AI workstations · on your network
Cloud AI bills you every time someone asks a question. An XAI machine sits in your building, runs the models on your own hardware and keeps your data on your network. Buy it once. The next question costs electricity.
Cloud AI bills you every time someone asks a question. An XAI machine runs the models in your building and keeps your data on your network.
Our own operating costs are running $2.8 million a year lower, with AI doing work in just about every department we have.
Our own operating costs are running $2.8 million a year lower, with AI doing work in just about every department.
XAI series
Core Ultra 9 285K · RTX 5090 32GB
64GB DDR5 · 2TB NVMe
Ryzen 7 9800X3D · RTX 5090 32GB
64GB DDR5 · 4TB NVMe
Ryzen 7 9800X3D · RTX 5090 32GB
64GB DDR5 · 4TB NVMe
Core i9 14900KF · RTX 5080 16GB
32GB DDR5 · 2TB NVMe
Ryzen 7 9800X3D · RTX 5080 16GB
32GB DDR5 · 2TB NVMe
Need a whole team on one machine? Up to four RTX PRO Blackwell cards and 384 GB of AI memory, or a GB300 deskside system.
See enterprise systemsFree tools
Estimate. Based on 50 working weeks and 2,080 hours per full-time person. Your numbers will be different.
Estimate. 50 working weeks, 2,080 hours per full-time person.
Estimate, not a quote. About 40,000 tokens in and 8,000 out per task at cloud list prices, 22 working days a month. Caching and volume deals can lower a cloud bill, and at low volume the cloud is cheaper.
Estimate, not a quote. ~40K tokens in, 8K out per task at cloud list prices, 22 working days.
Your customers' data
When someone pastes a customer list, a contract or payroll into a cloud AI, it's on somebody else's servers, under their rules. On an XAI machine the model runs in your building. You hold the keys, nobody outside sees it, and it keeps working when the internet goes down. For law firms, accounting practices, medical offices and government work, that's the whole ballgame.
On an XAI machine the model runs in your building. You hold the keys, nobody outside sees it, and it works when the internet goes down.
Writes ads, emails, posts and product descriptions. Shows what is working and where money is wasted.
Ads, emails, posts and what is working.
Writes quotes and follow-ups, preps reps before calls and flags deals going cold.
Quotes, follow-ups, deals going cold.
Codes expenses, catches duplicate and wrong charges and has reports done before you ask.
Codes expenses, catches wrong charges.
Compares vendors and prices, tracks every subscription and contract and finds what to cut.
Vendors, prices, contracts to cut.
Tracks shipments, catches carrier overcharges and spots problems before the customer calls.
Shipments and carrier overcharges.
Watches inventory, flags what is low or sitting too long and catches order mistakes before they ship.
Low stock, slow stock, order mistakes.
Checks the websites, pricing and compliance every morning and catches the expensive mistakes.
Pricing and compliance checked daily.
Shows where the money goes, where you are over or under staffed and what to change. Every day.
Where the money goes, every day.
AI workstations · NVIDIA Grace Blackwell
How big is the model? A desk-size system for building and testing, or a deskside supercomputer for the models that need one pool of memory. Both run on your network, in your building, with nothing leaving it.
Desk size · NVIDIA DGX Spark · GB10
Build, tune and test models beside your monitor. For one developer or a small team before they scale up.
Deskside · GB300 Grace Blackwell Ultra
One shared AI system for the whole team. It holds models too big to split across cards in one pool of memory, under a desk, with no rack.
A model's weights have to sit in memory with room left for context. Precision decides how much room they take. Most teams serve at 4-bit today.
| Model size | Ascent GX10 · 128 GB | ET900N G3 · 748 GB | ||||
|---|---|---|---|---|---|---|
| 4-bit | 8-bit | 16-bit | 4-bit | 8-bit | 16-bit | |
| 7–8B | Fits | Fits | Fits | Fits | Fits | Fits |
| 13–14B | Fits | Fits | Fits | Fits | Fits | Fits |
| 32–34B | Fits | Fits | Fits | Fits | Fits | Fits |
| 70B | Fits | Fits | Too big | Fits | Fits | Fits |
| 110B+ | Fits | Too big | Too big | Fits | Fits | Fits |
Approximate weight sizes for common open models, leaving about 15% of memory for context. Use it as a starting point. We confirm sizing with you before we quote.
