AION On-Premise PGX.Open the machine.
The book-sized AI computer that runs AION On-Premise. Scroll: we turn it around, lift the lid and go all the way down to the chip.
On the back, only what you need.
Under the lid, a single chip.
No add-in graphics cards, no separate RAM banks: CPU, GPU and memory sit around one superchip, the GB10.
NVIDIA GB10 Grace Blackwell.
150 × 150 × 50.5 mm · 1.2 kg · 240 W
One memory, all of it for the model.
In a traditional PC the GPU only works with its own VRAM, and data comes from RAM over the PCIe bus. In the GB10, CPU and GPU share the same 128 GB.
Two separate memories: the model has to fit in the GPU's VRAM, and every trip from RAM is a copy over PCIe.
- Unified memory
- 128 GB LPDDR5x
- Interface
- 256-bit
- Bandwidth
- 273 GB/s
- Models on one unit
- up to 200B parameters
Model capacity as stated by NVIDIA; two linked units reach 405 billion parameters.
A 125-billion-parameter model in a third of the space.
AION models are quantized: every weight goes from 16 to 4 bits. The model takes about 30% of its original space and leaves memory free for long contexts and several models at once.
Original weights · 16 bits per parameter
FP16 · 100%AION · ~30%
- Compression
- 69–70%
- Quality loss
- 1–3%
- Quality retained
- 97–99%
Example model with 125 billion parameters
Each square is 1 GB of memory.
FP16 · 16-bit
250GB
125 billion × 2 bytes (16-bit)
AION · 4-bit
≈ 75GB
125 billion × ~0.6 bytes (4-bit + scales)
Does not fit in memory. It needs 122 GB more than is available: the original model cannot run on the machine.
Over 30 open-source models, ready to launch.
AION Hub is the catalog that lives on your machine: it downloads, starts and stops, on the GPU, the models AION has optimized. We start from the leading open-source providers and prepare their models for a DGX Spark.
- No usage limits
- Query the models as much as you need: usage is not capped.
- Always up to date
- With the AION On-Premise license, the catalog models are always kept up to date.
How a catalog model is made
Open-source base
We start from the leading providers' models: open and verifiable.
- Qwen
- GLM
- Kimi
- Mistral
AION quantization
Weights drop to 4 bits, so large models fit in 128 GB.
Fine-tuning
Where it helps, we adapt them to your tasks and domain.
Optimized engines
Inference runs on engines AION has tuned for the GB10.
AION Hub
Demo example
- Stopped
Language model
Reasons and answers
- Stopped
Embeddings
Searches by meaning
- Stopped
Reranker
Ranks the results
- Stopped
OCR
Reads scans and images
OpenAI-compatible gateway
No service running
Illustrative example: the space used depends on the models you choose. Models are updated over VPN or, if the machine is air-gapped, with a USB drive.
Over 120 tokens per second. Locally.
The answer is generated on the machine, in your offices. No calls to outside services, no data leaving.
Indicative speed with quantized AION models on a single unit; varies with the model and context length.
Analyze the supply contract and identify the risk clauses.
I analyzed the supply contract (42 pages). Risk clauses found: 1. Art. 7.2 — Late-delivery penalties with no cap: potentially unlimited exposure. 2. Art. 11 — Automatic 36-month renewal, with only 30 days to cancel. 3. Art. 14.4 — Foreign jurisdiction: high legal costs in case of dispute. 4. Art. 18 — Unlisted subcontractors process personal data: GDPR review required. Suggestion: negotiate a penalty cap of 10% of annual value and cut the renewal to 12 months. Processing completed locally: no data left the company perimeter.
- Tokens
- 0
- Speed
- —
- Time
- —
- Data out
- 0 byte
From one to four units.
When you need more users or larger models, you add units linked at high speed. Memory adds up.
Total memory
128 GB
A single machine, on a desk or in a rack: the starting point for a department.
- Link
- None: single unit
- Sizing
- Level 1 · about 10 active users, up to 8 concurrent requests
- Models
- Up to 200 billion parameters (NVIDIA figure)
User counts are sizing estimates: the real number depends on models, contexts and usage.
Specifications.
Data for the reference unit, Lenovo ThinkStation PGX.
Sources: NVIDIA and Lenovo PSREF. Available configurations may vary.
NVIDIA GB10 superchip
- CPU
- Grace, 20 Arm cores (10 × Cortex-X925, 10 × Cortex-A725)
- GPU
- Blackwell, 6,144 CUDA cores, 5th-gen Tensor Cores, 4th-gen RT Cores
- AI performance
- Up to 1 PFLOP FP4 (with sparsity)
- Interconnect
- NVLink-C2C, 900 GB/s bidirectional
Memory and storage
- Memory
- 128 GB unified LPDDR5x, 256-bit, 273 GB/s
- Storage
- 1 TB or 4 TB M.2 NVMe SSD, self-encrypting
Connectivity
- Cluster
- 2 × QSFP, NVIDIA ConnectX-7 at 200 Gb/s
- Network
- 10 GbE Ethernet (RJ-45)
- USB
- 4 × USB-C (3 at 20 Gb/s, 1 for power)
- Video
- HDMI 2.1a
- Wireless
- Wi-Fi 7, Bluetooth 5.3
Unit
- Dimensions
- 150 × 150 × 50.5 mm
- Weight
- 1.2 kg
- Power
- 240 W, USB-C PD 3.1
Software
- Operating system
- NVIDIA DGX OS
- Stack
- NVIDIA AI Software Stack, CUDA 13
- With AION
- AION On-Premise and quantized AION models, installed by us
Let's talk about your machine.
We show you AION on DGX Spark with a real use case and size together how many units you need.
- [email protected]
- Office
- Corso Vercelli 89, 28100 Novara