J

ack Huynh stood on stage in Berlin and pitched a way for companies to stop paying the cloud a monthly bill for AI. AMD's senior vice president and general manager of Computing and Graphics called the Threadripper Halo Station "the most powerful workstation in the world," a desk sized computer meant to train and run large AI models on site instead of renting compute from a cloud provider (Storage Review, 2026).

Here is why that matters to a business. Cloud AI compute is a recurring cost that scales with how much a company uses it. A machine sitting in the office is a one time purchase. For a company running AI workloads constantly, owning the hardware can cost less over time than renting it by the hour. For companies handling sensitive data, healthcare records, financial data, government contracts, keeping AI processing on a machine the company controls also answers a compliance question that a cloud contract cannot always answer as cleanly.

Two Machines Built for the Same Job, on Different Timelines

AMD's headline figure is 2.6 terabytes of combined memory, more than three times what NVIDIA's competing machine, the DGX Station GB300, offers at 748GB. NVIDIA's machine already ships, through partners including ASUS, Dell, HP, Gigabyte, MSI, and Supermicro, priced between roughly $90,000 and $115,000 (VRLA Tech, 2026). AMD's machine is targeted for 2027, and the company has not published a price.

Comparing the two memory figures side by side takes one more step.

The two machines organize that memory differently, and the difference decides which one fits a given team's work.

How Each Machine Organizes Its Memory

Think of memory the way you'd think of storage space in an office. NVIDIA's machine gives you one large connected room, 748GB that the computer's brain and its AI chip share and reach equally. AMD's machine gives you four separate storage units instead, each one holding 144GB. Software can be built to spread a task across all four units, and companies already run workloads this way in production today. AMD's 576GB accelerator figure describes those four units added together, a different kind of total than NVIDIA's single connected room.

For a company comparing the two machines, the useful question is which arrangement matches the software your team already runs. A team whose tools split tasks across separate GPU pools gets full value from AMD's larger total. A team built around one shared memory pool gets that experience from NVIDIA's machine today.

Cooling and Facility Needs Differ Too

The parts inside AMD's machine draw 1,550 watts of power on their own, before counting fans, drives, or normal electrical losses (HWBusters, 2026). AMD cools that load with liquid, water lines running through the tower, to keep the system quiet on a desk. NVIDIA's competing machine draws about 1.6 kilowatts total and cools itself with regular air on a standard office outlet. Worth checking with facilities either way: liquid cooling in an office setting is newer ground for most IT departments than air cooling is.

"This is the most powerful workstation in the world."

What's Not Confirmed Yet

AMD has not announced a price. The company has not said what operating system the machine runs, how much storage it includes, or what software comes with it, all listed as to be confirmed in AMD's own early briefings (ServeTheHome, 2026). Huynh called the unit shown at IFA a prototype. AMD's own materials say two of the four planned AI chips are working today, with support for all four coming later. The 2027 target date comes from AMD's own roadmap, and multi-year hardware plans commonly shift as engineering work continues.

What to Watch For

Three things will clarify whether this machine fits a real purchasing decision. A published price, which sets the comparison against NVIDIA's current offer. Confirmation that the software runs well across all four AI chips, not the two shown at IFA. And whether the 2027 date holds as AMD moves from prototype to shipping product.