Why You Can't Get Kimi or DeepSeek From Any Major EU Cloud

A friend asked me a reasonable question: where do you get AI models inside the EU, and what do they cost? He wanted to run Kimi somewhere that isn’t a US hyperscaler.

Should have been a twenty-minute lookup. The catalogs didn’t cooperate.

Criteria: the model must be callable on a shared, serverless endpoint with EU data residency and a published price. GPU rental where you deploy the weights yourself doesn’t count, because at that point every provider “has” every model.

---

The Catalogs Don’t Line Up

Here is every text model the four big incumbents serve as of early August 2026, pulled from their own catalogs. Embedding, rerank, speech and image models are left out to keep the comparison honest.

Model	OVHcloud	Scaleway	IONOS	STACKIT
Qwen3.5-397B-A17B
Qwen3-VL-235B-A22B
Qwen3-235B-A22B (Instruct + Thinking)
Qwen3.5-122B-A10B
Qwen2.5-VL-72B
Qwen3-Coder-Next 80B
Qwen3.6-27B / -35B-A3B
Qwen3.5-9B / -35B-A3B
Qwen3-Coder-30B-A3B
Qwen2.5-Coder-32B
Qwen3Guard-Gen 8B / 0.6B
GLM-5.2
MiniMax-M2.5		dedicated only
gpt-oss-120b
gpt-oss-20b
Llama 3.3 70B
Llama 3.1 405B FP8
Llama 3.1 70B / 8B
Llama 3 70B / 8B
Llama-3.1-Nemotron-70B
DeepSeek-R1-Distill-Llama 70B / 8B
Gemma 3 27B
Gemma 4 31B / 26B-A4B
Mistral Large 3 675B
Mistral Medium 3.5 128B
Mistral Small 3.1 / 3.2 / 24B
Mistral Nemo 12B
Mistral 7B v0.3 / Mixtral 8x7B
Devstral 2 123B / Devstral Small
Magistral Small
Pixtral 12B
Molmo 72B
Holo2 30B-A3B
LightOnOCR-2-1B

The DeepSeek entries on Scaleway are R1 distills into Llama, not DeepSeek’s own architecture. GLM-5.2 is the only genuine frontier Chinese model any incumbent serves over a shared endpoint, at €1.80 in / €5.50 out. MiniMax-M2.5 sits in Scaleway’s catalog but only as a dedicated deployment on H100-SXM-4 or -8, not on the serverless API.

Meanwhile Nebius Token Factory serves Kimi K3, DeepSeek V4 Pro, GLM-5.2, MiniMax-M3 and Nemotron 3 Ultra. Artificial Analysis tracks 33 models on their endpoint. And Lyceum, a Berlin outfit most people haven’t heard of, lists Kimi K2.7 Code, GLM-5.2, DeepSeek V4 Pro and Flash, MiniMax M3 and Qwen, all EU-hosted with a DPA.

So it’s not an EU-versus-US thing. Both groups are European. They just serve different halves of the model space.

It’s Not a Ban on Chinese Models

That was my first theory and it’s wrong.

Every single one of those incumbents hosts Qwen, which is Alibaba’s. OVHcloud and Scaleway both run Qwen3.5-397B-A17B as their flagship, at the same price. STACKIT’s top model is Qwen3-VL-235B-A22B, also Alibaba’s. If there were a policy against Chinese weights, none of that would be on European price lists.

The licenses don’t explain it either. Qwen3.5 is Apache 2.0, but DeepSeek and GLM ship under MIT, which is just as permissive. Kimi’s modified-MIT has an attribution clause above certain scale thresholds, but that’s one model out of four.

It’s HBM per Node

Look at the parameter counts instead.

Qwen3.5-397B-A17B is 397 billion parameters total. At FP8 that’s roughly 400 GB of weights. An 8×H100-80G node gives you 640 GB of HBM. The model fits, with room left over for KV cache and a decent batch size.

Kimi K3 is 2.8 trillion parameters. Even at FP4 you’re looking at ~1.4 TB before you’ve cached a single token. That does not fit on an H100 node. You need Blackwell-class memory (8×B300 at 288 GB each gets you 2.3 TB), or you split across nodes and eat an InfiniBand all-reduce on every layer, which wrecks your latency.

Now check what the incumbents actually run inference on:

Scaleway sells B300-SXM GPU instances. But their Generative APIs dedicated deployment flavors top out at H100-SXM-8-80G. The Blackwell silicon exists in the building; it just isn’t behind the inference product. OVHcloud’s GPU line ends at H200 entirely.

That’s the whole explanation. European incumbents serve every open-weight model that fits on one node they already own. Everything above that line lives at the AI-native platforms, because parking 1.4 TB of weights per replica only pencils out if inference is your entire business and you can keep utilization high.

A secondary regulatory factor sits on top of that. IONOS documents in their own product reference that they act as Distributor for unmodified open-source models but as AI Provider, with additional transparency obligations, for quantized ones like Llama 3.1 405B-FP8. Since you cannot serve a trillion-parameter MoE economically without quantizing it, hosting the big models drags you into the heavier AI Act classification. That’s a real cost, but not the decisive one: these providers already serve plenty of FP8 and FP4 weights.

What It Costs

Prices as of August 2026. Currencies differ: the incumbents quote EUR, the AI-native platforms quote USD. Don’t compare these columns without converting.

gpt-oss-120b, the one model almost everyone carries, so a usable anchor:

Provider	Input	Output
OVHcloud	€0.08	€0.40
Scaleway	€0.15	€0.60
IONOS	€0.15	€0.65
STACKIT (LLM-Plus)	€0.45	€0.65

A factor of five between cheapest and dearest for identical weights. STACKIT is the outlier because it prices by plan tier, not by model, so a 120B model costs the same there as a 27B one.

