AI agents are creating a new challenge for data center operators: more of their work happens outside the AI models. Database queries, code execution, and other application tasks require CPU resources, even when GPUs handle the model responses.
AMD is targeting these computing requirements with Helion, an open rack-scale platform announced October 12. Its configurable CPU and GPU systems could help companies allocate processing capacity according to what their agents actually do, potentially reducing the cost of completing automated tasks.
Why AI agents need different computing resources
While GPUs process the calculations needed to generate model responses, CPUs handle operations such as database queries and code execution.
A personal assistant might execute hundreds of relatively simple actions, while a technical agent could spend more time running complex calculations. AMD’s Helion documentation distinguishes between agents that require high processing volumes and those that need faster individual execution.
Additional GPUs offer limited benefit when agents spend much of their time waiting for CPU-intensive operations to finish. Companies therefore need sufficient processing capacity for both activities rather than expanding inference hardware alone.
Two deployment options for different AI workloads
Hyperscale operators with existing inference infrastructure can deploy CPU-only Helion racks dedicated to agent execution. Enterprises can choose combined CPU/GPU racks to run agents alongside locally hosted models, with general-purpose servers supporting other app workloads.
According to AMD, Helion configurations can support up to 32,768 Zen 6 CPU cores, while GPU-equipped systems can include up to 16 Instinct MI350P GPUs. The company also lists networking bandwidth of up to 102 terabits per second. Note that these figures represent Helion’s maximum supported configurations rather than its performance in real-world deployments.
Helion also serves a different purpose from Helios, AMD’s accelerated computing platform announced in July. Helios combines 72 MI455X GPUs with EPYC processors for demanding AI model workloads. Helion focuses on models' processing requirements, with configurations that accommodate varying inference and agent-execution loads.
Model selection could influence operating costs
Helion also includes an integrated token router that directs model requests between cloud-based frontier models and open-source alternatives, including locally hosted models. Companies could use less expensive models for routine operations while reserving more capable models for tasks that require advanced reasoning.
For an organization running thousands of agents, the model chosen for each request could significantly affect inference spending. However, locally hosted models consume electricity and require infrastructure investment, while less capable models may need several attempts to complete a task correctly.
Companies need to evaluate the cost per completed task, including failed attempts and infrastructure expenses. As The Neuron previously reported, cheaper AI processing does not necessarily reduce total spending as usage increases.
AMD’s open approach enters a competitive market
AMD plans to contribute Helion’s specifications to the Open Compute Project so hardware manufacturers can develop their own configurations. This could give enterprises more supplier options and help them select systems based on their computing needs.
Dell and Cisco are among the companies supporting Helion. HPE and Lenovo also support the design, alongside Supermicro and other partners. AMD expects partner configurations to become available in November 2026, although the launch materials do not disclose commercial pricing.
Helion also positions AMD to compete with Nvidia in the growing market for agentic AI infrastructure. Nvidia offers Vera CPUs for agent execution and specialized inference hardware such as Groq 3 LPX.
Both companies address different agent workloads through specialized computing resources. AMD emphasizes an open reference design that manufacturers can customize, while Nvidia combines agent execution and inference products within its accelerated computing ecosystem. Buyers will need to assess software compatibility and vendor support alongside hardware costs.
Can Helion reduce the cost of running AI agents?
Helion’s economic case depends on whether its configurations can complete more agent tasks within a given computing and electricity budget. Dedicated CPU capacity could reduce delays between model calls, but any savings will depend on workload characteristics and deployment costs.
As of October 13, Helion's real-world performance and energy efficiency have yet to be demonstrated at commercial scale. Real-world throughput and energy consumption remain unresolved, as does availability at commercial scale.
Higher efficiency could let enterprises run more agents within existing power budgets, though increased adoption could also raise overall electricity consumption. Future data center purchases may increasingly depend on how many successful tasks each dollar and kilowatt-hour can support. Commercial deployments will need to show whether Helion delivers measurable savings.
