Let me set the scene.
A week ago, ASUS quietly launched a new desktop PC. Not a gaming rig with RGB fans and a glass side panel. Not a sleek all-in-one. Just a big black metal tower that looks like it could've shipped in 2014 with a Core i5 and 8GB of RAM.
The price tag? $99,999.
That's not a typo. And here's the thing: it might actually be a bargain.
Meet the ASUS ExpertCenter Pro ET900N G3, officially launched on June 15, 2026. Built on the NVIDIA DGX Station GB300 architecture, this is not a PC in any normal sense. It's a deskside AI supercomputer that happens to fit under your standing desk.
The heart of the machine is NVIDIA's GB300 Grace Blackwell Ultra Desktop Superchip — a Frankenstein's monster of silicon that bolts a 72-core Arm Neoverse V2 Grace CPU to a Blackwell Ultra GPU via NVLink-C2C, a high-bandwidth chip-to-chip interconnect.
Here's what that buys you:
| Spec | What You Get |
|---|---|
| Coherent Memory | 748GB total — 496GB LPDDR5X (CPU) + 252GB HBM3e (GPU) |
| CPU Memory Bandwidth | 396 GB/s |
| GPU Memory Bandwidth | 7.1 TB/s (≈18x faster than the CPU side) |
| AI Compute | Up to 20 PFLOPS (FP4) |
| Max Model Size | Up to 1 trillion parameters, locally |
| Networking | Dual ConnectX-8 SuperNICs (400 Gb/s each), plus 10GbE |
| Storage | 4x M.2 2280 NVMe (2x 2TB included) |
| Expansion | 3x PCIe 5.0 (one x16, two x8) |
| PSU | 1600W 80 PLUS Titanium (requires 240V) |
| OS | Ubuntu with NVIDIA AI Developer Tools (Windows support planned) |
To put 20 PFLOPS in perspective: that's 20,000,000,000,000,000 floating-point operations per second. In FP4, the precision format optimized for AI inference, this single desktop box out-computes most server racks from just a few years ago.
As Tom's Hardware put it: this "plain-looking desktop is more powerful than most server racks."
Here's the part that actually matters for AI developers: the 748GB of coherent unified memory.
In a traditional workstation, your CPU has its RAM and your GPU has its VRAM, and they're separated by a PCIe bottleneck. You can't easily run a model that needs more VRAM than your GPU has. If your GPU has 24GB, you're stuck with models under 24GB. Period.
The ET900N G3 tears down that wall. The NVLink-C2C interconnect means the CPU and GPU share a single coherent memory pool. The Grace CPU contributes 496GB of LPDDR5X (396 GB/s bandwidth), and the Blackwell Ultra GPU adds 252GB of HBM3e (7.1 TB/s bandwidth). That's 748GB total — enough to run frontier AI models with up to 1 trillion parameters entirely on-device, no cloud required.
ASUS benchmarked the system running the massive Qwen open-source model through vLLM:
That's not just "fast for a desktop." That's "previously required a data center" territory.
Let's talk about the price, because it's the first thing everyone fixates on.
In the US: $99,999 at retailers like Central Computer.
In the UK: £119,999.99 at Scan UK (that's about $152,000 at current exchange rates — the UK always gets the short end).
Is it worth it? Depends on your math.
A single NVIDIA H200 GPU (the previous-gen data center heavyweight) costs roughly $30,000–$40,000 on its own, and you'd need multiple to match this unified memory pool. Cloud GPU instances running 24/7 for LLM inference can easily hit $15,000–$20,000/month. If you're doing serious AI work, the ET900N G3 pays for itself in five to eight months of avoided cloud bills.
Plus, you own the hardware. No data leaves your office. No surprise overage charges. No "your instance was preempted."
ASUS isn't the only OEM building on NVIDIA's DGX Station platform. The GB300 launch has spawned a whole ecosystem:
| System | OEM | Key Differentiator |
|---|---|---|
| ASUS ExpertCenter Pro ET900N G3 | ASUS | First to market, Windows support planned |
| MSI XpertStation WS300 | MSI | Same GB300 architecture |
| GIGABYTE W775-V10-L01 | GIGABYTE | Tower server form factor |
| Exxact Valence VWS-158270643 | Exxact | Configure-to-order flexibility |
And then there's the little sibling: the ASUS Ascent GX10, based on the smaller GB10 Grace Blackwell platform (same chip as NVIDIA DGX Spark), starting around $2,999–$3,999. That's the "I want to play with AI models at home and not remortgage my house" option.
For $100K, you'd expect every port under the sun. The ET900N G3 delivers on the important stuff (dual 400Gb QSFP, 10GbE, plenty of USB 10Gbps), but there are some notable omissions:
The I/O choices make sense when you realize what ASUS optimized for: raw AI throughput. The dual ConnectX-8 SuperNICs give you 800 Gb/s of aggregate networking throughput. That's what matters when you're pushing terabyte-scale model checkpoints around. USB4 can wait.
Here's my favorite detail from the Tom's Hardware comments section:
"My gaming PC looks better than that."
And that's exactly the joke. The ET900N G3 is the ultimate sleeper build. No tempered glass. No RGB. No aggressive gamer aesthetics. Just a black metal box that looks like it belongs in a 2015 office cubicle — but inside, it's running models that would've required a small supercomputer a decade ago.
This is deliberate. ASUS isn't selling to gamers who want their PC to look like a cyberpunk prop. They're selling to AI labs, enterprise deployment teams, and researchers who need a machine that sits quietly under a desk and crushes trillion-parameter models without drama.
The ET900N G3 isn't just another expensive workstation. It represents a fundamental shift in how NVIDIA is approaching the market.
Historically, NVIDIA reserved its highest-end AI hardware for its own DGX systems — the DGX Station, DGX A100, etc. You bought from NVIDIA directly, and OEMs got the leftovers.
Now, NVIDIA is opening the GB300 platform to partners like ASUS, Dell, MSI, GIGABYTE, HP, Lambda, and Supermicro. This is NVIDIA saying: "We don't need to hoard the best silicon. Let the ecosystem compete on integration, support, and pricing."
It's also NVIDIA's quiet invasion of CPU territory. The Grace CPU is ARM-based, built specifically for AI and HPC workloads. It's not designed to replace your Intel or AMD processor for general computing — but for the AI development market, it doesn't need to. Every ET900N sold is one less Threadripper workstation deployed for AI work.
Let's be real: this is not for you. Unless "you" happen to be:
For everyone else? The ASUS Ascent GX10 ($2,999+) or a cloud GPU instance will serve you fine.
ASUS has built the ultimate wolf in sheep's clothing. A PC that looks boring enough to be invisible in any office, but packs enough AI compute to run trillion-parameter models locally. At $99,999, it's expensive — but against the cost of equivalent cloud compute over 12–18 months, it's genuinely economical for the right buyer.
The bigger story: NVIDIA is democratizing its most powerful silicon. The GB300 isn't locked inside a premium DGX box anymore. It's available from ASUS, MSI, GIGABYTE, and others. That means competition, which means (eventually) lower prices, better support, and more options.
We're entering an era where "desktop PC" and "supercomputer" are starting to mean the same thing. And honestly? That's pretty damn exciting.
Sources:
