Anthropic Plans Up to 2 GW of AMD AI Capacity. Nvidia Still Has the Lead.

Anthropic has committed to use up to two gigawatts of AMD MI450-series GPU capacity in the Helios rack platform, with the first gigawatt scheduled to begin in 2027. The commitment gives AMD a consequential prospective deployment for its next AI system. It also puts a hard operational question in front of the company: whether Helios can arrive on schedule and translate planned capacity into deployed revenue while Nvidia continues to sell AI infrastructure at a vastly larger scale.

A rack-scale bid for a larger role

Helios is AMD’s attempt to sell a complete rack rather than just an accelerator: it combines EPYC Venice CPUs, MI450 GPUs and Pensando networking. That design matters for a customer planning capacity at gigawatt scale, where computing components and networking must operate as one system. Anthropic’s commitment is therefore more than an order for chips; it is an early test of AMD’s ability to deliver a coordinated AI infrastructure platform.

AMD enters that test with fast growth. It reported $11.54 billion in second-quarter revenue, including $6.72 billion in data-center revenue, which rose 107% from a year earlier. AMD has also said it has committed deployments with OpenAI, Meta and Anthropic, which CEO Lisa Su called strategic anchor customers. The Anthropic plan adds a large, dated capacity target to that customer story, even though deployment is still set for 2027.

Nvidia’s lead is a platform, not a single chip

Nvidia’s position rests on more than its current revenue. The company said Blackwell is shipping to every major hyperscaler, while its wider stack includes networking and CUDA software alongside its processors. Nvidia also has the capacity leverage that comes with controlling a sought-after platform: one account characterized its ability to guarantee allocation and use access to capacity to secure major AI customers as an advantage AMD cannot yet match.

AMD is working on the software side of that contest. It said Rackham runs more than 3 million models out of the box, and that open-source contributions rose more than tenfold over the prior year. But Nvidia is already moving its roadmap forward: it said Vera Rubin production begins in the third quarter and claimed the platform can deliver up to 35 times Blackwell’s inference throughput. That performance figure is Nvidia’s claim, not an independent comparison, but it illustrates the pace AMD must contend with as it seeks to establish MI450.

The commitment now has to become a deployment

The immediate evidence supports momentum, not parity. AMD has guided to approximately $13 billion in revenue for the next quarter, and its data-center business has grown rapidly. Nvidia, however, is operating from a current data-center revenue base more than 11 times AMD’s reported figure. Anthropic’s commitment gives AMD a route into a very large future deployment, but it does not make the 2027 capacity operational today.

What Helios must prove before 2027

• MI450 systems must ship as planned rather than remain a capacity commitment.
• The rack’s software and hardware must perform for the workloads Anthropic intends to run.
• The planned capacity must convert into recognized revenue as deployments begin.

That is the concrete execution test for AMD. A successful Helios and MI450 rollout would show that major AI customers can deploy another rack-scale option alongside Nvidia’s systems. Delays in shipping, software performance or the revenue ramp would leave the Anthropic commitment as an important intention rather than proof of a durable competitive shift.