AMD puts up to $5B of its own stock on the line to land Anthropic as the first two-gigawatt customer for its open-standard Helios AI rack

Anthropic is committing to take up to two gigawatts of AMD's next AI accelerators, starting in the first half of 2027, in a deal that could include up to US$5 billion in AMD equity flowing back to the AI lab. The chips will sit inside Helios, a full rack-scale system AMD unveiled at its Advancing AI 2026 event in San Francisco on Wednesday. Helios pairs the new MI455X Instinct GPUs with 6th-generation EPYC "Venice" CPUs. Coverage of the event frames Helios as AMD's open-standard alternative to Nvidia's NVLink and Infinity Fabric lock-in. Nvidia responded the same week with technical details on its Vera CPU and Rubin GPU, its own next platform, pitched on performance per watt for AI workloads where models run continuously and call each other.

Until now, AMD's pitch to the largest AI labs was a chip: the GPU inside a server someone else built. Helios sells a designed rack instead, with the networking, CPU, accelerator, and memory architecture pre-integrated, and the customer buys the whole thing. The MI455X is the most visible part, but the Infinity Fabric that links it to Venice and to the rest of the rack is the part that has no Nvidia equivalent on the open market.

The lock-in is NVLink, the proprietary switch fabric, the high-bandwidth memory topology, and the inference software tuned to it. Helios makes the case that AMD can ship a comparable rack on open standards, with Infinity Fabric replacing NVLink, and that buyers tired of waiting for Nvidia's allocation can buy a second supplier at the platform level.

Anthropic is the first dated, falsifiable test of that argument. Two gigawatts is roughly the continuous power draw of a small city's worth of data centre space, enough to run frontier-scale training and serve a meaningful fraction of global Claude traffic. The H1 2027 start date is concrete enough to grade AMD on. If the racks land on time and the per-watt economics hold against Nvidia's Rubin generation, the deal becomes a template. If they slip, the open-platform story takes a hit.

Microsoft is already on board. Three days before the Moscone event, AMD announced a strategic partnership with Microsoft Azure to deploy Helios at scale, add two new Venice-powered virtual machine families, and integrate Pensando DPUs and Azure Boost into AMD's rack. Azure is the second-largest cloud provider in the world by capacity and a buyer with deep ties to both OpenAI and Anthropic.

Venice itself is a meaningful shift. AMD began ramping production of the chip on TSMC's 2nm process in May, making it the first high-performance computing product on that node. The smaller transistors translate directly into more compute per watt, which is the metric frontier model buyers care about most.

Behind the launch, a second customer is already on AMD's cap table. In October 2025, AMD and OpenAI announced a multi-year deal that includes tens of billions in annual revenue and an option for OpenAI to take roughly a ten percent stake in AMD. The OpenAI commitment is older, framed in conditional language, and was not the news of this week. It does mean that by the time Helios ships, both of the largest independent AI labs will have direct AMD exposure and both will be running on a rack AMD designed rather than assembled.

AMD has not matched Nvidia on any single benchmark. The open question is whether enough of the largest AI labs will fund a second supplier at the platform layer to make the open-rack approach viable. Nvidia's answer this week is to keep moving on its own platform: Vera is a custom CPU meant to share memory with Rubin-class GPUs in ways that mirror, deliberately, what AMD is now selling. The two products are converging in shape, while diverging in ownership.

A second-order signal points the other way. Earlier this week, IBM cut its annual revenue growth forecast and partly blamed AI capex rotation, the same spending pattern that has funded Nvidia's last two fiscal years. If hyperscaler AI budgets flatten, the willingness of any single lab to commit two gigawatts to AMD becomes a more expensive bet, because the alternative, running on whatever Nvidia can allocate in the same period, gets cheaper relative to the switching cost.

AMD says the first gigawatt of Helios lands at Anthropic in the first half of 2027. That is when the open-platform story gets tested.