AMD Unveils MI400 AI Chips, Backed by OpenAI CEO Sam Altman
New ‘Helios’ rack system challenges Nvidia with power efficiency, performance, and price
At a high-profile event in San Jose, AMD CEO Lisa Su introduced the company’s next-generation Instinct MI400 series AI chips, expected to launch in 2026. With backing from OpenAI CEO Sam Altman, who confirmed that OpenAI will use the new chips, AMD is aiming squarely at Nvidia’s dominance in the AI data center GPU market.
Helios: Rack-Scale Revolution
The MI400 chips will form the backbone of “Helios,” a full rack-scale system engineered to act as a unified compute engine.
- This setup mirrors Nvidia’s upcoming Vera Rubin racks, offering thousands of GPUs that behave as one system.
- AMD’s approach integrates CPUs, GPUs, and networking chips using open-source UALink for better modularity and interoperability.
Altman Endorsement and Early Adoption
Sam Altman, appearing alongside Su, praised the new system’s ambitious specs.
- OpenAI, which has traditionally used Nvidia hardware, will now incorporate AMD chips, reflecting a broader industry shift toward multi-vendor strategies.
- AMD says its AI hardware has now been adopted by 7 of the top 10 AI companies, including OpenAI, Tesla, xAI, Cohere, Meta, Microsoft, and Oracle.
Competing on Price and Efficiency
AMD is leveraging cost-efficiency as a primary advantage:
- The MI355X, currently shipping, delivers 40% more tokens per dollar than Nvidia’s chips due to lower power consumption.
- According to Andrew Dieckmann, AMD’s data center GPU GM, their solution offers double-digit savings in both acquisition and operational costs.
Market Share and Strategic Investments
Despite Nvidia holding over 90% market share, AMD expects rapid gains:
- AI chip revenue hit $5 billion in FY2024, with 60% growth projected this year.
- AMD sees a $500 billion AI chip market by 2028 and has invested in 25 AI companies in the past year, including ZT Systems for server integration.
Competitive Edge in Inference
AMD is emphasizing its edge in inference tasks, a fast-growing segment of AI deployment.
- The MI355X offers 7x performance over its predecessor, featuring more high-speed memory to support larger AI models on a single GPU.
- Meta uses AMD for its Llama model inference, and Microsoft deploys AMD chips to power Copilot features.
The Road Ahead
With annual chip updates now the norm, AMD plans to compete head-to-head with Nvidia’s B100 and B200 GPUs, and its rack-scale Helios systems are expected to unlock large-scale AI computing for cloud giants.
- Oracle alone will offer clusters with over 131,000 MI355X chips.
- AMD’s open, full-stack approach positions it as a serious alternative to Nvidia’s proprietary CUDA/NVLink ecosystem.