The models that fit nothing else
The flagship build
252 GB of HBM3e at 7.1 TB/s — four times the memory bandwidth of any card in the other four builds.
Built for
Price on request
Neither ASUS nor NVIDIA publishes a price for this machine, and we are not going to invent one. Every other build on this site shows its parts and their cost because we can source every figure; this one is a sealed system with no component-level bill of materials. We quote it from our supplier, and we show you that quote.
What this one is for
- Models in the 200-billion-parameter class, which no other build here holds
- Partitioning one machine into as many as seven isolated instances
- Work where the model is the constraint rather than the number of people using it
Where it stops
DeepSeek-V3 and R1 still do not fit. At 4-bit they are about 407 GB against 252 GB of HBM3e, so the remainder would spill into LPDDR5X at 396 GB/s — roughly eighteen times slower than the memory beside it. A model that technically loads and answers slowly is not a model you can put in front of staff.
Every part, and what it costs
Specifications from ASUS and NVIDIA product documentation, read 2026-08-22. No costs are shown because none are published.
| Part | Specification |
|---|---|
| Superchip | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip 20 petaFLOPS of FP4 compute. Partitions into up to seven isolated instances. |
| GPU memory | 252 GB HBM3e 7.1 TB/s. For comparison, an RTX PRO 6000 runs at 1.792 TB/s. |
| CPU | Grace 72-core Neoverse V2 ARM, not x86. Software expectations differ — we check yours before quoting. |
| CPU memory | 496 GB LPDDR5X 396 GB/s, coherent with the GPU memory across NVLink-C2C at 900 GB/s. |
| Storage | Up to 4 × M.2 NVMe PCIe 5.0, 8 TB total Configured to the workload. |
| Network | NVIDIA ConnectX-8 SuperNIC Up to 800 Gb/s, two QSFP 400G ports, plus 10 Gb LAN and a 1 Gb BMC port. |
| Chassis | ASUS ExpertCenter Pro ET900N G3 tower 565 × 584 × 232 mm. A tower, not a rack unit. |
Component prices moved sharply through 2026 — memory roughly doubled in the first quarter alone, and GPUs more than doubled over the year. A quote is good for the week it is written, and we re-price honestly rather than carrying an old number forward.
How it is built and configured
- Superchip
- NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
- GPU memory
- 252 GB HBM3e at 7.1 TB/s
- CPU memory
- 496 GB LPDDR5X at 396 GB/s, 748 GB coherent in total
- Interconnect
- NVLink-C2C at 900 GB/s
- Power
- 1,600 W — about 13 A at 120 V, so it fits a dedicated 20 A circuit
- Network
- NVIDIA ConnectX-8 SuperNIC, up to 800 Gb/s, plus 10 Gb LAN
Why the user counts have a context length on them
A model's weights are fixed, but every person talking to it needs their own context, and that is what fills the rest of the card. For a 70B model it costs about 320 KiB per token — so one person at 8K tokens uses 2.5 GiB and the same person reading a long contract at 32K uses 10 GiB. The same machine serves very different numbers of people depending on the work.
That is why every figure here names the context it assumes. A vendor quoting you a user count without one has not done the arithmetic.
See the full VRAM arithmeticThe other three
Tell us what you are trying to run
Describe the workload, how many people need it, and how sensitive the data is. You will get a straight answer about whether owning the hardware makes sense for that situation — including when it does not, and a subscription would serve you better.
Prefer to talk? Call James on 832-338-2926. Please do not send confidential, client-privileged, health, financial-account or credential information through this form.