Europe's Sovereign GPU Price Table: The Two Break-Evens That Decide Rent-vs-Own Inference

An H100 in Germany costs $2.80 an hour. The same class of GPU in France costs $3.20. Neither vendor will sell you an H200 or a B200 at any price, neither offers spot, and neither can be compelled by a US court order to hand over your data. That is the entire European sovereign-GPU market in one paragraph — and if you run inference for EU users, that paragraph is also your shortlist.

Spheron's September 2026 roundup of GPU cloud providers in Europe is the first comparison I have seen that treats EU data residency and CLOUD Act exposure as first-class columns next to per-GPU hourly rates. It is worth reading in full. But its real value compresses into a small table and one arithmetic question: at what utilization does renting that hourly H100 stop making sense next to owning a flat-rate dedicated GPU box in the same jurisdiction? The answer is calculable, it lands earlier than most teams expect, and it splits cleanly into two break-evens — one for each vendor.

The table, up front​

All rates below are per-GPU, on-demand hourly, from published September 2026 numbers. Hyperscaler rows are EU-region contrast, not recommendations.

Three things to read off before the detail. First, the sovereign discount is real: Hetzner's H100 rate undercuts AWS's EU-region H100 by roughly 60%. The broader market tells the same story — median on-demand H100 pricing runs about $4.17/hr on dedicated GPU clouds versus about $7.89 on hyperscalers, an ~89% premium for the same silicon.

Second, the sovereign catalog is shallow: H100 or nothing, no spot market, no newer generation on self-service. If your workload needs 141 GB of HBM3e on an H200 or 192 GB on a B200, the EU-sovereign self-service shelf is empty — with one exception, covered below.

Third, the last column is binary and jurisdictional, not technical: it follows the vendor's incorporation, not the datacenter's coordinates. An H100 in Frankfurt billed by a US hyperscaler and an H100 in the same city billed by Hetzner answer to different courts.

The break-even math, in one currency​

Currencies normalized to USD throughout so the comparison is checkable: Hetzner's dedicated GEX131 GPU server lists at €889/month, which Hetzner itself quotes as $989/month. A month is 730 hours. Stated once, used everywhere below.

The own-side candidate is the GEX131: a dedicated box with an NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB) at $989/month flat, no setup fee, IPv4 included — about $1.35/hr if you run it wall-to-wall. The rent-side candidates are the two hourly rows above. Full-month totals at 730 hours:

At full utilization, owning saves ~$1,055/month (52%) against Hetzner's own hourly rate and ~$1,347/month (58%) against OVHcloud. The break-evens — the utilization where flat beats metered — fall out by division:

• vs Hetzner hourly: $989 ÷ $2.80 ≈ 353 hours/month (~48% utilization). Past roughly 11–12 hours a day, own the box.
• vs OVHcloud hourly: $989 ÷ $3.20 ≈ 309 hours/month (~42% utilization). Past roughly 10 hours a day, own the box.

Two concrete workloads to pin the sensitivity. A 24/7 inference endpoint serving EU users (730 hrs/month) costs $2,044–$2,336 metered versus $989 owned — owning wins by four figures monthly, and the box keeps getting cheaper per token the busier it gets. A nightly batch job running 4 hours a night (120 hrs/month) costs ~$336 on Hetzner hourly versus $989 flat — renting wins by 3×, and owning would mean paying for 610 idle hours. The crossover band is narrow, roughly 42–48% utilization, which is exactly why this deserves arithmetic instead of vibes: most teams can forecast which side of 10-hours-a-day their inference fleet lives on.

Honest caveats, because the comparison is not apples-to-apples silicon. The GEX131 carries a Blackwell workstation-class GPU with 96 GB, not an H100 SXM with 80 GB — for inference throughput per dollar it is competitive, for Hopper-specific features or NVLink-scale training it is not the same part. Owning also means you operate the box: driver upgrades, RMA waits, and utilization risk are yours, while the hourly rate bundles someone else's spare capacity. And neither Hetzner nor OVHcloud self-service offers spot, so there is no third "cheap but interruptible" row to undercut both — the cheapest burstable H100 in the table is the on-demand hourly rate itself.

