192 GB. That's the unified memory ceiling of AMD's next-generation Ryzen AI Max 400 single-chip APU.
This means — a single APU in your PC can locally run large language models with over 300 billion parameters. It's a first for x86.
The Chip: Gorgon Halo
Codename "Gorgon Halo," officially the Ryzen AI Max 400 series, is a minor upgrade over last year's Strix Halo (Ryzen AI Max 300).
Specs at a glance:
- CPU: 16 Zen 5 cores / 32 threads, up to 5.2 GHz (vs. 5.1 GHz)
- GPU: 40 RDNA 3.5 CUs, up to 3.0 GHz (vs. 2.9 GHz)
- NPU: XDNA 2, 55 TOPS (vs. 50 TOPS)
- Max unified memory: 192 GB (vs. 128 GB)
- Usable as VRAM: 160 GB (vs. 96 GB)
The main upgrade is in the last two lines — memory ceiling raised by 64 GB. CPU, GPU, and NPU get modest frequency bumps with no architecture change. This was shaping up to be a "tick" chip, but the memory cut changes the story entirely.
Why 192 GB Is a Watershed
A 300B-parameter model, 4-bit quantized, takes roughly 150 GB of VRAM. Add KV cache and system overhead, and 160 GB of usable VRAM is just enough.
Previously, running a 300B model on an x86 client was nearly impossible:
- MacBook Pro M4 Max: max 128 GB unified memory — enough for 70B, not 300B
- Any RTX 5090/5080 setup: single card 24-32 GB; three or four cards needed for 100+ GB, costing $10,000+
- Ryzen AI Max 300: 128 GB ceiling, same tier as M4 Max
Gorgon Halo's 192 GB is the first time "running a 300B LLM locally" shifts from "you need a server" to "you can buy a laptop."
In plain terms: DeepSeek V4 Pro 1.6T is still out of reach, but Llama 3.3 405B, Qwen3.6-Max 27B, and GLM-5.1 mid-tier models are all covered by this single chip.
AMD's Real Play: Not Gaming, but AI Agents
This isn't about gaming laptops; it's about AI agent workloads.
When running agentic workloads, a long-running agent may keep a large model's weights in local memory for hours without unloading. Cloud API inference costs add up quickly per second. For professional users (developers, researchers, enterprise engineering teams), a local-capable device offers a compelling ROI.
Specific scenarios:
- Offline alternative to Claude Code / Codex: Developers run a 70B-300B model locally for code completion without sending data out
- Sensitive enterprise data analysis: Legal, medical, and financial industries that cannot use cloud models — local LLM is the only solution
- Long-running agents: A researcher running an automated literature review agent for 8 hours would burn through a cloud API budget
The XDNA 2 NPU delivers 55 TOPS, intended for "small model always-on" tasks — a 7B model handles routine jobs while the 300B main model is called on demand. It's a layered architecture.
A Less Friendly Detail
OEM shipments are scheduled for Q3 2026 — ASUS, HP, and Lenovo will be first. That means the earliest you can actually buy this chip is July.
And the 192 GB configuration won't be cheap. Strix Halo fully loaded currently retails around $3,000; Gorgon Halo fully loaded is expected at $4,000+. Add the rest of the system, and a laptop capable of running a 300B model will likely land in the $5,000–$6,000 range.
In plain terms: This chip is not for average consumers. It's for the small teams, researchers, and independent developers who previously had to buy H100s or RTX 6000 Adas.
But for that group, a $5,000 laptop that runs a 300B model is far more cost-effective than a $30,000 H100 card.
Ecosystem Implications
AMD's AI strategy over the past two years has become clear:
- Data center: MI300X / MI400 going head-to-head with Nvidia
- Consumer APU: Combining NPU + large GPU + large memory on one die to enable local model inference
Nvidia's consumer RTX 50 series still lacks a unified memory solution. AMD is effectively bypassing the frontal battle and capturing the new "local AI inference" market on a different path.
If Gorgon Halo launches in Q3 with good support from local LLM toolchains (llama.cpp, Ollama, LM Studio), this path could work.
Next thing to watch: Apple M5 Max's memory ceiling. If Apple raises the M5 Max to 192 GB or higher this fall, the local AI inference hardware war will truly begin.
Sources: CocoLoop, AMD Ryzen AI Max 400 'Gorgon Halo' packs up to 192GB of unified memory (Tom's Hardware), AMD Pushes Ryzen AI MAX 400 to 192GB Memory, Letting a Single Chip Run 300B+ Parameter LLMs Locally (WCCFTech), AMD Ryzen AI Max 400: New APU Unlocks 192GB Unified Memory (HotHardware), AMD confirms Ryzen AI MAX 400 will support up to 192GB memory and 160GB VRAM (VideoCardz)