Asus Unveils GB300-Powered AI Workstation With 748GB Coherent Memory and $120,000 Estimated Price

Asus has released the ExpertCenter Pro ET900N G3, a heavy-duty desktop workstation built around Nvidia's GB300 Grace Blackwell Ultra Superchip. Designed to execute data-center-scale artificial intelligence models on-premises without cloud infrastructure, the system carries an estimated price between $100,000 and $120,000, according to reporting by TechRadar Pro (https://www.techradar.com/pro/an-asus-homage-to-legendary-apple-mac-pro-gb300-dgx-workstation-gets-two-thumbs-up-from-reviewer-who-calls-usd120-000-ai-supercomputer-the-most-capable-thing-weve-ever-put-on-a-desk), though Asus has not publicly disclosed official pricing.

The ET900N G3 integrates 72 Arm CPU cores, a Blackwell Ultra GPU, and 748GB of coherent memory into a single chassis. Weighing 27 kilograms and measuring 584 by 232 by 565 millimeters, the tower sports a brushed silver exterior, a dark mesh front intake, and two top-mounted metal handles to assist personnel in moving the unit. The industrial chassis design has prompted visual comparisons from reviewers to Apple's classic Mac Pro tower.

To manage thermal output during extended compute cycles, Asus equipped the workstation with an enterprise-grade sealed liquid cooling loop using copper cold plates that cover the Superchip, adjacent memory modules, and the high-speed network card. A dedicated internal fan directs airflow straight across the network card's optical transceivers to prevent thermal throttling under sustained loads. Power is supplied by a 1,600-watt power supply certified at 80 PLUS Titanium efficiency.

Storage out of the box consists of a mirrored array of two 2TB NVMe drives to safeguard system software from single-drive failures. The internal layout includes three PCIe expansion slots and two extra drive bays for auxiliary storage or supplemental accelerator cards down the road.

For cluster scaling, the workstation includes two 400-gigabit networking ports providing up to 800Gb/s of combined interconnect bandwidth. Organizations can link two ET900N G3 towers directly together to establish a single 1,496GB coherent memory pool, doubling memory capacity for extra-large AI architectures.

Hardware benchmark testing highlighted substantial throughput on localized large language models. In tests running MiniMax M2.7 across 128 concurrent streams, the system achieved 4,793 output tokens per second. On GLM-5.2, the workstation routed 215GB of expert weight data directly across the coherent memory fabric. When evaluated on GPT-OSS-120B, output reached 9,280 tokens per second with total throughput surpassing 45,000 tokens per second, scaling smoothly from one to 128 simultaneous software agents.

Early independent evaluations highlighted the machine's multi-agent local processing capability. "The GB300 DGX Station remains the most capable thing we've ever put on a desk," noted Brian Beeler, a reviewer at StorageReview. In a separate hands-on test, technology YouTuber Alex Ziskind observed total power draw approach 1,000 watts while concurrently running dozens of coding, research, and writing agents. "Everything you're about to see, dozens of AIs all thinking at once, is running on this one machine, not a data center," Ziskind said.