SABER B300 · Large server

SABER B300.

Eight NVIDIA B300 GPUs and 2.3 TB of HBM3e memory: the power of an AI data center, all to yourself.

To run the largest open models at full precision, several at once, and to train them on your data without it leaving your infrastructure. Installed in a server room or a data center, and maintained by the AIR.TEK team.

With our partners GIGABYTE and DataForge · Member of the NVIDIA Inception program

Technical sheet

2.3 TBof HBM3e GPU memory, 8 × 288 GB
1.8 TB/sbetween GPUs, NVLink and NVLink Switch
128AMD EPYC cores, up to 5.0 GHz
2.3 TBof DDR5-6400 ECC system memory

The configuration.

Chassis
GIGABYTE G894-ZD3-AAX7NVIDIA HGX B300 platform
GPUs
8 × NVIDIA B300Blackwell Ultra, NVLink and NVLink Switch
Processors
2 × AMD EPYC 9575F2 × 64 cores, 3.3 GHz, up to 5.0 GHz
GPU memory
2.3 TB HBM3e8 × 288 GB
Memory
2.3 TB DDR5-6400 ECC RDIMM24 × 96 GB, 12 channels per processor
Storage
61 TB NVMe Gen5 for models4 × 15.36 TB, plus 1 TB NVMe for the system
GPU network
8 × 800 Gb/s OSFPConnectX-8 SuperNIC, InfiniBand XDR or 2 × 400 GbE
Network
2 × 10 GbENVIDIA BlueField-3 DPU compatible
Power
12 × 3,000 W 80 PLUS Titaniumup to about 14 kW: server room or data center
Software
AIR.TEK platform, Open WebUIOpenAI-compatible API, online monitoring

Models

Large open models, on your premises.

2.3 TB of GPU memory linked by NVLink: the largest open models run at full precision on a single server, and several can run together.

ModelSizePrecisionGPU memoryOn this server
GLM-5.3743B (MoE, 39B active)Native FP8≈ 743 GBYes, 1M-token context
GLM-5.3, full precision743B (MoE, 39B active)BF16≈ 1.5 TBYes, on a single server
DeepSeek V4.1 Flash552B (MoE, 16B active)Native FP8 / FP4≈ 510 GBYes, context up to 1M tokens
GLM-5.3 and DeepSeek V4.1 FlashBoth at the same timeFP8 and native≈ 1.25 TBYes, ≈ 1 TB free for context

Training too: the GPUs, linked by NVLink, also make it possible to fine-tune large models on your data, without it leaving your infrastructure. Sizes are indicative, based on the published versions; the configuration is validated with you according to your use cases.

Price

Bought or leased.

Purchase€490,420 excl. VAT
36-month lease€15,792 excl. VAT / month
48-month lease€12,359 excl. VAT / month
60-month lease€10,250 excl. VAT / month
Maintenance and service€4,500 to €6,900 excl. VAT / month

Delivery lead time: about 6 months.

Quote valid the same day: component prices and the dollar exchange rate fluctuate, so prices become firm only when the order is paid.

Delivery and deployment not included, quoted separately.

Prices excl. VAT, plus 20% VAT. Leasing subject to the financing company’s approval.

Prices as of October 13, 2026

Maintenance and service

Two plans, Essentiel and Premium.

Every SABER comes with a maintenance and service contract: online monitoring, updates, warranty management, on-site intervention and, with Premium, new open models tested and deployed within 30 days.

Essentiel

€4,500

excl. VAT / month

Premium

€6,900

excl. VAT / month

What each plan includes →

Guarantees

SABER users only

The power of an AI data center, on your premises.

Your data doesn’t leave

The models run on your server, on your network: your documents and your queries stay in the company.

A fixed cost, unlimited use

Within your server’s capacity. No bill per GPU hour or per token.

Access no one can cut off

Your AI depends on no remote service that could be taken away from you.

Questions

Your questions, answered.

Why a B300 rather than GPUs in the cloud?

To keep your data and your models on your premises, at a fixed cost, with no bill per GPU hour. Eight B300s rented continuously in the cloud cost several tens of thousands of euros per month; the SABER B300 is bought or leased, and its capacity stays yours.

Where should it be installed?

In a server room or a data center: the server draws up to about 14 kW and requires power and cooling sized accordingly. We help you choose the site.

Can it be used to train models?

Yes. The eight GPUs linked by NVLink make it possible to fine-tune large models on your data, without it leaving your infrastructure.

What is the lead time?

About 6 months for the B300 GPUs. The quote is valid the same day and the price is firm only when the order is paid.

What maintenance comes with this server?

Every SABER B300 comes with a maintenance and service contract, Essentiel or Premium, with the GPU network included in Premium, detailed on our Maintenance and service page.

Let’s talk about your server.