AMD's Halo Station answers DGX Station with 576GB of HBM3E

Share
AMD's Halo Station answers DGX Station with 576GB of HBM3E

AMD used its IFA 2026 keynote in Berlin to take a direct swing at Nvidia's DGX Station, and the pitch was unusually blunt about what the fight is actually over: memory.

AMD unveiled the Threadripper Halo Station, a deskside workstation that pairs a 96-core Threadripper PRO 9995WX with a pair of liquid-cooled Instinct MI350P accelerators and 2TB of DDR5 system memory. Jack Huynh, AMD's senior vice president and general manager of Computing and Graphics, called it "the most powerful workstation in the world" and claimed it is capable of running models with more than a trillion parameters — a claim AMD has not backed with a named model or a benchmark. The spec sheet is the argument: each MI350P carries 144GB of HBM3E at up to 4 TB/s, giving 288GB as shown, with what AMD calls a "path to four" accelerators for 576GB. AMD's own product page quotes up to 2.6TB of combined memory and 16.4TB/s of aggregate bandwidth, and puts the machine at up to 3.4x the total system memory of a DGX Station — a comparison it bases on a four-accelerator config against Nvidia's GB300 box with 252GB of HBM3E and 496GB of LPDDR5X.

Two details tell you how early this is. AMD's page says "coming in 2027," and its own FAQ describes Halo Station as "a prototype system AMD is showing for the first time at IFA 2026." The unit on the show floor was a conventional glass-sided tower with a standard ATX power supply and a custom loop on each card — the liquid cooling exists because, as Huynh put it, two 600W accelerators plus a 350W CPU is 1,550W of silicon that has to stay quiet on a desk. There is no price, no availability window, and no named OEM partner; AMD says nothing about how the accelerators talk to each other, what storage ships, or what the machine draws at the wall.

The economics are the part AMD didn't address. Tom's Hardware priced the core components at over $100,000 and noted the whole system could easily clear $150,000 — roughly $11,000 to $12,000 for the 9995WX, an estimated $20,000 per MI350P, and about $50,000 for the 2TB of DDR5 alone at today's memory prices. That last number is the story behind the story: AMD is selling capacity in the middle of a DRAM and HBM shortage that is already pushing workstation prices up across the board, and it is doing it a year before the thing ships. It arrives days after Lenovo puts a 120-billion-parameter model on a laptop with Nvidia's RTX Spark — Nvidia is attacking local AI from the laptop up, AMD from the server tray down, and both are betting that what teams want on-prem is not FLOPS but somewhere to put the model.

What to watch: which OEM picks this up, and at what price. A 2027 ship date means AMD is selling a roadmap against hardware Nvidia is shipping now.

Would you pay six figures to keep a trillion-parameter model on your own desk instead of renting it by the token? Tell us in the comments.

Read more

Akhetonics says its all-optical CPU reaches a customer in 2026

Akhetonics says its all-optical CPU reaches a customer in 2026

Light-based computing keeps promising more than it delivers — but one Munich startup has just put a date on its bet, and the interview laying it out is doing the rounds on Hacker News this week. Akhetonics says it will deploy its first commercial machine with a major customer by the end of 2026, with several more planned for 2027. The company, founded by Michael Kissner and Leonardo Del Bino, is building a computer where data enters as light, is switched as light, and circulates through memory

The Week in AI — October 5–11, 2026

The Week in AI — October 5–11, 2026

Every big claim this week turned out to rest on fine print more interesting than the headline: revenue only the company reporting it can define, safety tests sandboxed while the product keeps the web, and a Pentagon phase-out nobody would confirm until reporters kept asking. The week's top 5 1. OpenAI's revenue was $20 billion below the numbers everyone quoted — and the gap was definitional. The Financial Times reported Thursday that OpenAI's annualized revenue runs roughly $20 billion unde

Drone strike shuts a third Yandex data center, taking YandexGPT offline

Drone strike shuts a third Yandex data center, taking YandexGPT offline

Russia's largest tech company is learning what the AI era's infrastructure war looks like from the receiving end — three data centers in four days, and with them much of the cloud layer Russian businesses run on. A Ukrainian drone strike knocked out Yandex's data center in Vladimir early Sunday morning, the third of the company's facilities hit since October 8. The site — reported at roughly 50 MW and designed for about 2,880 server racks — stopped operating completely after the attack, Yandex

Agent teams cost up to 5x more, barely score higher

Agent teams cost up to 5x more, barely score higher

The multi-agent hype train hit a benchmark this weekend — and the grid and the trucking regulators had quiet weeks of their own. Vals AI put agent teams head-to-head with single agents on its Vibe Code Bench, and the teams cost between 1.8 and 5.1 times more for almost no extra quality. The evals company ran GPT-6 Sol and Claude Opus 5.5 solo and in teams across 50 apps at two reasoning efforts; out of four comparisons, only one was statistically significant — Sol at medium effort, where the t