Executive summary
- Open models are good, and still behind. The best self-hosted models trail closed models by three to four months. The gap did not close in 2026.1
- The strongest open models are Chinese. Local hosting stops data transfers to China but keeps trained-in bias.3 European and US alternatives sit one tier lower.
- "Open" no longer means unrestricted. Several leading models limit use by company revenue or region.
- Cost rarely justifies self-hosting. Hardware and GPU rental got dearer in 2026 while API prices kept falling. The case is control.
- Data law drives the regulatory case. High-risk AI Act duties start in December 2027. DORA, NIS2 and US data-access law press sooner.
1. Capability and origin
Epoch AI finds the best open-weight models have trailed the best closed models by four months on average since January 2026. In 2023 to 2025 the average was three months.1 On the Artificial Analysis Intelligence Index, the top closed model scores 58 and the top open model 46.2
Open models are good, and still behind
The best open model scores 46, the best closed model 58. All three leading open models are Chinese.
The strongest self-hosted model trails the frontier by 12 points.
Artificial Analysis Intelligence Index v4.3, score
Source: Artificial Analysis, index v4.3, September 2026. Scores are not comparable with earlier index versions.
View data table
| Category | Value | Group |
|---|---|---|
| Claude Opus 5.5 | 58 | 1 |
| GPT-6 Astra | 53 | 1 |
| Claude Fable 5.1 | 53 | 1 |
| MiMo-V2.6-Pro (Xiaomi) | 46 | 2 |
| GLM-5.3 (Z.ai) | 44 | 2 |
| Kimi K3 (Moonshot AI) | 44 | 2 |
Group 1: Closed model (API only). Group 2: Open-weight model (runs on your own infrastructure). Unit: score.
Today's best open model roughly matches the closed frontier of a few months ago. That is often enough for document processing, search and drafting. It is not yet enough for the hardest reasoning.
Every open model near the top comes from a Chinese lab. Italy, the Netherlands, Germany and the Czech Republic have acted against the hosted DeepSeek app and its data transfers. None targeted weights on an organisation's own servers. Local weights remove the transfer, not the behaviour trained into the model. NIST, the US standards body, found DeepSeek models echoed state narratives four times as often as US models.3
2. Licences
Open weights carry very different rights. Mistral Medium 3.5 grants no rights to companies above $20 million in monthly revenue without a commercial licence.4 Meta's Llama 4 policy withholds multimodal rights from EU-based companies.5 Alibaba moved its flagship Qwen model from Apache 2.0 to a custom licence.6 Mistral Small and Large, DeepSeek, Gemma, gpt-oss and Granite, among others, stay permissive. An open-weight licence review checks these terms before a model is chosen.
Permissive licences exist in every region, but not at the very top.
| Model | Licence | Minimum hardware |
|---|---|---|
| Europe | ||
| Mistral Small 4 Mistral AI (FR) | Apache 2.0 | 4× H100 (BF16) |
| Mistral Large 3 Mistral AI (FR) | Apache 2.0 | 8× H200 (FP8) |
| Mistral Medium 3.5 Mistral AI (FR) | Modified MIT | 2× H200 (FP8) |
| EuroLLM-22B EU research consortium | Apache 2.0 | 1 GPU |
| TildeOpen-30B Tilde (LV) | CC BY 4.0 | 1 GPU |
| United States | ||
| gpt-oss-120b OpenAI | Apache 2.0 | 1× 80 GB (MXFP4) |
| Gemma 4 Google | Apache 2.0 | 1 GPU |
| Nemotron 3 Ultra NVIDIA | OpenMDW 1.1 | 8× H200 (BF16) |
| Llama 4 Maverick Meta | Llama 4 Community | 8× H100 |
| Granite 4.2 IBM | Apache 2.0 | 1 GPU |
| China | ||
| MiMo-V2.6-Pro Xiaomi | MIT | 8× B200 |
| DeepSeek V4 Pro DeepSeek | MIT | 8× H200 or 8× B300 |
| Kimi K3 Moonshot AI | Kimi K3 License | 8× B200 minimum |
| Qwen3.8 (flagship) Alibaba | Qwen3.8-Max License | Multi-node |
| GLM-5.3 Z.ai | GLM-5.3 License | 8× H200 |
Source: model cards and licence texts, accessed 27 September 2026. Hardware is indicative for production inference. Full detail in the Model Index.
