QDNAAdvisory and architecture for LLM inference and training platforms, on-premises or hybrid

Local LLM and compliance in the public sector

The public sector falls under the GDPR and SecNumCloud framework. Serving a model on the client premises moves the question of data transfer.

By QDNA · Published on · Updated on

Short answer. A model served on the client premises processes data in place. No data leaves the information system, which answers the GDPR and SecNumCloud framework without depending on a transfer to a third party.

Which framework applies?

The sector falls under GDPR and SecNumCloud. The sensitive point is public procurement and sovereignty. Processing carried out on site removes the transfer of data to a provider, and with it the question of hosting location and of the sub-processor.

Which platforms qualify?

Indicative prices, not contractual, excluding integration; “on quotation” when no vendor publishes a price. Sources are at the end of the page. Sizing depends on the model chosen. See the sizing pages for the footprint of each model on each platform.

A production illustration: gironde.qdna.fr, a public-data site designed, deployed and operated by QDNA (case study).

What this page does not say

This page describes an architecture. It is not legal advice. Compliance is assessed on the actual processing, with the organisation’s data protection officer and, where applicable, its supervisory authority.

Official sources

Does a local LLM answer GDPR and SecNumCloud?

A model served on the client premises processes data in place. No data leaves the information system, which answers the GDPR and SecNumCloud framework without depending on a transfer to a third party.

Which platform should I start with?

The choice depends on the target model and the number of users. The sizing pages give the memory footprint of each model on each platform.

Method and sources

Memory footprints are calculated, not measured: a reading on real hardware may differ depending on the engine and the exact weight format. Every figure below carries its source and the date of its last check.

  • DGX Spark, 128 GB (128 GB LPDDR5X, 273 GB/s, 240 W power supply, 140 W GB10 TDP), manufacturer sheet checked on 2 September 2026. Price: €5,850 to €7,511 incl. VAT, indicative and not contractual. Best price €5,850 incl. VAT, €4,875 excl. VAT, listed on idealo.fr on 2 September 2026.
  • Mac Studio Ultra, 512 GB (M5 Ultra, up to 512 GB unified memory, 1.2 TB/s, 480 W continuous; the 512 GB variant is announced for late October 2026), manufacturer sheet checked on 2 September 2026. Price: €6,599 to €20,679 incl. VAT, indicative and not contractual. Apple Store France as reported by MacGeneration on 25 August 2026: €6,599 for the M5 Ultra with 96 GB, €20,679 with 256 GB and 16 TB; the 512 GB variant is not priced yet.
  • DGX Station, 748 GB (252 GB HBM3e (7.1 TB/s) and 496 GB LPDDR5X (396 GB/s), 1,600 W, 20 PFLOPS FP4 with sparsity (15 dense)), manufacturer sheet checked on 2 September 2026. Price: €93,000 to €110,500 excl. VAT, indicative and not contractual. Supermicro, MSI, Exxact, Gigabyte, HP and ASUS configurations listed by the reseller pi3g on 27 August 2026, excluding VAT and shipping.
  • RTX PRO 6000 server, 768 GB (96 GB GDDR7 and 1,597 GB/s per card, 600 W, hence 768 GB for 8 cards), manufacturer sheet checked on 2 September 2026. Price: on quotation. No manufacturer publishes a price for an 8-card server; the card alone is listed at $16,000 on the NVIDIA marketplace as of 1 September 2026 (as reported by thundercompute.com).
  • H200 SXM server, 1,128 GB (141 GB HBM3e and 4.8 TB/s per GPU, 900 GB/s NVLink, 700 W, hence 1,128 GB for 8 GPUs), manufacturer sheet checked on 2 September 2026. Price: about €453k excl. VAT, indicative and not contractual. 8-GPU H200 141 GB server with two Xeon Gold 6538Y+ listed at £387,785 by CTO Servers on 2 September 2026, converted at the ECB rate of £0.857 per €1.
  • B200 SXM, 1,440 GB (HGX B200, 8 Blackwell GPUs, 1.4 TB total memory, 144 PFLOPS FP4 with sparsity (72 dense), 1.8 TB/s NVLink per GPU), manufacturer sheet checked on 2 September 2026. Price: on quotation. NVIDIA publishes no price; Aivres quoted about $340,000 excl. VAT for an HGX B200 server (Arc Compute, updated 30 July 2026, flagged as outdated by the author).
  • B300 SXM, 2,304 GB (HGX B300, 8 Blackwell Ultra GPUs at 288 GB, hence 2,304 GB (the HGX sheet rounds to 2.1 TB), 144 PFLOPS FP4 with sparsity (108 dense), 1.8 TB/s NVLink per GPU), manufacturer sheet checked on 2 September 2026. Price: about €430k to €690k excl. VAT, indicative and not contractual. BIZON X9000 G5 at $496,728 (November 2026 delivery) and Supermicro AS-8126GS-NB3RT at $795,000 from Vipera, both listed on 2 September 2026, converted at the ECB rate of $1.159 per €1.
  • GB300 NVL72, 20,700 GB (72 Blackwell Ultra GPUs and 36 Grace CPUs, 20 TB of GPU memory per NVIDIA (72 × 288 GB = 20,736 GB), 37 TB fast memory, 130 TB/s NVLink, 1,440 PFLOPS FP4 with sparsity (1,080 dense); power draw not published), manufacturer sheet checked on 2 September 2026. Price: on quotation. NVIDIA publishes no price; analyst estimates range from $3.7M to $6.5M per rack (Loop Capital, Tom’s Hardware, as reported by Spheron on 16 August 2026).