AMD announced its Helios server rack on July 23 at an event in San Francisco. The system directly competes with Nvidia’s next-generation offerings. It contains AMD’s MI350X and MI355X GPUs, which are designed for AI inference—the part of AI computing where scale and speed matter most.

Helios will begin shipping at the end of Q3 2026. OpenAI plans to deploy it at massive scale starting later this year, accelerating throughout 2027. That is a significant customer win and a vote of confidence in AMD’s approach.

AMD’s Play

Nvidia controls 80-95% of the AI GPU market. That dominance is nearly absolute. AMD’s challenge is not to dethrone Nvidia but to take inference chips where Nvidia is weakest.

Training a model requires raw compute horsepower. Inference requires efficiency and scale. A data center runs thousands of inference queries. Power consumption matters. Latency matters. Cost per query matters.

AMD is marketing Helios as the inference workhorse. The company also landed a deal to sell up to 2 gigawatts of its Instinct MI450 chips to Anthropic beginning in H1 2027. That is another major customer win.

Nvidia’s Response

Nvidia will likely announce faster versions of its chips or better pricing. The company still owns the training market. New models are trained on Nvidia hardware almost exclusively. But the inference market is opening up.

Nvidia’s margins may compress slightly as competition increases. The company will adapt. Nvidia has been here before.

What Happens Next

Expect AMD to continue improving performance and price. OpenAI and Anthropic will deploy Helios and MI450 alongside Nvidia chips. The market is large enough for both. But Nvidia’s pricing power will diminish.

AMD’s Helios is the most credible Nvidia challenge to date. The market is ready for competition.