AMD AI Developer Day Lands in China: Lisa Su Bets on Agents as CPU-GPU Ratio Hits 1:1
On May 19, AMD held its inaugural AI Developer Day in China—the first time this event has ever left American soil.
Inside the venue, attendees without seats lined the walls in row after row. AMD Chair and CEO Lisa Su’s keynote boiled down to two things: agents and the return of the CPU.
Two Critical Judgments
Judgment One: CPUs Are Becoming More Critical Than Ever
Su delivered the numbers directly:
In traditional data centers, the CPU-to-GPU ratio is 1:4. By 2026, that ratio will become 1:1.
What does this mean?
For the past decade, AI compute narratives have been monopolized by GPUs. Training large models requires CUDA; inference acceleration depends on Tensor Cores. CPUs were marginalized as “control planes”—issuing orders but not participating in the main computation.
Su’s judgment: in the agent era, compute workloads are shifting.
- Agents aren’t single-shot inference; they’re multi-step, long-horizon task flows
- Every step requires scheduling, decision-making, and state management—CPUs excel at these
- When GPUs handle parallel computation, CPUs handle serial coordination. Both are indispensable
The 1:1 ratio isn’t a prediction; it’s a declaration. AMD is using its EPYC processor portfolio to reclaim data center discourse.
Judgment Two: China Is a Rich Vein of Compute Demand
Moving the developer conference to China isn’t market PR—it’s strategic positioning.
China’s AI application deployment velocity, enterprise agent adoption density, and compute procurement scale all rank in the global first tier. Su’s subtext on stage was clear: if you can’t win developers in China, you can’t win the AI infrastructure war globally.
AMD’s current challenges:
- CUDA’s ecosystem moat remains deep
- ROCm maturity still lags behind nvcc + cuDNN
- Developer migration costs are real
But AMD’s cards are also clear:
- EPYC’s data center CPU market share continues climbing
- MI-series GPUs offer inference cost-performance advantages
- The open-source route appeals to large enterprises seeking to avoid single-vendor lock-in
What Agents Change
Agents aren’t just “better chatbots.” Their technical characteristics drive underlying architectural shifts:
| Characteristic | Compute Impact |
|---|---|
| Multi-step tasks | Persistent state management required; CPU scheduling burden increases |
| Tool calling | Frequent I/O; memory bandwidth and latency sensitive |
| Long context | KV Cache bloat; memory capacity becomes a hard constraint |
| Multimodal input | Complex preprocessing pipelines; heterogeneous compute demand rises |
| Real-time decisions | Low-latency inference; can’t route through full GPU pipeline every time |
These characteristics collectively point to one conclusion: AI infrastructure is shifting from “GPU single-machine brute force” to “CPU+GPU collaborative fine architecture.”
AMD holds both EPYC and Instinct—this is what gives it the confidence to declare 1:1.
Competitive Landscape: Three-Way Standoff
| Vendor | CPU Advantage | GPU Advantage | Ecosystem |
|---|---|---|---|
| NVIDIA | Weak (Grace not yet widely deployed) | Dominant (CUDA ecosystem) | Closed but mature |
| AMD | Strong (EPYC) | Moderate (ROCm catching up) | Open-source route |
| Intel | Moderate (Xeon market share declining) | Weak (Gaudi limited traction) | oneAPI promotion ongoing |
AMD’s strategy is explicit: leverage CPU installed base to drive GPU incremental penetration. Agents are the perfect entry point—because in this scenario, the CPU is no longer a supporting actor.
Practical Impact for Developers
If you’re an AI application developer, what does this mean?
- Heterogeneous programming becomes standard. Pure CUDA skills are declining in value; engineers who can distribute workloads between CPU and GPU are more scarce
- Inference optimization is worth more than training optimization. In the agent era, 90% of compute consumption is inference and scheduling, not training
- Hardware lock-in risk needs hedging. ROCm, oneAPI, and CUDA will coexist for years. Don’t lock your code to a single platform
Conclusion
AMD moving AI Developer Day to China is both a signal and a bet.
The signal: in the agent era, compute architecture is being reshaped, and CPUs are no longer GPU accessories.
The bet: can ROCm, in CUDA’s shadow, leverage the agent tailwind to break through on ecosystem?
Su didn’t answer that question directly. But moving the conference to China is itself the answer—the demand here is large enough to redefine the rules.