 conference on July 23, AMD disclosed that Helios produces 30% more tokens per dollar invested than competitors.

The CUDA ecosystem is Nvidia's true moat, with about 6 million developers and over 400 accelerated libraries. However, Anthropic co-founder Tom Brown revealed at AMD's Advancing AI 2026 conference that his engineering team, using only the Claude model, automatically completed full adaptation and performance tuning of the AMD Instinct MI355 chip and ROCm platform over a weekend. Brown said the team initially expected 'a big project,' but the actual experience was completely different; an engineer had Claude perform the adaptation tasks and let the process run over the weekend, returning on Monday with performance curves.

An analyst at the conference commented on the spot: 'We have passed the era of the CUDA moat.' If this path is validated, Anthropic's compute diversification layout will become a reference: Google TPUs handle main training, AWS provides elastic scaling, Nvidia GPUs ensure critical workloads, and AMD MI450 fills the cost-performance gap.

The timing of this deal is closely tied to Anthropic's IPO process. On June 1, Anthropic confidentially filed an S-1 with the SEC, and Series H closed on May 28 with a post-money valuation of $965 billion. By July 15, Morgan Stanley, Goldman Sachs, and JPMorgan had begun arranging meetings between management and potential investors. If all goes well, Anthropic could be listed as early as October 2026.

AMD's previous deals with OpenAI and Meta both included substantial equity warrants. In October 2025, OpenAI signed a 6-gigawatt procurement agreement with AMD, receiving warrants for up to 160 million AMD shares. In early 2026, Meta reached a 6-gigawatt deal with AMD, obtaining warrants equivalent to about 10% of AMD's equity. However, the Anthropic deal includes no warrants. Barclays analyst Tom O'Malley interpreted this as a sign that in a tight supply market, chipmakers no longer need to give away equity to lock in large customers.

Nvidia's response has been equally sharp: investing $10 billion in OpenAI with an option to increase to $100 billion, and investing $5 billion in Intel for joint development. The competition among chip companies has extended from transistor density to balance sheets, and 'investment for orders' is solidifying as a new competitive rule in the AI chip market.

The MI450 is still a 'PPT chip.' The yield and production ramp of TSMC's 2nm process will not be clear until late 2026 or early 2027. At the end of January, SemiAnalysis released signals that MI450 progress was lagging, causing AMD's stock to drop 6% in a single day. Wells Fargo analysts later countered, noting that tape-out was complete, samples of the same-process Venice EPYC had shipped, and mass production was still on track for the second half of 2026. But the market reaction shows that the chip's schedule cannot tolerate any surprises.

Software ecosystem challenges remain. Although ROCm is improving rapidly, it still lags behind CUDA in large-scale distributed training scenarios. Key inference frameworks like vLLM are still experimental. The Claude adaptation of MI355 over a weekend is impressive, but MI355 is a mature chip, while MI450 is a new architecture, and full-stack adaptation is not on the same level of difficulty. Deployment infrastructure is another hurdle: AMD does not operate a cloud platform, and there are no public details on who will deploy the 2-gigawatt Helios rack solution on the customer side or where.

Geopolitics introduces uncertainty. The U.S. Department of Commerce has repeatedly tightened AI chip export controls between 2025 and 2026. If policies change again before 2027, the entire deployment plan could be forced to adjust. Key milestones include: whether AMD announces formal tape-out and performance validation results for MI450 in Q4 2026, management's updated timeline for MI450 in AMD's Q2 earnings report on August 4, and whether Anthropic can receive the first batch of MI450 as planned in the first half of 2027.