AMD's Advancing AI 2026 conference in San Francisco this week produced the clearest evidence yet that hyperscalers are building a real alternative to Nvidia: Meta and OpenAI together committed 6 gigawatts each of AMD accelerator capacity, while the company unveiled the industry's first x86 server CPU in volume production on TSMC's 2nm node.
Twelve gigawatts is a number that demands context. One gigawatt of AI compute translates to roughly 25,000 to 50,000 high-end GPUs, depending on the generation and cooling configuration. Meta and OpenAI are restructuring their infrastructure around AMD. Meta's 6GW commitment, confirmed alongside AMD CEO Lisa Su at the July 22 event, begins with a 1GW custom MI450 deployment in the second half of 2026. OpenAI's matching 6GW commitment follows a similar schedule. Both deals span multiple chip generations, which tells you something important: these aren't opportunistic hedge positions. They're long-term supply agreements, and they were negotiated before AMD had shipped a single Helios rack.
The hardware AMD put on stage justifies that confidence. The MI455X accelerator carries 432GB of HBM4 memory at 19.6 terabytes per second of bandwidth per chip. Pack 72 of them into a Helios rack alongside EPYC Venice CPUs and Pensando Vulcano 800Gbps networking cards and you get 31 terabytes of pooled HBM4, 1.4 petabytes per second of aggregate memory bandwidth, and 2.9 exaflops of FP4 inference throughput in a single chassis. AMD is pricing that at $5 million to $5.5 million per rack, a figure that, according to reporting from wccftech, sits roughly 40% above Nvidia's second-generation Rubin system. The fact that Microsoft Azure and Oracle signed on as early Helios customers anyway suggests the performance math works out.
EPYC Venice is the quieter story here, and the more significant one. AMD confirmed it is the first x86 server CPU in volume production on TSMC's 2nm process node, with future production ramp planned at TSMC's Arizona facility. Venice, built on the Zen 6 architecture, packs 256 cores: a 33% jump over the 192-core ceiling of its Turin predecessor. AMD claims a 70% performance improvement over the prior generation. Intel's comparable P-core Xeon response, Diamond Rapids, isn't confirmed before 2027. That's a 12-month window. Enterprise buyers choosing between Venice and a competitive Intel server platform have exactly one real option. One. That's the situation.
AMD's data center revenue hit $5.8 billion in Q1 2026, up 57% year over year. That growth rate has attracted attention, but the more interesting number is what's behind it: software. On July 23, AMD announced a $5 billion investment in Anthropic, which the company framed as a bid to close the gap with Nvidia's CUDA ecosystem. That ecosystem - the vast library of software optimised for Nvidia hardware - has kept AMD competitive on specs while often losing on practicality. A deep partnership with one of the leading AI labs addresses the problem directly. Anthropic declined to comment on specific deployment commitments at launch.
For anyone who has followed this space over the past three years, there's a temptation to read AMD's big-number announcements with some scepticism. The company has announced gigawatt-scale partnerships before that took longer to materialise than the press release suggested. What's different this time is the operational specificity. Meta's initial 1GW MI450 deployment is scheduled for H2 2026, not 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