AMD commercially launched Ryzen AI Max PRO 400 — "Gorgon Halo" — in the IFA 2026 Berlin opening keynote, with up to 192 GB of unified memory aimed at running larger than 300-billion-parameter language models locally without a cloud hop or a discrete GPU.

HP and Lenovo commercial systems are targeted for the third quarter of 2026, before the end of September, across AI PCs, mobile workstations, and small-form-factor desktops; a consumer non-PRO Gorgon Halo SKU has not been announced.

Call the product by the name on the commercial badge. Ryzen AI Max PRO 400. Gorgon Halo is the architecture label AMD used in the IFA opening keynote. TechTimes filed the launch on 4 September 2026 around one number that changes the client AI conversation: up to 192 GB of unified memory on a PRO mobile and desktop-class platform, enough — on AMD's claim — to hold and run language models north of 300 billion parameters on the machine in front of you. No cloud round trip for the weights. No second-slot dGPU required for that class of local load. That is the peg.

Tap to enlarge

Form factors are the delivery story. AMD is not pitching a single halo tower here. The PRO 400 brief covers AI PCs, mobile workstations, and small-form-factor desktops in the commercial channel. HP and Lenovo are the named system vendors in the TechTimes filing, with commercial systems targeted for Q3 2026 — meaning before the end of September on the calendar we are in. If those dates hold, the machines land inside the same quarter as the keynote, not as a 2027 brochure exercise.

What this story is not: the Threadripper Halo Station. That IFA object — Instinct accelerators, liquid loops, HBM in the hundreds of gigabytes, trillion-parameter workstation claims — is a separate filing with its own sockets. Gorgon Halo PRO 400 is the unified-memory client and pro path for local models on HP and Lenovo commercial metal. Keep the nouns apart. One is a workstation station with discrete Instinct. This one is Max PRO silicon with a 192 GB unified ceiling aimed at more than 300-billion-parameter local inference and work without cloud dependence.

Memory market reality is the cold water in the same brief. LPDDR5X has been short. Contract prices were reported around 146 percent higher year on year by mid-2026. Twenty-four-gigabyte packages have been tight enough that Apple cut high-memory Mac Studio configurations in the same shortage cycle. AMD can put 192 GB on a slide. Whether every HP and Lenovo config ships at that ceiling on day one is a supply question the keynote did not close. Pricing for the 192 GB configurations was not announced in the TechTimes launch file. I'm not inventing a street price.

Why commercial first is a product decision with a reason. PRO branding, HP and Lenovo commercial systems, AI PC plus mobile workstation plus SFF desktop — that is an IT buyer path, not a weekend Steam Machine path. No consumer non-PRO Gorgon Halo part was announced alongside the PRO 400 launch. Enthusiasts who want the architecture without the PRO badge will have to wait for a separate reveal if AMD chooses to make one. The September commercial window is what exists on the record.

Local models are why the memory number matters. A 300-billion-plus parameter model at usable precision eats DRAM the way older workstations ate hard-disk scratch. Unified memory on an AI Max-class APU lets the CPU and on-die GPU/NPU complex share one pool instead of copying tensors across a PCIe wall to a discrete card. AMD's commercial claim is that 192 GB is enough headroom to keep those weights local. TechTimes carried the greater-than-300B figure with the launch. Independent tokens-per-second charts were not the point of the keynote file — the point was the memory ceiling and the OEM timing.

The inventory, then. Unified memory: up to 192 GB. Model class claimed locally: greater than 300 billion parameters. OEM targets: HP and Lenovo. Calendar: Q3 2026 commercial systems, before end of September. Form factors: AI PCs, mobile workstations, SFF desktops. LPDDR5X contract price move cited in the shortage context: about +146 percent year on year by mid-2026. Mac Studio high-memory cuts: same shortage cycle, as reported. Consumer non-PRO Gorgon Halo: not announced. 192 GB config pricing: not announced.

A commercially launched PRO platform that puts nearly two hundred gigabytes of unified memory into AI PC and mobile-workstation shapes is a concrete step for private local inference. Enterprises that cannot ship every prompt to a public endpoint get a vendor-named path — AMD silicon, HP and Lenovo chassis, Q3 timing — instead of only a cloud invoice. That is infrastructure progress even before the first independent LLM bake-off lands.

Constraints stay honest. Memory shortage is real. Twenty-four-gigabyte LPDDR5X packages being tight is real. Unannounced pricing on the top memory configs is real. No consumer non-PRO SKU is real. None of that cancels the launch; it bounds what you can promise a purchasing manager on Monday morning. The honest brief is: commercial Gorgon Halo PRO 400 is launched, OEMs are named, the memory ceiling is the feature, and the bill of materials will argue with the spot market.

IFA Berlin's opening keynote is the venue stamp. TechTimes dated 4 September 2026 is the primary. AMD's commercial language is Ryzen AI Max PRO 400. The architecture nickname is Gorgon Halo. If you only remember one figure, remember 192 GB unified and the greater-than-300B local-model claim attached to it. If you only remember one calendar line, remember HP and Lenovo commercial systems targeted before the end of September 2026.

Software and framework posture will decide whether those 192 GB get used for production LoRA work or for demo reels. The launch file does not hand you a day-one model zoo or a latency table. It hands you silicon positioning, memory, OEMs, and form factors. I'll take that inventory. Review units and partner BIOS drops will write the second chapter.

AMD Ryzen AI Max PRO 400 (Gorgon Halo) commercially launched at IFA 2026 Berlin: up to 192 GB unified memory for greater-than-300B-parameter LLMs locally without cloud or dGPU — AMD's claim; HP and Lenovo commercial AI PCs, mobile workstations, and SFF desktops targeted Q3 2026 before end-September; no consumer non-PRO Gorgon Halo announced; LPDDR5X shortage context (about +146 percent YoY contract prices; tight 24 GB packages); 192 GB pricing not announced. This piece stays on PRO 400 and local models — not the separate Threadripper Halo Station story. Source: TechTimes, 4 September 2026. The memory ceiling is the news. The OEM calendar is the test.