The strongest open models carry the most licence conditions. Check the licence before choosing the model, not after.
3. Economics
2026 moved the economics against self-hosting. Memory shortages pushed an NVIDIA RTX PRO 6000 from its launch price of €7,511 to about €14,000 on NVIDIA's own store.7 The median H200 rental price rose about 25 percent in twelve months.8 Meanwhile, API prices for a given capability kept falling steeply.9
Renting GPUs is getting more expensive
Amsterdam-based Nebius raises its on-demand prices by 17 to 21 percent on 1 October 2026.
European rental prices are rising for every GPU generation.
On-demand price per GPU, € per hour
Source: verda.com/pricing and nebius.com/prices, accessed 27 September 2026; on-demand list prices; converted at the ECB reference rate of 25 September 2026
View data table
| Category | Series 1 | Series 2 | Series 3 |
|---|---|---|---|
| H100 | 3.09 | 3.38 | 3.95 |
| H200 | 4.03 | 3.95 | 4.74 |
| B200 | 5.98 | 6.27 | 7.45 |
| B300 | 7.34 | 6.88 | 8.33 |
Series 1: Verda (FI). Series 2: Nebius (NL), until 30 Sep. Series 3: Nebius (NL), from 1 Oct. Unit: € per hour
Independent academic research finds owned hardware pays back in about three months for small models, 6 to 24 months for medium ones and often over two years for large ones. It pays mainly at high, steady volumes or under strict data-residency rules.10 Lindstead's break-even analysis agrees: against a mid-priced closed model, an owned eight-GPU server breaks even only at several billion tokens per month, before staff costs.
4. Regulation and control
The AI Omnibus, in force since 27 July 2026, postponed high-risk obligations to 2 December 2027 for stand-alone systems. For AI embedded in regulated products, the date is 2 August 2028. Article 50 transparency duties apply from 2 August 2026.11 A company that only deploys an open model is usually a deployer under the Act.
Other rules press sooner:
- US data access. The US CLOUD Act can compel US providers to hand over data they control, wherever it is stored. Microsoft France told the French Senate under oath it could not guarantee this would never happen.12
- DORA. Financial entities must manage ICT third-party risk and plan exits. In November 2025 European supervisors designated 19 critical ICT providers, including AWS, Google Cloud and Microsoft.13
- NIS2. The Dutch Cyberbeveiligingswet, in force since 15 August 2026, brings supply-chain security duties to more than 8,000 organisations.14
- Transfers. The EU-US Data Privacy Framework stands, but an appeal is pending at the Court of Justice. The legal position is uncertain.15
The duties differ by sector: see AI in financial services under DORA and NIS2 for energy and critical infrastructure. Mapping them to each AI system is the core of AI governance and EU AI Act compliance.
5. Security
Self-hosting removes third-party access to prompts and outputs and adds audit and exit control. It does not make a model safe. Prompt injection tops the OWASP list of LLM risks,16 and the UK National Cyber Security Centre warns it may never be fully mitigated.17 Anthropic and the UK AI Security Institute showed that around 250 poisoned documents can backdoor models of very different sizes.18 Whoever runs the model owns these risks.
Demand for sovereign AI is broad, but rarely owned by the board
Demand is strongest in Germany and the Nordics. Only 16 percent of organisations treat it as a board-level issue.
Six in ten European organisations are looking for sovereign AI.
Organisations actively seeking sovereign solutions, %
1 Share of organisations actively seeking sovereign solutions; survey of 1,928 organisations in 28 countries, July to August 2025.
Source: Accenture, November 2025
View data table
| Category | Value |
|---|---|
| Europe | 62% |
| Germany | 72% |
| Denmark | 80% |
Value: Seeking sovereign AI solutions. Unit: %
Few boards own the question yet. Four steps put it on the agenda.
6. What leaders should do
- Classify workloads by data sensitivity. Most AI use can stay on APIs. Keep private deployment for data too sensitive to share.
