An Asus homage to legendary Apple Mac Pro: GB300 DGX workstation gets two thumbs up from reviewer who calls $120,000 AI supercomputer 'the most capable thing we’ve ever put on a desk'

September 27, 2026 at 07:20 PM

ASUS ExpertCenter Pro ET900N G3 weighs 27 kilograms net, built strictly for a desk It combines 72 ARM cores with a Blackwell Ultra GPU inside Coherent memory totals 748GB, unifying GPU and system memory together Asus has released the ExpertCenter Pro ET900N G3 workstation , powered by Nvidia’s GB300

<![CDATA[ <article> <ul><li>ASUS ExpertCenter Pro ET900N G3 weighs 27 kilograms net, built strictly for a desk</li><li>It combines 72 ARM cores with a Blackwell Ultra GPU inside</li><li>Coherent memory totals 748GB, unifying GPU and system memory together</li></ul>Asus has released the ExpertCenter Pro ET900N G3 <a href="https://www.techradar.com/best/best-workstations">workstation</a>, powered by Nvidia’s GB300 Grace Blackwell Ultra Superchip for elite compute performance.The tower features a net weight of 27kg and dimensions of 584 x 232 x 565 mm, making it ideal for a fixed desktop setup — or permanently positioned beneath the desk if the space is well-ventilated.Portability was never part of the design goal here, since the machine combines 72 Arm cores, a Blackwell Ultra <a href="https://www.techradar.com/news/computing-components/graphics-cards/best-graphics-cards-1291458">GPU</a>, and 748GB of coherent memory into one chassis.<h2 id="a-workstation-built-with-a-look-and-cooling-system-of-its-own">A workstation built with a look and cooling system of its own</h2>The workstation features a distinct industrial design, sporting a brushed silver exterior, a dark mesh front panel, and two prominent metal handles fixed to the top edge.These heavy-duty handles allow multiple staff members to safely relocate the unit without gripping the side panels directly.Beneath the hood, Asus deploys an enterprise-grade, sealed liquid cooling loop equipped with copper cold plates to chill the Superchip, its adjacent memory modules, and the high-speed network card.Because a liquid loop alone cannot reach every hot spot, a dedicated internal fan directs airflow straight onto the network card's optical transceivers to prevent thermal throttling under sustained networking loads.Delivering power to this hardware is a massive 1,600-watt power supply certified at 80 PLUS Titanium efficiency, which drastically minimizes wall-power waste during peak computational compute cycles.Out of the box, the system relies on a mirrored array of two 2TB NVMe drives to protect the operating system and software stack from single-drive failures.For future scaling, the internal layout leaves room for three PCIe slots and two additional expansion bays to accommodate extra storage or supplemental graphics cards down the line.During hardware testing, this architecture demonstrated remarkable data throughput on large-scale AI workloads, hitting 4,793 output tokens per second across 128 concurrent streams during the MiniMax M2.7 benchmark.Massive models like GLM-5.2 pushed 215GB of expert weight data directly through the coherent memory fabric, bypassing the severe latency bottlenecks typically seen on standard system architectures.<h2 id="data-center-performance-on-a-desktop">Data center performance on a desktop</h2>Initial evaluations from hardware reviewers draw strong visual comparisons between this tower and Apple’s classic, boxy Mac Pro <a href="https://www.techradar.com/news/computing/pc/10-of-the-best-desktop-pcs-of-2015-1304391">computer</a> chassis.Yet, the actual performance metrics place this workstation in a completely different class than any standard high-end desktop computer."The GB300 DGX Station remains the most capable thing we've ever put on a desk," said Brian Beeler, reviewer at <a href="https://www.storagereview.com/review/asus-expertcenter-pro-et900n-g3-review?type=Review" target="_blank" rel="nofollow">StorageReview</a>.Running dozens of coding agents simultaneously, the machine sustained thousands of tokens per second without individual agents slowing down noticeably during testing.The device ships with two 400-gigabit networking ports, supporting up to 800Gb/s combined bandwidth for linking directly with one additional GB300 station.Together, two linked stations form a single 1,496GB coherent memory pool, doubling the standalone unit's capacity for handling larger models."Everything you're about to see, dozens of AIs all thinking at once, is running on this one machine, not a data center," said Alex Ziskind, technology YouTuber.On GPT-OSS-120B, output climbed to 9,280 tokens per second, with total throughput reaching over 45,000 tokens per second.Concurrency tests showed the system scaling smoothly from a single agent up to 128 running simultaneously without any noticeable system crash or slowdown.Ziskind demonstrated this directly on video, watching power draw climb toward 1,000 watts while dozens of coding, research, and writing agents ran simultaneously.ASUS has not publicly disclosed exact pricing for this device, though estimates so far peg this device to cost between $100,000 and $120,000.<figure class="van-image-figure inline-layout" data-bordeaux-image-check ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:676px;"><img id="diM9tpwF2Lz85R8q85CT78" name="tr-g_news" alt="Google logo on a black background next to text reading 'Click to follow TechRadar'" src="https://cdn.mos.cms.futurecdn.net/diM9tpwF2Lz85R8q85CT78-1920-80.jpg" mos="" align="middle" fullscreen="" width="676" height="213" attribution="" endorsement="" class="inline"></div></div></figure> </article> ]]>