AMD: Advancing AI 2026 Keynote

研报英文原文证据摘录

AMD: Advancing AI 2026 Keynote

July 24, 2026

Customer System Co-Design Becomes the Real Differentiator

• AMD emphasized that large AI customers are moving from buying chips to co-designing

full systems across compute, power, cooling, networking, and software.

• Meta reinforced this point, saying AI infrastructure is no longer just a GPU game and

that CPUs and GPUs need to be treated as “conjoined things.”

• Meta’s MI450 comments were especially important because it described a move from

MI300 experimentation to production-scale deployment and deeper engineering co-

design.

Inference Segmenting Into Different Workloads

• AMD indicated that inference is not one market: some workloads need maximum

throughput, some need balanced responsiveness, and some require ultra-low latency.

• This supports a broader portfolio argument, with AMD positioning different compute

architectures for different inference use cases rather than a one-size-fits-all solution.

Cerebras Highlights Ultra-Low Latency Inference

• Cerebras was the clearest proof point for disaggregated inference, combining AMD

CPUs, Helios, and the Cerebras Wafer Scale Engine.

• The partnership is aimed at customers that need both high throughput and very low

latency, with Cerebras claiming 5x throughput while maintaining speed.

OpenAI Reinforces the Need for More Compute

• OpenAI framed models as evolving from chatbots to reasoners, agents, and eventually

“interns,” which keeps pushing demand for scaled compute.

• OpenAI also said exploding token budgets are showing up internally across engineering

and broader enterprise workflows, not just in research workloads.

Enterprise AI Moves From Use Cases to Workflow Rebuilds

• AT&T said AI is moving beyond point use cases into rebuilding entire workflows across