- Set a model-origin and licence policy. Decide at board level which origins and licence types are acceptable, before engineers decide by default.
- Start in a European cloud. Rent from an EU-owned provider to prove value before buying hardware at 2026 prices.
- Treat models as supply chain. Verify weights, prefer safe file formats, red-team before production, and plan for monthly model updates.
Accenture reports that 62 percent of European organisations are seeking sovereign AI solutions.19 HSBC chose to run Mistral models on its own infrastructure.20
AI deployment in a European cloud turns policy into a running system. It builds on an AI strategy that classifies workloads by data sensitivity. Model stewardship keeps it current as models and licences change.
Sources
- Epoch AI, Open-closed ECI gap, 29 May 2026. epoch.ai/data-insights/open-closed-eci-gap
- Artificial Analysis, Intelligence Index v4.3 changelog and model pages, September 2026. artificialanalysis.ai/changelog
- NIST CAISI, Evaluation of DeepSeek AI models, 30 September 2025. www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks
- Hugging Face, Mistral-Medium-3.5-128B licence. huggingface.co/mistralai/Mistral-Medium-3.5-128B
- Meta, Llama 4 Acceptable Use Policy. dev.meta.ai/llama/llama4/use-policy
- Qwen3.8-2.4T-A95B licence, Hugging Face. huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- Thunder Compute, NVIDIA RTX PRO 6000 pricing, September 2026. www.thundercompute.com/blog/nvidia-rtx-pro-6000-pricing
- getdeploying.com, NVIDIA H200 rental prices, accessed 27 September 2026. getdeploying.com/gpus/nvidia-h200
- Epoch AI, LLM inference price trends, 12 March 2025. epoch.ai/data-insights/llm-inference-price-trends
- Pan et al. (Carnegie Mellon University and independent researchers), cost-benefit analysis of on-premise LLM deployment, arXiv 2509.18101. arxiv.org/abs/2509.18101
- Lewis Silkin, The Digital Omnibus on AI enters into force, 27 July 2026. www.lewissilkin.com/insights/2026/07/27/the-digital-omnibus-on-ai-enters-into-force-today-102nedo
- The Register, Microsoft exec admits it cannot guarantee data sovereignty, 25 July 2025. www.theregister.com/off-prem/2025/07/25/microsoft-exec-admits-it-cannot-guarantee-data-sovereignty/458553
- EIOPA, ESAs designate critical ICT third-party providers under DORA, 18 November 2025. www.eiopa.europa.eu/european-supervisory-authorities-designate-critical-ict-third-party-providers-under-digital-2025-11-18_en
- Rijksoverheid, Cyberbeveiligingswet in force from 15 August 2026. www.rijksoverheid.nl/actueel/nieuws/2026/07/07/cyberbeveiligingswet-en-wet-weerbaarheid-kritieke-entiteiten-vanaf-15-augustus-2026-van-kracht
- WilmerHale, Court of Justice to review challenge to EU-US Data Privacy Framework, 1 December 2025. www.wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/20251201-european-court-of-justice-to-review-challenge-to-eu-us-data-privacy-framework
- OWASP, Top 10 for LLM Applications 2025. genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/
- UK NCSC, on prompt injection, December 2025. www.ncsc.gov.uk/news/mistaking-ai-vulnerability-could-lead-to-large-scale-breaches
- Anthropic, UK AI Security Institute and Alan Turing Institute, A small number of samples can poison LLMs of any size, October 2025. www.anthropic.com/research/small-samples-poison
- Accenture, Europe seeking greater AI sovereignty, November 2025. newsroom.accenture.com/news/2025/europe-seeking-greater-ai-sovereignty-accenture-report-finds
- HSBC and Mistral AI partnership, 1 December 2025. www.hsbc.com/news-and-views/news/media-releases/2025/hsbc-and-mistral-ai-join-forces-to-accelerate-ai-adoption-across-global-bank
Method: desk research of public sources, accessed 27 September 2026. Vendor-sponsored data is labelled as such. This briefing contains no paid content. Corrections: contact@lindstead.com.