Dr Lisa Su’s CES 2026 roadmap rethinks AI infrastructure at planetary scale, a bold challenge to Nvidia’s dominance
At CES 2026 in Las Vegas, AMD Chair and CEO Dr. Lisa Su unveiled a breathtaking vision for the future of computing, one that moves beyond zettaflops and into the era of yotta-scale AI infrastructure. This isn’t incremental progress, but a bold architectural gamble: energizing data centers with modular rack-scale systems capable of trillions of parameters, exaflops of performance, and the agility to train the next generation of massive AI models. With its Helios platform, new Instinct MI400-series GPUs, and a roadmap stretching toward the MI500 family, AMD is positioning itself as a counterweight to Nvidia’s dominance, and redefining what it means to build infrastructure for artificial intelligence at planetary scale.
2026 Matters: The Yotta-Scale Imperative
In 2025, industry watchers began using a new word in AI compute forecasting: yottaflops. A yottaflop equals 10²⁴ floating-point operations per second, a million times larger than a zettaflop and 10,000× more than all global AI compute in 2022. AMD’s leadership says that within five years, demand for AI performance could exceed 10 yottaflops, driven by trillion-parameter models and increasingly ambitious workloads.
Yet raw power is only part of the story. Today’s compute landscapes, dominated by GPU clusters and emerging CPUs, struggle to keep pace with training requirements for next-gen generative models, multi-modal AI, digital agents, and real-time robotics intelligence. AMD’s response is an architectural one: Helios, a rack-scale blueprint that unifies compute, memory, networking, and software into a modular, scalable system capable of delivering up to 3 AI exaflops per rack, enough to approach the yottaflops era when clusters are interconnected across hundreds of racks.
CES 2026 Unveiling: Helios, MI455X, and the Blueprint for Rack-Scale AI
At the heart of AMD’s roadmap is Helios, the company’s rack-scale AI platform. Unlike discrete server nodes, rack-scale systems integrate CPUs, GPUs, high-bandwidth memory (HBM4), and network fabrics into a cohesive compute continuum. Helios racks combine 72 Instinct MI455X accelerators with AMD EPYC “Venice” CPUs and AMD Pensando™ Vulcano NICs to produce exceptional bandwidth and AI throughput.
Helios isn’t just brute force, it’s connected force. Through open, modular design and high-speed networking, Helios can link multiple racks into unified fabrics capable of supporting trillion-parameter model training and inference workloads. The significance is clear: as AI models balloon in size and complexity, infrastructures that treat compute as isolated boxes will fall behind those that see whole racks as unified supercomputers.
Instinct MI400-Series: A New GPU Foundation
Supporting Helios and AMD’s broader vision are the Instinct MI400-series accelerators, including the MI430X, MI440X, and MI455X GPUs. These next-generation chips are built on advanced process nodes and optimized for diverse AI workloads: from high-precision scientific computing to low-precision inference and large-scale training.
- MI430X targets scientific and high-precision HPC.
- MI440X is designed for enterprise AI deployments on standard server infrastructure.
- MI455X fuels Helios, combining massive memory and compute density for data center AI performance.
By segmenting the MI400 portfolio, AMD addresses both hyperscale training clusters and more modest enterprise needs. Equally important is AMD’s continued investment in the ROCm software ecosystem, ensuring that developers can tap into GPU acceleration across heterogeneous workloads, a critical factor for adoption in real-world systems.
4. From Edge to Exaflops: Full-Stack AI Strategy
Unlike some competitors focused narrowly on GPUs, AMD’s strategy spans end-to-end AI infrastructure:
- Edge and client AI with ~Ryzen AI~ processors delivering robust on-device inference for PCs and embedded systems.
- Rack-scale AI with Helios and MI400 accelerators powering massive data center workloads.
- Next-gen MI500 series previewed for 2027, aiming for orders-of-magnitude performance leaps over current generations.
This breadth reflects a recognition that AI isn’t just about training giant models, it’s about making intelligent systems ubiquitous. From autonomous robots and intelligent healthcare diagnostics to immersive real-time applications and customizable personal agents, AMD’s roadmap touches every layer of the computing stack.
Competitor Dynamics: AMD vs Nvidia and the AI Hardware Landscape
AMD’s yotta-scale announcement arrives amid a broader competitive battle with Nvidia, which at CES also unveiled its Vera Rubin platform, a unified AI supercomputer leveraging custom CPUs, GPUs, and networking designed to vastly accelerate training workflows.
Traditionally, Nvidia has dominated the AI GPU market, with its Blackwell architecture driving much of the recent data center growth. But AMD’s architecture, coupled with open standards and diverse CPU/GPU integration, signals a shift toward more choice and modularity in AI infrastructure. Moreover, partnerships with OpenAI and other ecosystem players reinforce AMD’s relevance in hyperscale AI deployments.
The hardware competition is not just about speed; it’s about ecosystem and openness. AMD’s emphasis on ROCm support, open rack designs, and broad hardware segmentation challenges a historically more proprietary landscape, encouraging broader innovation and flexibility for cloud providers and enterprises alike.
Market Implications: Enterprise Adoption and Sovereign AI
The implications of AMD’s roadmap extend far beyond CES buzz. As AI becomes mission-critical in industries like healthcare, drug discovery, finance, autonomous vehicles, and scientific research, the underlying compute infrastructure determines who gets to innovate and how fast.
AMD’s Helios platform is already being adopted by partners such as OpenAI and Blue Origin, and collaborations with firms like HPE indicate that open rack-scale systems will soon power research, cloud services, and sovereign AI projects across global markets.
In Europe, AMD’s rack-scale approach, integrated into systems like HPE’s supercomputers, ighlights a trend toward sovereign computing, where countries seek more control over their AI infrastructure rather than relying on multinational cloud providers alone.
Energy and Sustainability: The Unseen Constraint on AI Growth
As AMD’s leadership points out, global infrastructure is hurtling toward yottaflops, but the path isn’t just a matter of chip count. Energy and thermal management are critical bottlenecks. Today’s high-performance AI clusters already create intense power demands that strain existing grid infrastructure. Scaling to yotta-level compute will demand new approaches to energy efficiency, cooling, and data center design.
AMD’s Helios platform addresses this in part through efficient GPU memory designs, optimized CPU acceleration, and modular, scalable designs that reuse and redistribute power more effectively than traditional server islands. Yet broader energy innovation, from liquid cooling to renewable sourcing, will need to follow if the industry hopes to sustain unfettered growth.
Software Imperative: ROCm and Developer Engagement
Hardware alone cannot unlock AI’s potential; software ecosystems matter just as much. AMD’s open ROCm software stack, now supported across servers, PCs, and embedded systems, enables developers to harness GPU acceleration with familiar tools like PyTorch and TensorFlow.
This open roadmap lowers barriers to entry for enterprises, researchers, and startups looking to build AI solutions without being locked into proprietary hardware stacks. In a world where AI models and data pipelines evolve rapidly, software portability and ecosystem support often dictate who wins and who falls behind.
From Vision to Reality: AI Everywhere, For Everyone?
AMD’s mantra at CES 2026, “AI Everywhere, for Everyone”, reflects a holistic vision of a future where artificial intelligence is embedded across devices, cloud systems, and industrial applications. It’s not just about training a bigger model, it’s about enabling a new class of AI that is accessible, scalable, and sustainable across use cases.
Whether these promises materialize will depend on the broader industry’s ability to move beyond prototype demos into production deployments that deliver real economic value. But one thing is clear: AMD’s roadmap shifts the narrative in AI hardware, from single-chip performance to rack-scale ecosystems capable of yotta-level compute.
