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

Advanced Micro Devices connected Thursday unveiled caller details astir its next-generation AI chips, the Instinct MI400 series, that volition vessel adjacent year.

The MI400 chips volition beryllium capable to beryllium assembled into a afloat server rack called Helios, AMD said, which volition alteration thousands of the chips to beryllium tied unneurotic successful a mode that they tin beryllium utilized arsenic 1 "rack-scale" system.

"For the archetypal time, we architected each portion of the rack arsenic a unified system," AMD CEO Lisa Su said astatine a motorboat lawsuit successful San Jose, California, connected Thursday.

OpenAI CEO Sam Altman appeared connected signifier connected with Su and said his institution would usage the AMD chips.

"When you archetypal started telling maine astir the specs, I was like, there's nary way, that conscionable sounds wholly crazy," Altman said. "It's gonna beryllium an astonishing thing."

AMD's rack-scale setup volition marque the chips look to a idiosyncratic similar 1 system, which is important for astir artificial quality customers similar unreality providers and companies that make ample connection models. Those customers privation "hyperscale" clusters of AI computers that tin span full information centers and usage monolithic amounts of power.

"Think of Helios arsenic truly a rack that functions similar a single, monolithic compute engine," said Su, comparing it against Nvidia's Vera Rubin racks, which are expected to beryllium released adjacent year.

AMD's rack-scale exertion besides enables its latest chips to vie with Nvidia's Blackwell chips, which already travel successful configurations with 72 graphics-processing units stitched together. Nvidia is AMD's superior and lone rival successful large information halfway GPUs for processing and deploying AI applications.

OpenAI — a notable Nvidia lawsuit — has been giving AMD feedback connected its MI400 roadmap, the spot institution said. With the MI400 chips and this year's MI355X chips, AMD is readying to vie against rival Nvidia connected price, with a institution enforcement telling reporters connected Wednesday that the chips volition outgo little to run acknowledgment to little powerfulness consumption, and that AMD is undercutting Nvidia with "aggressive" prices.

So far, Nvidia has dominated the marketplace for information halfway GPUs, partially due to the fact that it was the archetypal institution to make the benignant of bundle needed for AI developers to instrumentality vantage of chips primitively designed to show graphics for 3D games. Over the past decade, earlier the AI boom, AMD focused connected competing against Intel successful server CPUs.

Su said that AMD's MI355X tin outperform Nvidia's Blackwell chips, contempt Nvidia utilizing its "proprietary" CUDA software.

"It says that we person truly beardown hardware, which we ever knew, but it besides shows that the unfastened bundle frameworks person made tremendous progress," Su said.

AMD shares are level truthful acold successful 2025, signaling that Wall Street doesn't yet spot it arsenic a large menace to Nvidia's dominance.

Andrew Dieckmann, AMD's wide manger for information halfway GPUs, said Wednesday that AMD's AI chips would outgo little to run and little to acquire.

"Across the board, determination is simply a meaningful outgo of acquisition delta that we past furniture connected our show competitory vantage connected apical of, truthful important double-digit percent savings," Dieckmann said.

Over the adjacent fewer years, large unreality companies and countries alike are poised to walk hundreds of billions of dollars to physique caller information halfway clusters astir GPUs successful bid to accelerate the improvement of cutting-edge AI models. That includes $300 billion this twelvemonth unsocial successful planned superior expenditures from megacap exertion companies.

AMD is expecting the full marketplace for AI chips to transcend $500 cardinal by 2028, though it hasn't said however overmuch of that marketplace it tin assertion — Nvidia has implicit 90% of the marketplace currently, according to analyst estimates.

Both companies person committed to releasing caller AI chips connected an yearly basis, arsenic opposed to a biannual basis, emphasizing however fierce contention has go and however important bleeding-edge AI spot exertion is for companies similar Microsoft, Oracle and Amazon.

AMD has bought oregon invested successful 25 AI companies successful the past year, Su said, including the purchase of ZT Systems earlier this year, a server shaper that developed the exertion AMD needed to physique its rack-sized systems.

"These AI systems are getting ace complicated, and full-stack solutions are truly critical," Su said.

Currently, the astir precocious AMD AI spot being installed from unreality providers is its Instinct MI355X, which the institution said started shipping successful accumulation past month. AMD said that it would beryllium disposable for rent from unreality providers starting successful the 3rd quarter.

Companies gathering ample information halfway clusters for AI privation alternatives to Nvidia, not lone to support costs down and supply flexibility, but besides to capable a increasing request for "inference," oregon the computing powerfulness needed for really deploying a chatbot oregon generative AI application, which tin usage overmuch much processing powerfulness than accepted server applications.

"What has truly changed is the request for inference has grown significantly," Su said.

AMD officials said Thursday that they judge their caller chips are superior for inference to Nvidia's. That's due to the fact that AMD's chips are equipped with much high-speed memory, which allows bigger AI models to tally connected a azygous GPU.

The MI355X has 7 times the magnitude of computing powerfulness arsenic its predecessor, AMD said. Those chips volition beryllium capable to vie with Nvidia's B100 and B200 chips, which person been shipping since precocious past year.

AMD said that its Instinct chips person been adopted by 7 of the 10 largest AI customers, including OpenAI, Tesla, xAI, and Cohere.

Oracle plans to connection clusters with implicit 131,000 MI355X chips to its customers, AMD said.

Officials from Meta said Thursday that they were utilizing clusters of AMD's CPUs and GPUs to tally inference for its Llama model, and that it plans to bargain AMD's next-generation servers.

A Microsoft typical said that it uses AMD chips to service its Copilot AI features.

AMD declined to accidental however overmuch its chips outgo — it doesn't merchantability chips by themselves, and end-users usually bargain them done a hardware institution similar Dell oregon Super Micro Computer — but the institution is readying for the MI400 chips to vie connected price.

The Santa Clara institution is pairing its GPUs alongside its CPUs and networking chips from its 2022 acquisition of Pensando to physique its Helios racks. That means greater adoption of its AI chips should besides payment the remainder of AMD's business. It's besides utilizing an open-source networking exertion to intimately integrate its rack systems, called UALink, versus Nvidia's proprietary NVLink.

AMD claims its MI355X tin present 40% much tokens — a measurement of AI output — per dollar than Nvidia's chips due to the fact that its chips usage little powerfulness than its rival's.

Data halfway GPUs tin outgo tens of thousands of dollars per chip, and unreality companies usually bargain them successful ample quantities.

AMD's AI spot concern is inactive overmuch smaller than Nvidia's. It said it had $5 cardinal successful AI income successful its fiscal 2024, but JP Morgan analysts are expecting 60% maturation successful the class this year.

WATCH: AMD CEO Lisa Su: Chip export controls are a headwind but we inactive spot maturation opportunity