Advanced Micro Devices (NASDAQ:AMD) used its CES 2026 keynote to outline an aggressive vision for scaling AI compute across cloud infrastructure, PCs, healthcare, robotics, and scientific supercomputing, with Chair and CEO Dr. Lisa Su repeatedly emphasizing that “you ain’t seen nothing yet” as AI adoption accelerates.
AI adoption, compute demand, and AMD’s “Yotta scale” message
Su framed AI as “the most important technology of the last 50 years” and AMD’s “number one priority,” arguing that AI is moving from early excitement to broad utility across industries. She said AI usage has grown from “a million people using AI to now more than a billion active users,” and that AMD projects adoption rising to “over five billion active users” as AI becomes “indispensable.”
AMD, Su said, is positioned to address AI across “cloud,” “PCs,” and “the edge,” highlighting the company’s mix of CPUs, GPUs, NPUs, and custom accelerators.
Cloud infrastructure: Helios rack and Instinct MI455
In the cloud segment, Su said “every major cloud provider runs on AMD EPYC CPUs,” and that “eight of the top 10 AI companies use Instinct accelerators.” She described training demand rising sharply and inference surging, with token volumes increasing “100 times” over the last two years.
To address what she called “AI infrastructure at Yotta scale,” Su introduced Helios, AMD’s next-generation rack-scale platform. She described Helios as a double-wide rack built on the OCP Open Rack Wide standard developed “in collaboration with Meta,” weighing “nearly 7,000 pounds.” Helios uses an open modular rack design, high-speed networking, and turnkey deployment to scale clusters.
At the core of Helios, AMD unveiled its next-generation Instinct MI455 accelerators and detailed the compute tray configuration:
- Each tray includes four MI455 GPUs paired with the next-generation EPYC CPU code-named “Venice” and Pensando networking chips, with liquid cooling.
- MI455 is built with “2-nanometer and 3-nanometer process technologies,” advanced “3D chiplet packaging,” and “HBM4 memory.”
- Su said MI455 includes “320 billion transistors,” “432 gigabytes” of HBM4, and “12” compute and I/O chiplets.
- Venice is built on “2-nanometer process technology” and features “up to 256” Zen 6 cores, with doubled memory and GPU bandwidth versus the prior generation.
- Helios uses Ultra Accelerator Link tunneled over Ethernet for rack-scale GPU cohesion, alongside “800-gig Ethernet” Pensando chips and “Ultra Ethernet NICs.”
Su also provided a set of rack-scale specifications, including “up to 2.9 exaflops of performance” per rack, “31 terabytes of HBM4 memory,” “260 terabytes per second” of scale-up bandwidth, and “43 terabytes per second” of aggregate scale-out bandwidth. She said Helios is “on track to launch later this year.”
Performance claims were a centerpiece: Su said MI355 (launched roughly six months earlier) delivered “up to 3x” more inference throughput versus the prior generation, while MI455 is expected to deliver “up to 10 times more performance across a wide range of models and workloads.”
OpenAI and AMD: compute constraints and “agentic workflows”
OpenAI President and co-founder Greg Brockman joined Su on stage and described AI shifting from simple prompts to “agentic workflows,” where models may run “for minutes or hours or soon even days,” and users might operate “a fleet of agents.” Brockman said OpenAI has been “tripling our compute every single year for the past couple of years,” and also “tripled our revenue,” while remaining “compute constrained.”
Brockman argued that increasingly capable agents will require dramatically more hardware, saying he would “love to have a GPU running in the background for every single person in the world,” but noted “no one has a plan to build that kind of scale.” He also discussed AI use in healthcare through anecdotal examples where ChatGPT helped identify potential medical risks and prompted follow-up with clinicians.
Looking ahead, Brockman said he expects compute availability to increasingly influence economic outcomes, stating he believes “GDP growth will itself be driven by the amount of compute that is available in a particular country.” He also described an OpenAI experiment connecting “GPT-5” to a wet lab setup, claiming a “79x” improvement in efficiency for a protocol based on iterative model suggestions and human validation.
Software, partners, and the AI PC push
Su said AMD’s AI platform strategy depends on an “open ecosystem,” centered on ROCm. She called ROCm “the industry’s highest performance open software stack for AI,” and said it includes day-zero support for major frameworks and is supported by projects such as PyTorch, vLLM, SGLang, and Hugging Face.
