AMD Unveils MI400 AI Chips and Helios Rack System to Challenge Nvidia’s Dominance
LATEST
A New Era for AMD’s AI Ambitions
San Jose, California — June 12, 2025 — In a major announcement on Thursday, Advanced Micro Devices (AMD) revealed new details about its next-generation AI chips, the Instinct MI400 series, set to ship in 2026. AMD also introduced a novel rack-scale system called Helios, designed to integrate thousands of MI400 chips into a unified computing unit that can function as a single massive AI engine.
“For the first time, we architected every part of the rack as a unified system,” said AMD CEO Lisa Su during a launch event in San Jose. The new Helios racks mark a shift from individual chip performance to system-level computing, targeting cloud providers and AI developers who demand hyperscale infrastructure to train and deploy large language models.
A Strategic Partnership With OpenAI
A surprise appearance by OpenAI CEO Sam Altman underscored the significance of the announcement. Altman confirmed OpenAI’s plans to use AMD’s MI400 chips, saying, “When you first started telling me about the specs, I was like, there’s no way, that just sounds totally crazy. It’s gonna be an amazing thing.”
OpenAI, historically a core Nvidia customer, has also been providing AMD with direct feedback on the MI400 roadmap, according to the chipmaker.
Competing Against Nvidia’s GPU Empire
The Helios system is AMD’s most ambitious attempt yet to rival Nvidia’s dominance in the data center GPU market, which analysts estimate Nvidia currently controls with over 90% share. AMD’s rack-scale system aims to directly compete with Nvidia’s Vera Rubin racks and Blackwell chips, which are expected to ship in 2025.
Lisa Su compared Helios to Nvidia’s offerings, emphasizing, “Think of Helios as really a rack that functions like a single, massive compute engine.” AMD believes its architecture—pairing CPUs, GPUs, and networking hardware acquired from Pensando in 2022—offers a competitive edge in power efficiency and cost-effectiveness.
A Focus on Inference and Efficiency
AMD is targeting a critical growth area in AI: inference. Su noted that demand for inference — the computational process of running AI models — has surged. AMD’s MI355X chips, launched earlier this year and currently being deployed by cloud providers, are designed with more high-speed memory to support larger AI models on a single GPU.
The MI355X delivers seven times the computing power of its predecessor and, according to AMD, provides up to 40% more AI tokens per dollar than Nvidia’s chips, largely due to lower power consumption. AMD argues that its chips are particularly well-suited for inference workloads that require energy-efficient, high-throughput systems.
Adoption by Major AI Players
AMD reported that its Instinct chips have already been adopted by 7 of the 10 largest AI customers, including OpenAI, Tesla, xAI, Cohere, Meta, Oracle, and Microsoft. Oracle, for instance, plans to deploy clusters with over 131,000 MI355X chips. Microsoft said it currently uses AMD hardware to support its Copilot AI features.
Meta confirmed it is leveraging AMD CPUs and GPUs for running inference with its Llama model and plans to purchase AMD’s next-gen servers as well.
Aggressive Pricing Strategy
Though AMD declined to disclose chip prices, executives emphasized that the company is aggressively undercutting Nvidia in terms of total cost of ownership. Andrew Dieckmann, general manager of AMD’s data center GPU business, stated that AMD offers “significant double-digit percentage savings” both in acquisition and operation costs.
“Across the board, there is a meaningful cost of acquisition delta that we then layer on our performance competitive advantage,” Dieckmann said.
Given that high-end GPUs for data centers can cost tens of thousands of dollars apiece, AMD’s pricing approach could be critical in gaining market share.
Open Software Versus Proprietary Ecosystems
A long-standing challenge for AMD has been software compatibility. Nvidia’s CUDA ecosystem is deeply entrenched among AI developers, offering a proprietary but mature framework for deploying machine learning models. In contrast, AMD has backed open-source alternatives and is working to prove their viability at scale.
Su insisted that the MI355X outperforms Nvidia’s Blackwell chips, even without CUDA. “It says that we have really strong hardware, which we always knew, but it also shows that the open software frameworks have made tremendous progress,” she said.
Massive Market Opportunity Ahead
The stakes are high. AMD estimates the total addressable market for AI chips will exceed $500 billion by 2028. Tech giants and governments alike are investing heavily, with projected $300 billion in AI infrastructure spending in 2025 alone from the world’s largest technology companies.
To strengthen its AI portfolio, AMD has made or invested in 25 AI-related companies over the past year, including its acquisition of ZT Systems, a server manufacturer that helped enable Helios’ development.
“These AI systems are getting super complicated, and full-stack solutions are really critical,” Su said.
Aiming for Annual Chip Releases
Both AMD and Nvidia are now committed to releasing new AI chips on an annual cycle, signaling a rapid acceleration in innovation and intense market competition. This pivot from a biannual schedule underscores how critical AI has become for companies like Microsoft, Amazon, Oracle, and others building out AI-first strategies.
Though AMD’s AI business remains significantly smaller than Nvidia’s, it is growing quickly. The company reported $5 billion in AI-related sales for fiscal 2024. JP Morgan analysts expect that figure to rise by 60% in 2025.
Wall Street Remains Cautious
Despite the ambitious plans and technological advancements, AMD’s stock has remained flat in 2025. Analysts suggest that while AMD is narrowing the performance and efficiency gap, Wall Street is waiting to see whether it can chip away at Nvidia’s overwhelming lead.