AMD is betting big on inference, the phase of AI where models answer real-world questions after training ends, and its new Helios rack-scale systems are designed to make that workload dramatically cheaper to run at scale. The company unveiled its next-generation AI infrastructure portfolio at Advancing AI 2026, including the Helios rackscale solution, which combines 72 high-performance AMD Instinct MI455X GPUs (graphics processing units, the specialized chips that accelerate AI workloads) and 18 sixth-generation AMD EPYC CPUs (central processing units) in a single integrated rack.

The timing reflects a fundamental shift in the AI industry. While companies like OpenAI and Meta spent billions on training frontier models, the real economic pressure now comes from serving those models to millions of users. That's where inference costs dominate. AMD's Helios delivers up to 30% more inference tokens per dollar than leading competitors, meaning cloud providers can answer more user queries without proportionally increasing their hardware spending.

What's Driving the Shift From Training to Inference?

Training a large language model (LLM), the type of AI system that powers ChatGPT or Claude, is expensive but happens once. Inference, by contrast, happens continuously. Every time a user types a prompt and waits for an answer, that's an inference operation. As AI adoption accelerates, inference workloads are growing exponentially, and the economics of running them efficiently have become critical to profitability.

AMD's announcement reflects this reality. The company noted that leading AI labs and cloud providers are already choosing Helios for its performance, including OpenAI, Anthropic, Meta, Microsoft, Oracle, and several infrastructure specialists like Tensorwave, Vultr, and Cirrascale. OpenAI expects to bring Helios systems online beginning in the fourth quarter of 2026, with deployments accelerating throughout 2027.

How Are Major AI Companies Planning to Deploy These Systems?

• Anthropic Partnership: Anthropic and AMD announced a strategic partnership to deploy up to 2 gigawatts of AMD Instinct MI455X GPUs in Helios rackscale solutions, with a multiyear engineering collaboration to optimize workloads for AMD hardware and accelerate ROCm software development using Anthropic's Claude AI model.
• OpenAI Optimization: OpenAI and AMD are partnering to optimize the full AI stack from silicon to software, leveraging OpenAI's Triton framework with AMD ROCm software to optimize GPT-class workloads on AMD Instinct MI455X GPUs.
• Meta Co-Design: Meta and AMD are co-designing for gigawatt-scale deployments, with Meta now validating sixth-generation EPYC CPU platforms in its labs and testing workloads on AMD Helios racks as it prepares for large-scale deployment.
• Cerebras Integration: Cerebras and AMD are collaborating to combine Cerebras ultra-low-latency AI compute with AMD Helios high-throughput infrastructure to improve inference efficiency and economics for ultra-low-latency inference serving.

These partnerships signal confidence in AMD's approach. Rather than building proprietary, closed ecosystems, AMD is positioning itself as an open platform that works with the industry's leading AI companies. This contrasts with Nvidia's dominant position in AI chips, where the company controls both the hardware and much of the software stack.

"The next phase of AI will span frontier models, agents and physical AI, creating new opportunities to bring intelligence everywhere. Realizing that potential will take the entire industry working together. AMD is partnering across the ecosystem to deliver leadership compute and open platforms that give customers the performance, flexibility and choice to scale AI from the data center to the edge," said Dr. Lisa Su, chair and CEO of AMD.

Dr. Lisa Su, Chair and CEO, AMD

What Hardware Improvements Does AMD's New Lineup Offer?

AMD's new hardware stack addresses multiple layers of the inference problem. The sixth-generation EPYC processors deliver the broadest server CPU portfolio for agentic AI (systems that can take actions autonomously), spanning cloud, enterprise, and high-performance computing workloads. These CPUs are designed to keep GPU accelerators fully fed with data, preventing bottlenecks that slow inference.

On the GPU side, the AMD Instinct MI455X delivers 34 times higher token throughput compared to the previous-generation MI355X, a dramatic performance jump. For high-precision scientific computing, the AMD Instinct MI430X accelerator offers up to 288 TFLOPS (trillion floating-point operations per second) of hardware-based FP64 performance, making it the most advanced option for sovereign AI and HPC (high-performance computing) applications.

AMD also introduced the Instinct MI350P GPU, which brings AI acceleration to existing infrastructure with leadership token economics. The MI350P delivers up to 4.2 times more tokens per second per dollar than competing solutions, making it attractive for organizations that want to upgrade without replacing entire data center infrastructure.

How Is AMD Building Out Its Software Ecosystem?

Hardware alone doesn't win in AI infrastructure. Software matters equally. AMD is advancing its ROCm (Radeon Open Compute) open software platform, which provides developers with the tools to build and deploy AI on AMD hardware. The company introduced ROCm.ai, an AI-driven development platform that helps developers build, optimize, and deploy GPU software faster across AMD platforms.

ROCm.ai enables popular coding agents such as Claude, Codex, and Cursor to understand AMD platforms and ROCm natively, essentially allowing AI to help programmers write better code for AMD chips. Leading open-source frameworks including PyTorch, Hugging Face, vLLM, and SGLang are already enabled on the MI455X and showing strong results.

This software-first approach matters because it lowers the barrier to adoption. Developers don't need to learn proprietary languages or tools; they can use familiar open-source frameworks and get AI assistance to optimize for AMD hardware.

What's AMD's Long-Term Roadmap for AI Compute?

AMD is committing to an annual cadence of CPU, GPU, networking, and rack-scale innovation through 2030. The company shared details on upcoming generations that signal sustained investment in the inference market.

Next-generation EPYC server CPUs based on the "Zen 7" architecture are coming in 2028, with processors codenamed "Florence," "Ferrara," and "Fidenza" expected to extend AMD's leadership in density, performance, and performance-per-watt. The "Ravenna" CPU based on the "Zen 8" architecture will follow in 2030. On the GPU side, the next-generation AMD Instinct MI500 Series GPUs are coming in 2027, followed by the MI600 Series in 2028. The next-generation AMD Helios 500 rackscale solution will be powered by MI500 Series GPUs and EPYC "Verano" CPUs, with the Helios 600 following later.

This roadmap extends AMD's competitive window. By committing to annual updates, AMD is signaling that it won't cede the inference market to Nvidia, which has dominated AI chip sales for years. The company is also betting that the shift from training to inference will create a different competitive dynamic, one where cost-per-token matters more than raw peak performance.

Why Does This Matter for Enterprise AI Adoption?

For enterprises considering AI deployment, AMD's strategy has practical implications. The company's emphasis on open platforms and ecosystem partnerships means customers have more choice in how they build and deploy AI systems. Rather than being locked into a single vendor's stack, enterprises can mix and match components from different manufacturers while maintaining compatibility through open standards like ROCm.

The 30% efficiency gain in inference tokens per dollar also translates directly to lower operational costs. For a large cloud provider running millions of inference queries daily, that efficiency difference compounds into significant savings. It also means smaller companies and startups can afford to run larger models or serve more users with the same hardware budget.

AMD's partnerships with Anthropic, OpenAI, and Meta also signal that these companies see real value in the platform. These aren't small players testing new hardware; they're the companies defining the frontier of AI development. Their willingness to invest engineering resources into optimizing for AMD hardware suggests confidence that AMD's approach will remain competitive for years to come.