Qwen3.5-397B-A17B, the flagship. Only two providers carry it:

Provider	Input	Output
Scaleway	€0.60	€3.60
OVHcloud	€0.60	€3.60

Identical to the cent on both sides. I have no explanation for it.

Each provider’s largest model, which is the number that matters if you’re picking one:

Provider	Top model	Input	Output
Scaleway	GLM-5.2	€1.80	€5.50
IONOS	Llama 3.1 405B FP8	€1.75	€1.75
STACKIT	Qwen3-VL-235B (Premium)	€1.50	€1.75
OVHcloud	Qwen3.5-397B-A17B	€0.60	€3.60

The models you actually came here for, at Lyceum (USD, per 1M):

Model	Input	Cached	Output
Kimi K2.7 Code	$1.25	$0.31	$4.50
GLM-5.2	$1.50	$0.38	$4.50
DeepSeek V4 Pro	$1.75	$0.44	$3.50
DeepSeek V4 Flash	$0.15	$0.04	$0.30
MiniMax M3	$0.40	$0.10	$2.00

STACKIT prices differently from the rest, which is why its numbers above are tier rates rather than model rates. Their Leistungsschein sorts every model into a service plan by size and bills against the general price list. The product page carries the plan rates and the model docs carry the plan each model sits in, so you have to join two documents to get a price:

Plan	Input	Output	Models
LLM-Standard	€0.15	€0.25	gpt-oss-20b
LLM-Plus	€0.45	€0.65	Llama 3.3 70B, gpt-oss-120b, Gemma 3 27B, Gemma 4 31B, Qwen3.6 27B
LLM-Premium	€1.50	€1.75	Qwen3-VL-235B-A22B
Embedding-Standard	€0.02	n/a	E5 Mistral 7B
Embedding-Plus	€0.08	n/a	Qwen3-VL-Embedding-8B

Note the Leistungsschein still calls the tiers Base, Plus and Premium while the docs and the billing SKUs say Standard. The PDF is dated September 2025 and hasn’t caught up.

Then Check the Ownership Chain

Everything I wrote in the GDPR and cloud hosting post applies here, and the answers are less tidy than they look.

Nebius is the one to look at carefully. The entity is Nebius Group N.V., a Dutch company, so there’s no US parent and no CLOUD Act exposure in the AWS European Sovereign Cloud sense. But it’s Nasdaq-listed, the current business was launched in July 2024 after the divestment of Yandex’s Russia-based operations, and the data center footprint spans Finland, France, the UK, Iceland, New Jersey and Missouri. EU residency is available and documented. It is not the same proposition as OVHcloud, and if your compliance team is going to ask questions about anything on this page, it’ll be this.

Lyceum is a German company with EU residency, zero-retention mode and a DPA/AVV on request. Small, which cuts both ways: responsive support, but you’re betting on a startup for production inference.

And watch for providers where residency is per-model rather than per-account. llmbase.ai, for example, states plainly that the processing location depends on which model you pick. EU legal entity, EU-operated API boundary, compute wherever capacity happens to be. That’s a meaningfully weaker guarantee and it’s easy to miss.

One more operational note: Scaleway deprecated five models on 1 July 2026 and routes requests for retired models to something with “similar pricing and capabilities”. For a production workload that matters more than ten cents per million tokens, and nobody puts it in a comparison table.

My Take

If your models fit on one node, meaning anything up to roughly the 400B MoE class, the incumbents are the better buy. Real DPAs, real certifications, real support contracts, and OVHcloud is the cheapest of the four. Start there.

If you need Kimi, DeepSeek V4 Pro or GLM at the top end, the incumbents don’t have it and won’t until their inference fleets get Blackwell. Nebius or Lyceum, with the ownership caveats above.

If you’re on Mistral, note that you’re on a single vendor’s models by design. La Plateforme runs in the EU and Large 3 is priced aggressively, but there’s no DeepSeek or Kimi path. That’s a portfolio decision, not a limitation you can work around.

And if you’re German public sector or anything BSI C5 adjacent, IONOS and STACKIT are the pragmatic answer even though the catalogs are thin. IONOS pins everything to Germany with no cross-region replication, which is exactly what you want or exactly what breaks your multi-region plan, depending. Their built-in vector database reaches end of life on 31 August 2026, so plan the move to Managed PostgreSQL with pgvector now if you’re using it.

The Questions That Matter

1. Does the model you need fit on the hardware your provider actually owns? Check their GPU line, not their marketing.
2. Is residency guaranteed per account or per model? Ask specifically.
3. Is zero retention contractual, or a sentence on a landing page?
4. What’s the model lifecycle policy, and what happens to your requests when a model is retired?
5. Who signs as AI Act Provider for the quantized weights you’re calling?

Prices here are a snapshot. Token pricing moves faster than instance pricing, model names rotate every few weeks, and half the third-party trackers are months behind. Mistral Large 3 was still listed at its pre-cut rate on at least one aggregator four months after the change. Verify against the vendor before you commit volume.

---

Last validated: August 2026, against each vendor’s own catalog and price list. Model catalogs change monthly. If you spot an error, let us know.

Related: Which European providers offer GPU instances | GDPR and cloud hosting | Provider details | Compliance matrix