AMD and OpenAI CEO Sam Altman jointly released the next generation of AI chips

AMD CEO Lisa Su

AMD announced new details on its next-generation AI chip Instinct MI400 series on Thursday, which will be shipped next year.

AMD said that the MI400 chips can be assembled into a complete server rack called Helios, which enables thousands of chips to be connected together in the form of a "rack-level" system.

"For the first time, we designed every part of the rack as a unified system," AMD CEO Lisa Su said at the launch event in San Jose, California on Thursday.

OpenAI CEO Sam Altman took the stage with Lisa Su and said that his company would use AMD chips.

"When you first told me the specs, I thought, this is impossible, it sounds crazy," Altman said. "This is going to be something."

AMD's rack-scale setup will make the chip look like a unified system to users, which is critical for most AI customers, such as cloud service providers and companies developing large language models. These customers need "hyperscale" AI computer clusters that cover entire data centers and consume a lot of electricity.

"You can think of Helios as a rack that really acts like a single large computing engine," Lisa Su said, comparing it to Nvidia's Vera Rubin rack expected to be released next year.

Sam Altman, CEO of OpenAI

AMD's rack-scale technology also enables its latest chip to compete with Nvidia's Blackwell chip, which has 72 graphics processing units (GPUs). Nvidia is AMD's main and only competitor in the field of large data center GPUs for developing and deploying AI applications.

AMD said that as an important customer of Nvidia, OpenAI has been providing feedback on its MI400 roadmap. With the MI400 chip and this year's MI355X chip, AMD plans to compete on price with rival Nvidia. The chips will cost less to run because they use less power, and AMD is undercutting Nvidia's market share with "aggressive" pricing, a company executive told reporters on Wednesday.

Nvidia has dominated the data center GPU market so far in part because it was the first company to develop the software AI developers need to make chips originally designed to display graphics for 3D games work. Before the AI ​​boom of the past decade, AMD focused on competing with Intel in server CPUs.

Lisa Su said AMD's MI355X can outperform Nvidia's Blackwell chip even though Nvidia uses its "proprietary" CUDA software.

"It shows that we have really powerful hardware, which we always knew, but it also shows that the open source software framework has made great progress," Lisa Su said.

AMD shares are flat so far in 2025, suggesting Wall Street doesn’t yet see it as a major threat to Nvidia’s dominance.

AMD “Helios” AI Rack

AMD’s AI chips will be cheaper to run and cheaper to acquire, Andrew Dieckmann, general manager of AMD’s data center GPUs, said Wednesday.

“Overall, we have a significant price advantage, and combined with our competitive performance advantage, we can achieve a pretty significant double-digit percentage cost savings,” Dieckmann said.

In the coming years, large cloud companies and countries are preparing to spend hundreds of billions of dollars to build new data center clusters around GPUs to accelerate the development of cutting-edge AI models. This year alone, planned capital expenditures by large tech companies include $300 billion.

AMD expects the total AI chip market to exceed $500 billion by 2028, although it hasn’t said how much of the market it will capture — Nvidia currently has more than 90% of the market, according to analyst estimates.

Both companies have pledged to release new AI chips every year, rather than every two years, highlighting the intensity of the competition and the importance of cutting-edge AI chip technology to companies such as Microsoft, Oracle and Amazon.

AMD has acquired or invested in 25 AI companies in the past year, including earlier this year in server maker ZT Systems, which developed the technology AMD needs to build rack-scale systems, Su said.

"These AI systems are getting extremely complex, and full-stack solutions are really critical," Su said.

Where AMD is selling now

The most advanced AMD AI chip being installed by cloud service providers right now is its Instinct MI355X, which the company said it began shipping in volume last month. AMD said cloud service providers will be able to rent the chip starting in the third quarter.

Companies building large data center clusters for AI want an alternative to Nvidia not only for lower costs and flexibility, but also to meet growing demand for “inference” — the computing power needed to actually deploy chatbots or generative AI applications — which require more processing power than traditional server applications.

“What’s really changed is that the demand for inference has grown significantly,” said Lisa Su.

AMD officials said Thursday that they believe its new chip outperforms Nvidia’s at inference. That’s because AMD’s chip is equipped with more high-speed memory, enabling larger AI models to run on a single GPU.

AMD says the MI355X has seven times the computing power of its predecessor. The chips will be able to compete with Nvidia’s B100 and B200 chips, which Nvidia began shipping late last year.

AMD said its Instinct chips have been adopted by seven of the 10 largest AI customers, including OpenAI, Tesla, xAI and Cohere.

Oracle plans to offer its customers clusters of more than 131,000 MI355X chips, AMD said.

Meta officials said Thursday they are using clusters of AMD CPUs and GPUs to run inference for its Llama model and plan to buy AMD's next-generation servers.

Microsoft representatives said it uses AMD chips for its Copilot AI feature.

Price competition

AMD declined to disclose the cost of its chips -- it doesn't sell them individually, and end users typically buy them through hardware companies such as Dell or Super Micro Computer -- but the company plans to make the MI400 chip competitive on price.

The Santa Clara-based company pairs its GPUs with CPUs and networking chips it acquired through its 2022 acquisition of Pensando to build its Helios racks. That means wider adoption of its AI chips will benefit AMD's other businesses, too. It also uses an open-source networking technology called UALink to tightly integrate its rack systems, unlike Nvidia’s proprietary NVLink.

AMD claims its MI355X delivers 40% more tokens (a measure of AI output) per dollar than Nvidia’s chips because its chips use less power than rivals.

Data center GPUs can cost tens of thousands of dollars per chip, and cloud companies typically buy them in bulk.

AMD’s AI chip business is still much smaller than Nvidia’s. The company said it will have $5 billion in AI sales in fiscal 2024, but JPMorgan analysts expect the category to grow 60% this year.