Several partners highlighted AMD-based deployments. Luma AI CEO and co-founder Amit Jain said 60% of Luma’s “rapidly growing inference workloads” run on AMD GPUs, and that Luma plans to expand its partnership “about 10 times” in 2026. Jain introduced Ray 3, which he described as the “world’s first reasoning video model” and the “world’s first model” able to generate “in 4K and HDR.” He said Luma is also developing “agent models” for end-to-end creative workflows.
On PCs, Su announced the Ryzen AI 400 Series, describing it as the “broadest and most advanced family of AI PC processors,” with up to 12 Zen 5 CPU cores, 16 RDNA 3.5 GPU cores, an XDNA 2 NPU delivering “up to 60 TOPS,” and faster memory support. She said the first systems begin shipping “later this month,” with more than “120” ultra-thin gaming and commercial PCs launching throughout the year from major OEMs.
Liquid AI CEO and co-founder Ramin Hasani announced Liquid Foundation Models 2.5, describing them as “tiny class” on-device models at “1.2 billion parameters,” optimized for AMD Ryzen AI CPUs, GPUs, and NPUs. He also previewed LFM 3, a natively multimodal model designed for text, vision, and audio input with sub-100-millisecond latency for audio-visual data, expected “later in the year.”
Su also introduced “Ryzen AI Halo,” a compact reference platform for local AI deployment powered by Ryzen AI Max with 128GB unified memory, capable of running models up to 200 billion parameters locally. She said Halo launches in the “second quarter of this year.”
Healthcare, robotics, space, and U.S. supercomputing initiatives
In healthcare, Su hosted a panel with Absci CEO Sean McClain, Illumina CEO Jacob Thyssen, and AstraZeneca’s Head of Molecular AI Ola Engkvist. McClain described using generative AI to “engineer biology” for drug discovery and said Absci has been able to “screen over a million drugs in one single day,” citing AMD-backed compute scaling. Thyssen said Illumina’s sequencing generates more data than YouTube produces daily and noted the company uses AMD FPGA and EPYC processors in its sequencers. Engkvist said AstraZeneca is applying AI end-to-end and claimed it can deliver candidate drugs “50% faster” by using generative models to virtually assess and narrow millions of candidates before lab validation.
In physical AI, Generative Bionics CEO and co-founder Daniela Pucci introduced “Gene One,” a tactile humanoid robot designed for human-robot collaboration and healthcare use cases. Pucci said the first commercial humanoid will be manufactured in the “second half of 2026,” and that the company is working with industrial partners, including a “leading steel manufacturer,” for deployments in safety-critical environments.
Blue Origin Senior Vice President of Lunar Permanence John Kulluris described space as “the ultimate edge environment” and said AMD’s embedded architecture helps reduce mass and power while addressing radiation constraints. He said Blue Origin began working with AMD on the Versal 2 platform for flight computers, and within a few months incorporated units into development flight computers that are flying in a vehicle test bed. Kulluris said Blue Origin’s Mark 2 lunar lander could land astronauts “as early as 2028,” and highlighted interest in edge AI for hazard identification, landing site selection, and far-side lunar radio astronomy exploration.
Su also discussed AMD’s position in supercomputing, saying AMD powers “the two fastest supercomputers in the world” and “more than half of the 50 most energy-efficient systems.” She cited examples including climate modeling improvements on Finland’s LUMI system, energy research at Eni, and virus-mutation modeling on El Capitan at Lawrence Livermore National Laboratory.
Finally, Su highlighted the “Genesis” mission, a U.S. program launched “late last year” to accelerate convergence of AI, supercomputing, and quantum computing. White House science and technology policy advisor Michael Kratsios said Genesis is intended to “double the productivity and impact of American science within a decade,” and outlined priorities from President Trump’s AI action plan, including removing regulatory barriers, addressing AI infrastructure and energy production, promoting “AI diplomacy” through an “American AI Export Program,” and expanding AI education initiatives. Su said AMD committed “$150 million” to programs that bring AI into classrooms and pledged free AI courses aimed at reaching “over 150,000 students” this year.
About Advanced Micro Devices (NASDAQ:AMD)
Advanced Micro Devices, Inc (NASDAQ: AMD) is a global semiconductor company that designs and sells microprocessors, graphics processors, chipsets and adaptive computing solutions for a broad set of markets. The company’s product portfolio includes consumer and commercial CPUs under the Ryzen and Threadripper brands, data center processors under the EPYC brand, and Radeon graphics processing units for gaming and professional visualization. AMD also offers semi-custom system-on-chip (SoC) products for gaming consoles and other specialized applications, and provides supporting software and platform technologies for OEMs, cloud service providers and end users.
Founded in 1969, AMD has evolved from a supplier of logic chips into a diversified, fabless semiconductor designer.