The missing rows: H200, B200, and the one exception​

The most consequential cell in the table is "Not available." Neither Hetzner Cloud nor OVHcloud self-service rents H200 or B200 capacity as of September 2026; OVHcloud's H200 footprint (Paris GA, Milan) sits behind its managed AI products, not the public-cloud hourly meter. Market-wide, the generation ladder is steep — H200 on-demand averages ~$4.00/GPU-hr with a 25–30% premium over H100, B200 averages ~$6.11, and hyperscaler B200 rows reach $8.60–$14.24 — but none of that matters to a team that cannot clear compliance on a US-incorporated vendor, because the sovereign shelf stops at Hopper's first rung.

The exception is Scaleway: its EU-first Blackwell posture puts B300-class capacity on a French-incorporated rate card (~€2.99/hr for H100 PCIe alongside it), making it the only sovereign vendor in this comparison with a newer-than-Hopper story. The practical consequence is a two-tier sovereign market — Hetzner and OVHcloud for cheap, legible H100 hours; Scaleway when the workload genuinely needs more than 80 GB of VRAM per GPU and cannot leave EU jurisdiction. Teams sizing models against 70B-plus long-context inference should check that VRAM ceiling before they check prices: if the model does not fit in 80 GB without sharding tricks, the Hetzner-vs-OVHcloud debate is moot and the shortlist has one name on it.

The sovereignty column is not a price​

Why does a column about American surveillance law sit in a GPU price table? Because for EU teams serving EU users, it prices a risk the hourly rate omits. The CLOUD Act compels US-incorporated providers to produce data stored anywhere in the world — including their EU regions.

An H100 in Frankfurt rented from a US hyperscaler and an H100 in the same city rented from Hetzner or OVHcloud are technically adjacent and legally different: only the former answers to a US court order. EU data residency (where the bits sit) and CLOUD Act exposure (who can compel them) are independent axes, and the Spheron roundup is right to score them separately.

GDPR sharpens the point for inference specifically. An EU user querying a model with personal details creates a processing event; routing that query to a US-East cluster needs a transfer mechanism (SCCs, BCRs, or the Data Privacy Framework), while keeping it in-EU needs none. Model weights themselves are generally not personal data, so training location is the softer constraint — inference serving location is the hard one.

The compliance-correct architecture (EU users, EU inference, EU-incorporated vendor) is therefore also, per the table above, the cheapest architecture. That alignment is unusual enough to say plainly: sovereignty here is not a tax. At H100 tier, it is a discount.

What a self-hosted PaaS does with this​

The playbook writes itself once the numbers are normalized. Rent Hetzner or OVHcloud H100 hours for spiky, sub-300-hour-month inference — prototypes, eval runs, bursty batch — where the meter's legibility beats the flat rate. Own dedicated GEX-class boxes the month sustained utilization crosses ~45%, and treat that crossing as a scheduled decision: instrument GPU hours per tenant, alert at 300 hours, convert.

Keep Scaleway in the pocket for the workload that outgrows 80 GB of VRAM rather than re-architecting residency around a US vendor. And never let a US-region default survive contact with EU users — the transfer paperwork costs more in lawyer hours than the hourly delta ever saved.

The deeper lesson is structural. The sovereign market's weakness (shallow catalog, no spot, no H200 on self-service) and its strength (half-price H100s with no CLOUD Act exposure) are the same fact: small EU vendors sell what they can provision simply. A platform that already provisions its own machines — declarative fleet lifecycle on owned hardware, the Cluster-API-shaped bet — slots into that market's strong side naturally: rent the meter while you are small, convert to flat boxes on a formula, and never re-platform to do it.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.

Sources​

• Spheron, "GPU Cloud Providers in Europe 2026: Sovereign AI, B200 and H200 Availability, EU Data Residency" (Sept 2026) — the Hetzner ~$2.80/hr, OVHcloud ~$3.20/hr table, CLOUD Act column, no-H200/B200 and no-spot findings.
• Spheron, "GPU Cloud Pricing Comparison 2026" (July 2026) — hyperscaler EU-region contrast rows (AWS ~$6.88, Azure ~$12.29 per H100/hr).
• Thundercompute, "AI GPU Rental Market Trends, September 2026" — median $4.17/hr dedicated vs $7.89 hyperscaler H100 figures.
• Hetzner press room, GEX131 announcement — €889/$989 per month, no setup fee, RTX PRO 6000 Blackwell Max-Q specs.
• OVHcloud AI portfolio announcements; Saturn Cloud GPU comparison report — H100 PCIe from €2.80/hr public cloud, managed tiers, April 2026 price rises, no InfiniBand on self-service.
• Gart, "Scaleway vs OVH: the definitive comparison" — Scaleway H100 PCIe ~€2.99/hr, B300 EU-first availability.