For thirty years, the word "laptop" has meant roughly the same thing: a portable computer that runs the software you point it at, one task at a time, waiting for your input. That definition is now wrong for a growing class of hardware, and the industry has not caught up with vocabulary that reflects it.
What is arriving on desks and in bags right now is not a faster laptop. It is a machine that can run multiple autonomous AI agents concurrently, persistently, and entirely offline — agents that plan, retrieve, execute tool calls, and hand work off to each other without a round trip to a cloud API. Call it what it is: an agent computer. And the silicon making this possible — AMD's Ryzen AI Max and Ryzen AI Embedded X100/X199 series, alongside Intel's Core Ultra with its own AI acceleration path — is the actual story, even though most coverage is still talking about "AI PCs" as if they are laptops with a chatbot bolted on.
The Shift Nobody Named Correctly
"AI PC" was always a marketing category, not an architectural one. It described a machine with an NPU present, not a machine capable of doing anything qualitatively different from what came before. An agent computer is a different claim: it is a device where the memory capacity, the memory bandwidth, and the heterogeneous compute (CPU + GPU + NPU) are sized specifically for running large language models as standing infrastructure, not as an occasional cloud API replacement.
The distinction matters because agentic workflows are not single-shot inference. A multi-agent system running locally might have a planning agent, a retrieval agent, a code-execution agent, and a critique agent all resident in memory, each potentially backed by a different model size or specialization, all expected to respond in the seconds-not-minutes range a human collaborator would tolerate. That is a fundamentally different workload profile than "type a prompt, wait for a paragraph." It looks more like running a small department than running an app.
The Memory Wall Just Moved
The single biggest constraint on local LLM inference has never really been raw compute — it has been memory capacity and bandwidth. A 70B parameter model needs on the order of 140GB at FP16, or something in the 35–40GB range aggressively quantized. A 300B+ parameter model, the range where genuinely capable reasoning and coding performance starts to show up even in open-weight releases, pushes past what any consumer discrete GPU has ever offered, let alone a laptop.
AMD's Ryzen AI Max and the newer Ryzen AI Embedded X100/X199 series attack this problem directly with unified memory architecture, configurable up to 192GB, shared coherently between the CPU, the RDNA integrated GPU, and the XDNA 2 NPU. That is not a typo relative to what discrete GPU buyers are used to — a single high-end discrete card tops out well below that figure, and stacking multiple cards to get there means a workstation chassis, a server-grade power supply, and a very different capital outlay than a device you close and put in a bag.
What 192GB of unified memory actually buys you is the ability to load a 300B+ parameter model, quantized to a reasonable bit depth, entirely in memory with headroom left for KV cache growth across a long agentic context window — the kind of context window multi-step tool-calling workflows chew through quickly. This is the architectural unlock that makes "agent computer" a meaningful category rather than a rebrand: the model does not have to be small enough to fit a phone, and it does not have to phone home to a datacenter to be capable.
Performance at the Edge: Where Zen 5 and XDNA 2 Actually Win
Memory capacity gets you in the door. Token generation speed determines whether the experience is usable. This is where AMD's current-generation combination of Zen 5 CPU cores and the XDNA 2 NPU has been posting a real, measurable advantage over Intel's Core Ultra parts in local inference benchmarks — specifically in llama.cpp and Vulkan-backed inference paths, which is where most of the open-source local-AI ecosystem (Ollama, LM Studio, and the broader GGUF-model tooling) actually runs today.
The reason is not mysterious. XDNA 2 was designed with a substantially higher INT8/BF16 throughput ceiling than the NPU blocks in prior Ryzen AI generations, and Zen 5's wider vector units and improved memory controller give the CPU-side fallback path — which matters whenever an inference framework has not yet been optimized to route through the NPU — meaningfully better throughput than Intel's current Lunar Lake and Arrow Lake-based Core Ultra parts on the same workload. In practical terms, this shows up as faster tokens-per-second in decode, which is the phase agentic workflows spend the most wall-clock time in, since it scales with response length rather than prompt length.
This is a genuine, benchmarkable edge today. It is worth stating plainly rather than hedging it into vagueness: if your priority is local token generation speed on the Vulkan/llama.cpp stack right now, AMD's current silicon is ahead.
Intel's Different Bet: Open, Heterogeneous, Multi-Vendor
Intel is not chasing AMD's unified-memory-capacity number, and it would be a mistake to read that as Intel losing the race rather than declining to run it. Intel's Core Ultra strategy centers on a prefill/decode split across heterogeneous accelerators — using the NPU for the compute-bound prefill phase and the integrated GPU (or a discrete Arc GPU, or in datacenter contexts, a completely separate accelerator) for the memory-bandwidth-bound decode phase — built explicitly on open standards (OpenVINO, ONNX Runtime) rather than a single-vendor stack.
The bet here is architectural flexibility over raw local capacity: Intel's approach is designed to scale from a laptop NPU up through datacenter-class heterogeneous infrastructure using the same software abstraction layer, and to let the prefill and decode stages be served by different hardware — potentially different vendors' hardware — without rewriting the inference pipeline. For an enterprise that wants a coherent AI infrastructure story from edge to datacenter, rather than a single very capable laptop, that is a real and defensible strategy. It optimizes for portability of the software stack rather than for maximizing what fits under one chassis lid.
What This Means If You Are Choosing Hardware Now
If your requirement is running the largest possible local models — 300B+ parameter class, multi-agent, entirely offline, no cloud dependency — the unified memory capacity on AMD's Ryzen AI Max and Ryzen AI Embedded X100/X199 platforms is currently unmatched in the mobile/workstation category, and the Zen 5 + XDNA 2 combination gives you real throughput advantages on the tooling most local-AI practitioners already use.
If your requirement is building infrastructure that spans laptop to datacenter with a consistent, vendor-neutral software stack, and you are willing to trade some raw local memory ceiling for architectural openness, Intel's heterogeneous prefill/decode approach is the more coherent long-term bet.
Either way, the underlying claim is the same: the device on your desk is no longer just running software. It is capable of running a standing team of agents, persistently, locally, at parameter counts that would have required a rack of GPUs two years ago. The vocabulary will catch up eventually. The hardware already has.