AMD now has Meta, OpenAI, Microsoft and Anthropic lining up behind its AI racks. Nvidia still owns the market, but buyers with billion-dollar power plans clearly don't want one supplier deciding the whole compute future.
Twelve gigawatts gets your attention. Meta and OpenAI have each committed to deploy up to 6 gigawatts of AMD Instinct GPUs across multiple chip generations, according to AMD's own announcements. Meta's agreement was announced on February 24, 2026, with shipments for the first 1 gigawatt expected in the second half of 2026. OpenAI's deal was announced on October 6, 2025, with its first 1 gigawatt deployment of MI450 Series GPUs also set to begin in the second half of 2026.
That isn't this week's news by itself. The current development is bigger than a re-run of two old press releases. AMD used the week around its Advancing AI 2026 event in San Francisco to show that those commitments now sit inside a wider customer list: Microsoft is bringing Helios to Azure, and Anthropic has agreed to deploy up to 2 gigawatts of AMD Instinct MI450 GPUs while AMD invests up to $5 billion in the Claude maker. That's the part you should focus on.
Nvidia hasn't been beaten. Not close. But AMD has moved from trying to win benchmark arguments to assembling a customer map that includes Meta, OpenAI, Microsoft, Oracle and Anthropic. That's a different kind of threat. Specs are one argument. Purchase commitments are another.
The Rack Is The Product Now
The Helios pitch is not subtle. AMD's rack-scale reference design puts 72 Instinct MI455X GPUs together with EPYC Venice CPUs and Pensando networking. AMD lists the rack at 31TB of HBM4 memory, 19.6TB per second of memory bandwidth per GPU, 1.4 exaflops of FP8 compute and 2.9 exaflops of FP4 compute. Those are peak figures, not proof that every customer workload will run better. Still, you can see why memory-hungry AI labs are paying attention.
Wccftech reported, citing Futurum estimates, that Helios racks could be priced at $5 million to $5.5 million each, around 40% above expected second-generation Nvidia Rubin rack pricing. Treat that number carefully, because AI rack pricing depends on volume, service agreements, networking, cooling and the customer's own negotiations. Even so, it punctures the old idea that AMD's only role is to be the cheaper alternative.
Microsoft's July 20 blog post gives the cleanest customer signal. Azure is adding AMD's Helios platform for large-scale inference through ND MI455X v7 virtual machines, while also adding HDv2 and HXv2 VM families powered by sixth-generation EPYC Venice CPUs. Microsoft didn't disclose capacity or financial terms. The missing number is frustrating, but the deployment target is concrete: inference, AI data systems and chip design workloads inside Azure.
Oracle was already on the Helios customer list. Anthropic now adds a different kind of pressure. The Wall Street Journal reported that Anthropic will acquire up to 2 gigawatts of AMD's latest Instinct chips beginning in early 2027, with AMD's investment tied to milestones. That's not a casual supplier announcement. It's hardware and financing and engineering all moving together.
Venice Gives AMD Another Opening
EPYC Venice is the quieter story, and it matters. AMD announced in May that Venice had begun production ramp on TSMC's 2nm process technology in Taiwan, with future production planned at TSMC's Arizona facility. The company described it as the first high-performance computing product to enter production on TSMC's advanced 2nm node. Worth noting. The chip is built on Zen 6 and supports up to 256 cores, compared with the 192-core ceiling on Turin. AMD has also claimed up to 70% more compute performance than the current EPYC Turin generation. You don't have to accept every vendor benchmark as destiny. You do have to take the timing seriously. If cloud buyers are planning AI capacity for late 2026 and 2027, Venice gives AMD a CPU story to sit beside the accelerator story.
Here's the thing: Nvidia's moat has never been only silicon. CUDA, developer familiarity and the habit of buying Nvidia first are still powerful. AMD's ROCm stack has improved, but it still has to prove itself across the messy workloads customers actually run, not only the clean cases shown on stage. A rack spec sheet won't fix that on its own.
That is why the Anthropic deal is more than another customer logo. The Verge reported that AMD will incorporate Anthropic's Claude models into its own software development and product workflows as part of the partnership. If AMD wants customers to believe ROCm can support serious frontier AI work, it needs labs like Anthropic pushing on the software early and in production - not just the polished demos.
For buyers, the useful lesson is plain. Nvidia remains the default, but default no longer means uncontested. Meta has a 6GW AMD agreement. OpenAI has one too. Microsoft is bringing Helios into Azure. Anthropic is taking up to 2GW and AMD money comes with it. These are not small tests hiding in a lab.
Power is the real scoreboard now. If AMD can turn gigawatt commitments into delivered racks and working software that customers actually reorder, Nvidia's grip on AI data centers loosens in a practical way. If it can't, this week's excitement becomes another reminder that the AI hardware market is full of promises scheduled for next year.
Also read: Microsoft is replacing OpenAI's image models in PowerPoint and Bing with its own MAI-Image-2 • Nvidia's Jetson chips are becoming the default AI layer for Moon missions • Black Forest Labs collapses image, video, and robot control into a single model with FLUX 3
