Self-Hosted LLMs — buy or rent?
What does operating your own LLM infrastructure really cost — as a purchase (DGX Spark, DGX Station GB300, GB300 NVL72) or as bare-metal rental with full control during the rental period?
The key figures
All values EUR net (excl. VAT), calculation basis November 2026. Every figure comes from the generated calculation (calc/tco.md), every performance figure from public receipts (vLLM, SGLang, TensorFold) — nothing is estimated where a receipt exists.
Method: receipts first
Three use-case profiles with shared requirements — the scenarios differ only in peak factor and model class.
A1 — Document research / RAG (NDA)
Peak factor 50%. Half the users run a concurrent research session at peak. Flash class (GLM-5.3-Flash, DeepSeek-V4.1-Flash).
A2 — Software & firmware development
Peak factor 80%. Agentic, long generations, large contexts — the cost driver. Frontier models (Kimi K3) only on GB300/NVL72 classes.
A3 — Internal IT automation
Peak factor 90%. Agents and workflows run continuously at high concurrency — flash class.
Mixed-load power model: Weff = idle + (load − idle) × duty cycle. Cooling +15%. Electricity is included in the rent for every rental option — no separate power line there.
Buying: the four candidates
Prices November 2026 (re-verified 2026-11-22; DGX Spark +9.6% since October, RTX 6000 −1.7%). Device classes come as pairs (N+1 inherent).
| Configuration | CAPEX net | Month idle | Month full load |
|---|---|---|---|
| DGX Spark GB10 2× cluster (256 GB) | € 9,575 | € 301–316 | € 332–373 |
| DGX Spark GB10 4× cluster (512 GB) | € 19,150 | € 602–632 | € 665–745 |
| 2× DGX Station GB300 + 2× RTX 6000 (pair) | € 210,418 | € 6,835–6,926 | € 7,393–7,931 |
| 2× GB300 NVL72 (purchase, redundant) | ~€ 8.6M | ~€ 330,129–374,455 | |
Which model runs where?
| Class | Largest models | Does not run |
|---|---|---|
| Spark 2× (256 GB) | GLM-5.3-Flash, Qwen3.8-Flash-Next, DS-V4.1-Flash only in low quants (fewer users) | DS-V4.1-Flash Q3_K_M (347 GB) and everything above |
| Spark 4× (512 GB) | + DS-V4.1-Flash Q3_K_M/Q4_K_M, GLM-5.3 Int4-Int8Mix | DS-V4.1-Flash FP8 (510 GB), GLM-5.3 FP8 (756 GB), Kimi K3 |
| GB300 pair (1,496 GB) | All flash quants (FP8/NVFP4), GLM-5.3, DS-V4-Pro, Kimi K3 int3/1-bit | GLM-5.3 BF16 (1,507 GB), Kimi K3 MXFP4 (1,561 GB) |
| 2× NVL72 (~20 TB) | Everything, incl. Kimi K3 MXFP4 and GLM-5.3 BF16 | — |
Classification: DGX Spark is the entry tier (pilot, 16–64 users, flash workloads) with a planned exit at ~128 users. The team platform is the GB300 pair. Frontier-A2 scales only on 2× NVL72 from ~90 users.
Renting: bare metal with full control
Only dedicated bare-metal servers with full root access count — no hypervisors, no shared GPUs. 19 qualifying providers reviewed (full research).
| Rental node | VRAM | Month net | One-time setup | Location / provider |
|---|---|---|---|---|
| Hetzner GEX430 (4× RTX PRO 6000 Max-Q) | 384 GB | € 3,599 | € 3,599 | Helsinki, Hetzner |
| 8× RTX PRO 6000 Blackwell | 768 GB | ~€ 9,629 | n.v. | Amsterdam, DataPacket |
| 8× HGX B300 (Blackwell Ultra) | 2,304 GB | ~€ 35,040 | n.v. | Gcore (no receipts — price reference) |
Billing reality: Hetzner charges the one-time setup fee always — hourly billing included. Cancellation is to month end: a rental effectively runs 1–2 months minimum (setup + ~2 months ≈ € 10,797 for the GEX430). The hourly rate (€5.77/h) is almost theoretical for short projects.
NDA/IP data leaves the premises: Rental runs in the provider's data center (e.g. Helsinki). The same concern that excluded colocation applies here — rental is documented as a parallel scenario, not a replacement for the on-prem decision.
N+1 applies to rental too: every additional node is an additional rental contract (with its own setup fee). At small stages, redundancy roughly doubles the cost per user.
Buying versus renting: the break-even
Cost per user/month (mixed load, 36-month basis), A1 document research. The GB300 pair gets cheaper with occupancy; rental scales linearly.
| A1 users | GB300 pair (buy) | Hetzner GEX430 (rent) | Spark 2× (buy) | 8× RTX PRO 6000 (rent) |
|---|---|---|---|---|
| 16 | € 454 | € 462 | € 41 | € 1,204 |
| 64 | € 114 | € 116 | € 41 | € 301 |
| 128 | € 57 | € 116 | € 41 | € 150 |
| 256 | € 28 | € 116 | € 41 | € 150 |
At 64 A1 users, the purchased pair (€113.61) and the GEX430 rental (€115.59) are practically tied — the break-even. At 128 users the rental costs twice as much per user, at 256 four times. Note: the Spark cluster wins the per-user math at small stages but only carries flash classes and scales by device count — 128 A1 users need 32 GB10 devices.
Sources
- Calculation model: selfhosted-lm-tco —
calc/parameter.json(source of truth) →calc/tco.md/tco.csv(generated, as of 2026-11-22). - Price re-verification November 2026: research/preis-update-2026-11-22.md (Geizhals, buyzero/pi3g, Apple Store DE, Iodis).
- Rental research (19 providers): research/miete-anbieter.md — Hetzner GEX430 product page, DataPacket, Gcore, OVHcloud and others.
- Performance receipts: research/hardware/rtx-pro-6000.md — vLLM/SGLang/TensorFold measurements on 4×/8× RTX PRO 6000 (GLM-5.3-Flash, DeepSeek-V4.1-Flash), localmaxxing matrices, SGLang 4× build with Engram on NVMe.
- GB300 receipts: RonanLabsAI/dgx-station-frontier (2026-10-10) — in
research/hardware/dgx-station.md.
All monetary values EUR net (excl. VAT). Rental prices net as published by the providers; USD quotes converted at the ECB reference rate 0.90992 (2026-11-20). This page is a private, non-commercial presentation; no guarantee is given for completeness or currency.