AMD (AMD) AMD Advancing AI 2026 Keynote summary

Industry and Market Trends

• AI demand is accelerating rapidly, with a 158x increase in tokens consumed monthly over two years and training compute (FLOPS) increasing 5x annually since 2020, driving a projected $2T total compute TAM by 2030.

• The AI accelerator market is projected to reach $1.4 trillion by 2030, with data center AI accelerator and CPU TAMs growing at >45% and >50% CAGR respectively, and server CPU market forecasted to grow over 50% to $200 billion by 2030.

• Inference workloads have surpassed training as the primary AI workload, with agentic AI expected to account for 60% of global AI compute in 2026 and further increasing compute needs.

• AI is transforming industries including healthcare, manufacturing, finance, education, science, software, agriculture, and media, with workloads distributed across cloud, data center, client, and edge environments.

• Collaboration, open ecosystems, and partnerships are emphasized as essential for scaling AI infrastructure and innovation.

Product and Technology Announcements

• Helios, a rackscale AI infrastructure solution, was launched, offering up to 2.9 exaFLOPS FP4 compute, 31TB HBM capacity, 15% more compute, 50% more memory capacity and bandwidth, and 50% more scale-out bandwidth than competitors, with shipments starting late Q3 and ramping in Q4.

• The Instinct MI455/MI455X accelerator, built on advanced 2nm/3nm process, delivers industry-leading GPU performance, up to 34x higher inference throughput, and up to 18x lower token cost versus previous generations, integrated with CPUs and networking for rack-scale AI.

• Venice, the 6th-gen EPYC CPU family, introduces Zen 6 cores, up to 256 cores per chip, up to 1.8x performance, 512 threads, and 1.6TB/s memory bandwidth, targeting agentic AI, cloud, and enterprise workloads.

• ROCm AI/ROCm.ai, an agentic AI software platform, enables AI-assisted GPU programming, automates kernel optimization, delivers 3.3x average inference and 2.4x average training performance improvements, and supports 3M+ models on Hugging Face.

• MI350P, an air-cooled GPU for enterprise AI, supports up to 260B parameter models and delivers 4x more tokens per dollar than competitors, enabling LLM-scale inference in existing data centers.

Strategic Partnerships and Customer Deployments

• Anthropic will deploy up to 2 GW of Helios, OpenAI up to 6 GW of AMD infrastructure, and Meta is a lead partner for Venice CPUs and MI450 accelerators, all emphasizing co-design and early collaboration.

• Cerebras and AMD announced a partnership to combine Helios racks with Wafer Scale Engine for ultra-low latency inference, delivering 5x throughput improvements.

• Partnerships with leading AI companies and infrastructure providers deepen, including collaborations with Microsoft, Oracle, HPE, Lenovo, and Cisco.

• AT&T shared its experience deploying open source models on AMD, launching OTel 2.0, and achieving enterprise-scale AI with reduced token costs and data sovereignty.

• Case studies show up to 43% cost savings and 2.9x faster responses by routing workloads to the optimal model and infrastructure.