| Specification | ASUS Ascent GX10 | ASUS ExpertCenter Pro ET900N G3 |
|---|---|---|
| Superchip | NVIDIA GB10 Grace Blackwell (NVIDIA DGX Spark) | NVIDIA GB300 Grace Blackwell Ultra |
| CPU | 20-core Arm (10 Cortex-X925 + 10 Cortex-A725) | 72-core Arm Neoverse V2 (Grace) |
| Memory | 128 GB unified LPDDR5x · 273 GB/s | 748 GB coherent: up to 252 GB HBM3e at 7.1 TB/s + up to 496 GB LPDDR5X |
| Compute | Blackwell GPU · up to 1 petaFLOP at FP4 | NVIDIA Blackwell Ultra GPU + Grace 72-core CPU · 1 PCIe Gen5 x16 slot for an optional RTX PRO 6000 Blackwell Max-Q |
| Storage | 1 TB PCIe 4.0 or 4 TB PCIe 5.0 NVMe | 2 × 2 TB NVMe (RAID 1) · 4 × M.2 Gen5 slots |
| Networking | ConnectX-7 SmartNIC (hardware TLS, IPsec, MACsec) · 10 GbE · Wi-Fi 7 · BT 5.4 · links two GX10s | 2 × ConnectX-8 SuperNIC 400G · 10 GbE |
| Management | No BMC | BMC with IPMI 2.0 and Redfish · TPM 2.0 |
| Software | NVIDIA DGX OS (Ubuntu) · CUDA, PyTorch, TensorFlow, Jupyter, TensorRT, NIM preloaded | Ubuntu with NVIDIA AI developer tools |
| Warranty | 1 year ASUS limited (3-year extension from ASUS) | 3 years ASUS manufacturer |
| Lead time | Confirmed on your quote | 2–3 weeks · NVIDIA end-user paperwork required before it ships |
| Size · power | 150 × 150 × 51 mm · 1.48 kg · USB-C power (180 W) | 584 × 232 × 565 mm · 27 kg · 1,600 W Titanium · needs a 20A outlet (NEMA 5-20) |
Vendor specifications and model-size ratings (ASUS, NVIDIA), not our measurements. Warranty terms per ASUS US.
Same NVIDIA silicon anyone can sell you. What changes is what happens before it ships and after it lands.
Set up and burned in by our specialists, and we keep the image so a rebuild comes back running your setup.
Tell us the model, the precision and how many people use it. A specialist matches it to the right system before you buy.
The people who set these up answer the phone, on top of ASUS’s warranty: 1 year on the GX10, 3 years on the ET900.
Purchase orders, Net 30 for approved accounts, tax-exempt certificates. Quotes hold 30 days.
Yes, at 4-bit or 8-bit precision. At full 16-bit a 70B model needs more than its 128 GB of memory, so it moves to the ET900N G3. Linking two GX10s over ConnectX-7 takes you to 405B-class models.
The GX10 is a desk-size system with 128 GB of unified memory for one developer or a small team building and testing models. The ET900N G3 is a deskside GB300 system with 748 GB of coherent memory, shared by a whole team, for models too big to split across cards.
It is ASUS's system on the NVIDIA DGX Spark platform: the GB10 Grace Blackwell superchip, running NVIDIA DGX OS with CUDA, PyTorch, TensorFlow, Jupyter, TensorRT and NIM preloaded.
No. The GX10 is 150 x 150 mm and runs on a 180 W USB-C supply. The ET900N G3 sits under a desk; it has a 1,600 W supply and needs a 20A outlet (NEMA 5-20).
No. Both systems run on your own network, in your building. Models and data stay on hardware you own.
Both are quoted. Tell us the model, the precision and how many people use it, and a specialist sizes the system first. GX10 lead time is confirmed on your quote; the ET900N G3 runs 2 to 3 weeks plus NVIDIA end-user paperwork. We take purchase orders, Net 30 for approved accounts and tax-exempt certificates.
ASUS covers the GX10 for 1 year (a 3-year extension is available from ASUS) and the ET900N G3 for 3 years. On top of that, the people who set your system up answer the phone, in the US.
Tell us the model, the precision and how many people use it at once. We'll tell you which system fits, or that neither does.
Designed & supported in Utah
Your cart is empty
Shop builds
