AMD reveals next-generation AI chips with Openai CEO Sam Altman

Advanced Micro Devices On Thursday, we announced new details about the next generation AI chip, the Instinct MI400 series, which will be shipped next year.

The Mi400 chip can be assembled into a full server rack called Helios, AMD said.

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

Openai CEO Sam Altman appeared on stage with Su and said his company would use AMD chips.

“When you first started talking to me about specs, I didn't seem to have a way to sound crazy at all,” Altman said. “That would be amazing.”

AMD's rack scale setup makes the chip look like a single system user. This is important for most artificial intelligence customers, such as cloud providers and companies developing large language models. These customers want a “hyperscale” cluster of AI computers that can use huge amounts of power across data centers.

“Think of Helios as a rack that works like a single, huge calculation engine,” SU said. nvidia's Vera Rubin Racks is scheduled to be released next year.

AMD's rack-scale technology also allows the latest chips to compete with Nvidia's Blackwell chips. This is already composed of 72 graphics process units sewn together. Nvidia is AMD's primary for developing and deploying AI applications and is the sole rival of big data center GPUs.

Openai, a well-known Nvidia customer, provides AMD feedback on the MI400 roadmap, says Chip Company. With the MI400 chip and this year's MI355X chip, AMD plans to compete with rival Nvidia at price points, with company executives reporting on Wednesday that Chips will be expensive to operate thanks to lower power consumption, and AMD is lowering NVIDIA at a “active” price.

So far, Nvidia has dominated the market for data center GPUs. This is because they were the first company to develop the type of software needed for AI developers to utilize chips designed to display graphics in 3D games. For the past decade, AMD has focused on competition before the AI ​​boom Intel On the server CPU.

Su said AMD's Mi355X could surpass Nvidia's Blackwell chip, despite Nvidia's “owned” CUDA software.

“We say we have very strong hardware, which we've always known, but it also shows that open software frameworks have made great strides,” Su said.

AMD's stock has been flat so far in 2025, indicating that Wall Street still doesn't consider it a major threat to Nvidia's control.

Andrew Dieckmann, AMD's data center GPU total mangar, said on Wednesday that AMD's AI chips are few to work and few to get.

“All in the end, there's a meaningful cost of the acquisition delta, and then we've layered the competitive advantage of performance in addition to a very double-digit percentage saving,” Dickman said.

Over the next few years, large companies and countries are ready to spend hundreds of billions of dollars building new data center clusters around GPUs to accelerate the development of cutting-edge AI models. This includes $300 billion in planned capital expenditures from Megacap technology companies this year alone.

AMD expects the total market for AI chips to exceed $500 billion by 2028, but analyst estimates Nvidia currently has more than 90% of the market today, and more than 90% of the market today.

The companies are committed to releasing new AI chips annually on a yearly basis, highlighting how intense the competition is and how important bleeding edge edge chip technology is. Microsoft, Oracle and Amazon.

AMD has purchased or invested in 25 AI companies over the past year, including the server manufacturers that developed AMD earlier this year, which are needed to build rack-sized systems, Su said.

“These AI systems are becoming extremely complex and full stack solutions are really important,” Su said.

The cutting-edge AMD AI chip currently installed from a cloud provider is the instinct MI355X, which it says began shipping in production last month. AMD said it can be rented from cloud providers starting in the third quarter.

Companies building large data center clusters for AI can use far more processing power than traditional server applications to not only keep costs down, maintain flexibility and provide flexibility, but also to increase the need for “inference” or to meet the computing power needed to actually deploy chatbots or generated AI applications.

“What really changed is the huge increase in demand for reasoning,” Su said.

AMD officials said Thursday that they believe the new chip is suitable for Nvidia's reasoning. This is because AMD's chips are equipped with faster memory, allowing larger AI models to run on a single GPU.

The MI355X has seven times the computing capabilities of its predecessor, AMD said. These chips can compete with Nvidia's B100 and B200 chips that have been shipped since the end of last year.

AMD said its instinct chips have been adopted by seven of its 10 largest AI customers, including Openai. TeslaXai, and Cohere.

Oracle plans to provide its customers with clusters with over 131,000 MI355X chips, AMD said.

Staff Meta On Thursday they said they plan to use AMD's CPU and GPU clusters to perform Llama model inference and purchase AMD's next-generation servers.

A Microsoft representative said it will use AMD chips to provide Copilot AI capabilities.

AMD refused to say what the chips cost. It does not sell chips on its own. End users usually purchase through hardware companies. Dell or Super Microcomputer -However, the company plans to compete for the MI400 chips by price.

Santa Clara Company has been building a Helios rack by combining GPUs along with CPUs and networking chips since its acquisition of Pensando in 2022. In other words, facilitating adoption of AI chips should also benefit AMD's remaining businesses. It also uses open source networking technology to closely integrate Nvidia's own NVLink with rack systems called Ualink.

AMD claims that the Mi355X can use 40% more tokens (a measure of AI output) than Nvidia's chips, and less power than rival chips.

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

AMD's AI chip business is much smaller than Nvidia's. Although AI sales were $5 billion in 2024, JP Morgan analysts are hoping to grow 60% in the category this year.

clock: AMD CEO Lisa Su: Chip export control is headwind, but there are still opportunities for growth