AMD Fires Back at Nvidia with Helios AI System and Next-Gen Epyc CPUs

Four months after Nvidia unveiled its Vera Rubin AI infrastructure, AMD has launched a direct counteroffensive with its sixth-generation Epyc CPUs and the Helios rack-scale AI solution, powered by the new Instinct MI455X GPUs. The announcements, made at AMD’s Advancing AI conference in San Francisco, mark a pivotal moment in the competitive AI infrastructure landscape as the company challenges Nvidia’s market dominance with an integrated strategy encompassing both GPUs and CPUs.

AMD CEO Lisa Su announced that Helios is now in full production, with shipments from hardware partners scheduled to begin by the end of the third quarter of 2026. Helios is a fully integrated rack-scale AI system combining Epyc processors, Instinct MI455X GPUs, Pensando networking, and ROCm software. When compared to Nvidia’s Vera Rubin NVL72 system, Su said Helios delivers 15% more AI compute, 50% more memory, and 50% more scale-out bandwidth, while offering up to 30% more tokens per dollar. “Every Helios can deliver more performance for the largest models, more capacity for longer context, and the bandwidth to scale across thousands of racks,” Su said during her keynote address.

The launch comes as the shift to agentic AI is fueling significant growth in the AI infrastructure market. Su revised her previous market forecast upward, projecting the AI infrastructure market will reach $1.4 trillion by 2030, up from an earlier estimate of $500 billion by 2028. She noted that agentic AI is also creating a new growth vector for server CPUs, with the market expected to grow from $25 billion today to $200 billion by 2030. “What we’re seeing is that the rate and pace of agentic AI adoption is much faster than any of us thought,” Su said. “Agents are scaling from millions to billions, driving unprecedented infrastructure needs.”

AMD also launched its Epyc 9006 Series server CPUs, code-named Venice, a family of four purpose-built processors for enterprise, cloud, high-performance computing, and AI workloads. The lineup includes the Epyc 9006 SP7 with up to 256 cores, designed for high-density agent sandbox execution and shipping in the fourth quarter of 2026; the Epyc 9006 SP8, available in 8- to 128-core configurations with low-power options for general-purpose enterprise workloads, shipping in the first half of 2027; the Epyc 9006X SP7 supporting up to 256 cores; and the Epyc 9006 LP, formerly code-named Verano, built as an AI host node CPU with up to 72 cores, shipping in the second half of 2027. AMD claimed that a 256-core Venice CPU offers 2.2 times the throughput of Nvidia’s 88-core Vera CPU, while a 96-core Venice chip delivers 20% better per-core performance.

Industry analysts described AMD’s announcements as highly significant. IDC analyst Ashish Nadkarni said AMD is the only company that can match what Nvidia offers in data center and AI infrastructure. “They are making steady progress to be the strong No. 2 in the market,” he said. “There is nobody else in the semiconductor industry today that can offer what AMD is offering other than Nvidia.” Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, said Helios is a very strong response to Nvidia’s NVL72 platforms. “It doesn’t matter how you measure Epyc versus Vera,” Kimball said. “AMD is saying across any of those measurement vectors, not only do we win, but we win decisively. That’s a strong story.”

Helios is already gaining traction with major customers. Microsoft announced it will deploy Helios to power AI workloads, including frontier model AI inference and Azure AI services. Anthropic also said it will deploy up to 2 GW of the GPUs in Helios solutions. AMD stated that its MI455X GPU delivers 34 times higher token throughput and up to 18 times lower token cost than its previous MI355X GPU. During the keynote, Su revealed that AMD’s Epyc CPUs have reached 46% revenue share of the server market, with more than 60% of the Fortune 100 running on Epyc. Looking ahead, Su previewed seventh-generation CPUs codenamed Florence for 2028 and eighth-generation CPUs for 2030, with plans for a new GPU and Helios system every year, including the Instinct MI500 in 2027 and MI600 in 2028.