Enterprise
Agentic AI workloads shift compute demand back to CPUs
Power constraints and concurrency requirements are forcing chipmakers and server builders to rethink the GPU-centric architecture that dominated early AI deployments.
Key takeaways
- Agentic AI workloads require massive concurrency rather than raw GPU throughput, shifting compute demand back to CPUs after years of GPU dominance.
- Enterprise data centers are hitting 15-20 kilowatt rack power limits, forcing server designs that prioritize efficiency over maximum performance.
- AMD's EPYC 9006-based servers can replace 13 legacy machines while cutting cooling power by 80 percent, freeing capacity for new workloads.
- Early AI architectures that hung too many GPUs off single CPUs created data movement bottlenecks that left expensive accelerators idle.
- Agentic systems can turn one researcher's input into thousands of parallel compute jobs, multiplying infrastructure demands without adding headcount.
The rise of agentic AI is forcing a fundamental reassessment of data center architecture, with central processing units reclaiming a starring role after years of GPU dominance. The shift reflects hard lessons about power budgets, data movement bottlenecks, and the specific computational patterns that autonomous AI agents create.
Scott Tease, vice president of business development for high-performance computing and sovereign AI at Advanced Micro Devices, told SiliconANGLE that early AI deployments often starved expensive graphics accelerators by attaching too many GPUs to a single CPU. The result was idle silicon and wasted capital.
"People have realized how important that CPU is in a GPU architecture," Tease said during an interview at SC26. "But beyond that, agentic is not necessarily just a GPU thing. Agentic is going to be very CPU heavy because with agentic, what you're looking at, you want to go very fast, but you also want concurrency."
That concurrency requirement stems from how agentic systems work. Unlike monolithic AI models that process one query at a time, agents spawn thousands of parallel tasks from a single researcher's input, creating compute patterns that favor CPU architectures designed for simultaneous workloads.
Power budgets drive architecture decisions
The CPU renaissance is colliding with a hard physical constraint: data centers are running out of power. David Schmidt, vice president of PowerEdge product management at Dell Technologies, said enterprise customers told Dell 2.5 years ago that rack power would hold at 15 to 20 kilowatts — and that they often exhausted their power allocation before filling a rack with equipment.
Dell's response was the PowerEdge R9825, a 3U server built on AMD's sixth-generation EPYC 9006 processors and optimized for power and cooling efficiency. Schmidt said the design returns power to customers by reducing the energy spent on cooling itself.
The efficiency gains compound during hardware refresh cycles. A single EPYC 9006 server can replace 13 legacy machines while cutting cooling power by 80 percent, Schmidt noted. That reclaimed power becomes available for new workloads — particularly the agent-driven compute jobs that Tease expects to multiply exponentially.
"Same number of researchers, massive increase in the amount of compute," Tease said. "So agentic is going to continue to drive more and more IT, and we've got to find a way to power it."
Why it matters
The shift from GPU-first to balanced CPU-GPU architectures represents more than a technical correction. It signals that the industry is moving past the initial AI infrastructure buildout phase and confronting the operational realities of running production AI systems at scale. Companies that optimized for training throughput are now redesigning for inference concurrency, agent orchestration, and power efficiency — constraints that favor different silicon and different vendors. For chipmakers, it means business development strategies built around GPU attach rates need to account for workloads where the CPU is the primary engine.
Details of the Dell and AMD executives' remarks were first reported by SiliconANGLE during coverage of SC26.