AMD's new business PC chips are built to run AI agents without the cloud

What AMD is shipping

As IT Brief Australia reported, new business PCs built on AMD's Ryzen AI Max and Ryzen AI Max PRO 400 Series processors are now available from AMD's partners. The pitch is that these machines can run AI agents on the device itself instead of sending every request to a cloud service.

An AI agent differs from a chatbot. A chatbot answers a question. An agent can carry out multi-step tasks across several applications, and if set up that way, it can take some actions on its own and keep working on a long project while you do something else.

The chips have up to 16 "Zen 5" CPU cores, 40 graphics compute units and an NPU rated at 55 TOPS. An NPU is a part of the chip designed for AI calculations. They also share one large pool of memory between the processor and the graphics unit.

Why memory is the headline

Agent workloads can be hungry for memory. One job may combine several models, a very long context window (the amount of text a model can consider at once), company data and multiple tools. All of that has to fit somewhere.

The previous generation topped out at 128GB of total memory, with up to 96GB reserved for graphics work. The 400 Series goes up to 192GB, with up to 160GB for the GPU. AMD describes that as an increase of 1.67x.

AMD also says the PRO 400 Series is the first x86 client processor able to run models with more than 300B parameters at 4-bit quantization, with no cloud offload. Parameters are the internal settings that make up a model, and quantization shrinks them to lower precision so the model fits in less memory. This is AMD's own claim, and the source gives no speed figures, so how usable such a large model feels on one PC is still unknown.

Building and testing agents on one machine

The article says a working business agent needs more than a model. Teams may need to link it to internal documents and code, build retrieval-augmented generation apps (where the model looks up your documents before answering), compare models, create connectors, and test how the agent behaves.

AMD positions these processors as a single place for that whole cycle. The platform works with Windows and Linux. The wider software ecosystem includes well-known tools: PyTorch, vLLM, llama.cpp, Ollama, ComfyUI and LM Studio.

For readers who run open-weights models, which are models whose files you can download and run yourself, that tool list matters. It suggests that the usual local-model workflow is meant to work on this hardware without special software.

Local versus cloud: the hybrid argument

The piece does not argue for dropping the cloud. It says top-tier cloud models still help when a task needs abilities local models lack. The idea is a mix, with some steps local and others in the cloud.

HP is the example. It is pairing the Ryzen AI Max+ PRO 495 with Perplexity Portable Computer in the HP ZBook Ultra G3a 16. That laptop can run local models and agents next to professional applications, and users can call a cloud model for steps that need frontier-level capability.

The privacy case is that an agent working on internal documents, engineering files or source code can handle suitable operations on the machine, instead of sending each request and document to an outside service.

The cost case is that agents use many more tokens than chatbots. Tokens are the small chunks of text models read and write. An agent may reason, revise, retrieve and call models repeatedly in one job, and in the cloud each of those is a recurring charge.

What to take from it, and what to doubt

AMD offers a Tokenomics Calculator that estimates costs for cloud-only, local-only and hybrid setups, using your user numbers, token use, hardware needs and cloud pricing. The article says cloud costs can pass local costs fairly quickly. That is AMD's illustration, not an independent test, so plug in your own numbers before believing it.

Local also does not mean safe by default. An agent that can open files and operate apps still needs limits. Before any rollout, ask what data it can read, what actions it can take, whether you can see a log, and whether you can switch it off.

The story is partly a product announcement, so treat the performance claims as a starting point for your own tests.