AMD launches Helios rackscale AI system with Anthropic and OpenAI deals
AMD has used its Advancing AI 2026 conference in San Francisco to launch a sweeping update to its AI infrastructure portfolio, anchored by the AMD Helios rackscale solution now entering production deployment. The announcement signals AMD's most direct challenge yet to Nvidia's dominance in the AI accelerator market, backed by named commitments from some of the largest buyers in the industry.
The Helios system integrates 72 Instinct MI455X GPUs and 18 sixth-generation EPYC "Venice" CPUs, connected through AMD's Pensando networking fabric and driven by the ROCm open software stack. AMD says the platform delivers up to 30% more inference tokens per dollar than the leading competitive solution, a figure based on internal benchmarks run against Nvidia's Vera Rubin NVL72 rack using the Kimi K2 Thinking workload. Buyers will want to scrutinise the full methodology, as the comparison relies on AMD's own performance lab estimates and projected hourly pricing rather than independently audited results.
Landmark partnerships and deployment timelines
The scale of the ecosystem commitments announced alongside Helios is notable. Anthropic has agreed to deploy up to two gigawatts of MI455X GPUs in Helios racks, with a multiyear engineering collaboration to use Claude to optimise AMD's ROCm software development. OpenAI is optimising GPT-class workloads on Helios using its Triton framework with ROCm, and expects to bring Helios online in the fourth quarter of 2026, with deployments accelerating through 2027. Meta is validating sixth-generation EPYC platforms in its labs and has begun testing Helios racks ahead of gigawatt-scale deployment. Cerebras is combining its low-latency inference compute with Helios high-throughput racks for inference serving use cases.
System availability will come through OEM partners including HPE, Lenovo and Supermicro, as well as infrastructure partners Sanmina and Wiwynn. AT&T and Cisco are collaborating with AMD on enterprise and edge AI deployments, with AT&T using Instinct GPUs and ROCm to power its open-source OTel 2.0 telecom AI model.
Beyond Helios, AMD launched the Instinct MI430X accelerator for high-precision scientific computing, claiming up to 288 TFLOPS of FP64 performance, and the Instinct MI350P GPU, which AMD says delivers up to 4.2 times more tokens per second per dollar than Nvidia's H200 NVL, again based on internal testing. AMD also introduced ROCm.ai, a developer platform integrating coding agents including Claude, Codex and Cursor to simplify GPU software development on AMD hardware.
Market context and competitive read
AMD's launch comes at a critical inflection point in the AI accelerator market. Nvidia has built a commanding position through its CUDA ecosystem, supply chain scale and its own NVL rack architecture. However, hyperscalers and frontier AI labs have publicly expressed interest in multi-vendor strategies to manage supply concentration risk and negotiate on price, creating structural demand for credible alternatives.
AMD's ROCm software stack has historically been cited as a gap relative to CUDA's maturity, but the Anthropic and OpenAI engineering collaborations, if they deliver on their stated goals, would represent meaningful progress on closing that gap. The addition of ROCm.ai and native support across PyTorch, Hugging Face and vLLM frameworks reinforces that trajectory.
AMD also shared a multi-year product roadmap through 2030, covering Instinct MI500 Series GPUs in 2027, MI600 Series in 2028, and future EPYC "Zen 7" and "Zen 8" CPU architectures. The company projects its total addressable market reaching approximately two trillion dollars by 2030, driven by AI demand across data centre, edge and embedded markets. That figure is a forward-looking estimate and should be treated as illustrative rather than a committed forecast.
From a regulatory standpoint, AMD's growing AI infrastructure business is exposed to US export controls administered by the Bureau of Industry and Security. The company acknowledged this risk in its cautionary disclosures; any tightening of restrictions on GPU exports to certain markets could affect the scale of international Helios deployments.