AMD says its next AI PCs could run agents locally, including models above 300 billion parameters

AMD is pitching the AI PC as a place for agents to keep working, not just a machine for occasional prompts. The company says its latest Ryzen AI Max PRO 400 Series can run models with more than 300 billion parameters at 4-bit quantisation without cloud offload, a claim that puts memory capacity at the centre of its agent-computing push.

AMD Watch analysis

What happened

In an interview with CRN Asia, AMD senior vice-president Vinay Awasthi described a shift from AI assistants towards “Agentic Computers”: PCs running agents locally and continuously. He said AMD has more than 50 Agentic PC designs in market and that the Ryzen AI Max PRO 400 Series supports up to 192GB of system memory.

Awasthi also argued that businesses will need to budget for the computing agents consume, not only for staff numbers or software seats. In his view, multiple agents can run in parallel, increasing demand for compute and tokens. AMD’s proposed answer is a mix of local and cloud processing, with workloads placed according to cost, latency, data sensitivity and scale.

Why it matters

Local inference could reduce cloud costs and keep some sensitive processing closer to the organisation. But AMD’s vision also points to a less obvious expense: agents that keep reasoning and calling tools use computing capacity even when nobody is sitting at the keyboard. The practical question for buyers is not simply whether a PC has an AI label, but what it can run, for how long, and at what cost.

Our read

The memory figure and model-size claim make this more than another forecast about agents taking over the office. They are AMD’s claims, not independent performance comparisons, and the interview does not provide a workload-by-workload cost test. Still, it gives buyers something concrete to ask about: whether local systems can run useful agent workflows without shifting the bill from a cloud provider to a more expensive PC fleet.

What to watch

• Independent tests of the Ryzen AI Max PRO 400 Series on real agent workflows.
• Whether AMD’s 50-plus designs translate into available systems and customer deployments.
• How organisations compare local running costs with cloud inference and the cost of upgrading PCs.

Discussion spark: For enterprise AI agents, should companies prioritise running workloads locally for control and lower cloud use, or keep them in the cloud until local systems prove cheaper in practice?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.