Enterprise AI no longer has to rent the cloud; at least, those who do not want to now have a deskside answer. The ASUS ExpertCenter Pro ET900N G3 carries NVIDIA's GB300 Grace Blackwell Ultra Desktop Superchip into a tower chassis. As TechRadar's review notes, the brushed-silver design nods to the legendary Mac Pro, and while pricing is unconfirmed, the 100,000 to 120,000 dollar band is being discussed.
The technical core impresses: a 72-core Arm Neoverse V2 Grace CPU, a Blackwell Ultra GPU, and a 748 GB unified pool combining 252 GB of HBM3e with 496 GB of LPDDR5X. Per ASUS's official spec page, a 900 GB/s bidirectional NVLink-C2C link between CPU and GPU removes the bottleneck, and the system reaches 20 PFLOPS of AI performance.
Thermals and power explain why this class is no home PC. Closed-loop enterprise liquid cooling covers not just the superchip but memory modules and the network card; a vertical fan blowing on optical transceivers prevents thermal throttling under sustained traffic. Energy comes from an 80 PLUS Titanium 1,600-watt PSU, with RAID 1 dual 2 TB NVMe PCIe 5.0 SSDs for storage.
Independent tests back the claims. In StorageReview's lab the box produced 4,793 output tokens per second on the MiniMax M2.7 test across 128 concurrent streams; the 215 GB GLM-5.2 ran straight from unified memory; GPT-OSS-120B hit 9,280 tokens per second with aggregate throughput above 45,000. Power draw was observed reaching 1,000-watt territory under load.
Scalability arrives via dual 400 Gb/s ConnectX-8 links: two towers cabled together form a 1,496 GB unified pool. As the ASUS product page stresses, the Ubuntu plus NVIDIA software stack targets NemoClaw agent workflows and local trillion-parameter models; Windows support is planned. As the ASUS launch announcement stresses, the box belongs to a new deskside-supercomputer class targeting local trillion-parameter models with NemoClaw support.
Same Silicon, Two Different Towers
StorageReview's comparison with the MSI WS300 is instructive: two chassis on the same GB300 baseboard landed within 2 percent on 24 of 28 tests. The difference is practical: ASUS's twin carry handles let two people move the 27-kilo giant, and the titanium supply saves a few watts at the wall. Silicon sets the ceiling, chassis sets the experience.
Whose Desk, at What Price?
Bitter questions remain: whose desk fits a 27-kilo tower needing a 20-amp circuit at a six-figure price? The answer is labs, film studios and teams fleeing cloud bills. For anyone sensitive to data sovereignty and latency, the ET900N G3 is not an alternative to the cloud but its branch office on the desk.
AI commentary
"I read StorageReview's lab tests alongside ASUS's spec sheets, and what impressed me was not raw power but engineering details like carry handles and an optics fan. My take: this machine is not a product, it is a declaration that the data center is moving to the desk."
AI assessment
The strongest counter-view says this class stays on paper: a 100,000-dollar Arm-based Ubuntu box demanding a 20-amp circuit, with a shared power budget that tightens once a big add-in GPU is installed and no display output outside BMC. Against hourly cloud rent the break-even stretches over years; for most teams renting still wins.
The second limitation is platform reality: tests ran on one vLLM build with the same models, while real enterprise loads are messy and impatient. The 1,496 GB two-tower pool looks great on paper but adds 800 Gb/s optics, a second tower and electrical work to the bill. StorageReview itself has not yet tested the high-power RTX PRO scenario.
The practical takeaway is buyer profile: if local inference, data privacy and deterministic latency matter more to you than the rental bill, this tower makes sense; otherwise the cloud wins. My measure: GB300 reached the desk, but the desk needs a 20-amp socket underneath and a six-figure budget on top.
Sources
5 links; no other published story cites them. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
asus · gb300 · blackwell ultra · workstation · local ai · dgx station