Research
German AI companies and their model providers
Of German AI firms that disclose a model provider, about half run a US frontier model. Mistral reaches 1 in 10; no German-hosted frontier model exists.
Summary · We examined 514 publicly identifiable German AI companies; 221 disclose which AI model they run, and those 221 are the basis for every percentage here. About half (~53%) depend on a US frontier provider (OpenAI, Anthropic, Google, Microsoft or AWS), and a similar share (~43%) build their own models. Mistral, the only EU-native frontier lab, reaches about 1 in 10 (~10%). No German-hosted frontier model fills the sovereignty gap. Germany sits close to the European mean on provider origin; its distinctiveness is mix (more Google and Anthropic), not a different sovereignty posture.
Market research report · June 2026 · 514 German firms examined, 221 disclosing.
Executive summary
Of German AI companies that publicly disclose which AI model they run, about half (~53%) depend on a US frontier provider: OpenAI, Anthropic, Google, Microsoft or AWS. A similar share (~43%) build their own models. Mistral, the only EU-native frontier lab, reaches about 1 in 10 (~10%). And there is no German-hosted frontier model to fill the sovereignty gap.
How to read these numbers (please read before quoting)
Sampling frame. No public register of German AI companies exists, so this is a census of publicly identifiable firms, not a random sample of a known population. The figures describe firms that disclose their provider, not all German AI firms. Correct phrasing: “of German AI companies that publicly disclose a provider, ~X% use Y”. Avoid “X% of German AI companies use Y”.
Denominators. 514 German AI firms were examined; 221 disclose a provider. Those 221 are the basis for all percentages here.
Conservative-headline check. Restricting to the 185 firms with the strongest evidence (provider named on the company’s own site) raises the US-dependence figure to ~60%, because the lower-confidence cases lean toward own and open models. The ~53% headline is therefore the conservative end, not the high end.
What “own model” means. A firm counts as running an own proprietary model when it trains or develops the model itself rather than calling a third-party LLM API: think computer-vision and predictive-maintenance models in manufacturing, risk and fraud models in finance, or validated on-prem medical models. These are usually domain-specific systems, not frontier LLMs, and many firms run them alongside a third-party API, which is why the own-model and US-provider shares can both be high at once.
Overlapping sets. A firm may use several providers and is counted under each. Percentages do not sum to 100% and must never be added. “Any US proprietary” is the de-duplicated union of the five US labs.
Independent audit. 40 disclosed German claims were re-checked blind against the companies’ live websites: 34 fully confirmed, 5 partially confirmed (a secondary provider over-stated), 1 unverifiable, 0 contradicted. No claim named the wrong primary provider or the wrong provider origin.
Snapshot and comparison caveat. Provider relationships change quickly; this is a mid-2026 snapshot. The European benchmark blends dedicated country studies (UK, France, Nordics) with fresh 2026 research (Germany, Netherlands, Spain, Italy); both sides of the comparison use all disclosing firms for fairness. It is robust directionally, not a single controlled panel.
1. Germany vs the rest of Western Europe
Germany: 221 disclosing firms. Europe (15 countries, Iceland excluded): 1725 disclosing firms. Both columns use all disclosing firms, an apples-to-apples comparison.
| Provider (% of disclosing firms) | Germany | Europe (excl. DE) | Difference |
|---|---|---|---|
| Any US proprietary | ~53% | ~49% | +4 pp |
| OpenAI (any, incl. Azure) | ~41% | ~43% | -2 pp |
| Anthropic | ~18% | ~14% | +4 pp |
| Google/Vertex | ~20% | ~14% | +6 pp |
| Microsoft Copilot | ~5% | ~8% | -3 pp |
| AWS Bedrock | ~10% | ~7% | +3 pp |
| Mistral | ~10% | ~8% | +2 pp |
| Own proprietary model | ~43% | ~46% | -3 pp |
| Open-source / self-hosted | ~19% | ~19% | +0 pp |
| Meta Llama | ~6% | ~4% | +2 pp |
Germany sits close to the European mean on both US dependence and own-model rate: a representative, not exceptional, market. Its one clear tilt is toward the challenger US labs, with more Google/Vertex and Anthropic than the OpenAI-dominated European average. Mistral is slightly above the EU mean but still marginal.
2. By industry
Shares are of disclosing firms within each industry. Cells with fewer than ~20 disclosing firms (marked *) are directional only. Read them as “leans toward”, not precise figures.
| Industry | Firms | Disclosing | Any US | OpenAI | Mistral | Own model |
|---|---|---|---|---|---|---|
| Cyber/Data/Infra/Labs | 74 | 29 | ~31% | ~21% | ~14% | ~52% |
| Finance/Fintech | 69 | 22 | ~64% | ~41% | ~5% | ~41% |
| Manufacturing/Industrial | 64 | 21 | ~14% | ~14% | ~0% | ~71% |
| Retail/Media/Marketing | 63 | 30 | ~63% | ~43% | ~10% | ~47% |
| EdTech/Other | 59 | 33 | ~70% | ~64% | ~12% | ~24% |
| Legal/HR/Gov/CX | 52 | 28 | ~71% | ~57% | ~11% | ~25% |
| Energy/Mobility/Logistics | 49 | 13* | ~62% | ~23% | ~8% | ~38% |
| Healthcare/Biotech | 47 | 20 | ~10% | ~5% | ~0% | ~85% |
| Other | 37 | 25 | ~76% | ~72% | ~28% | ~24% |
* fewer than ~20 disclosing firms, directional only.
- Cyber/Data/Infra/Labs. The sovereign cluster: mostly own and open-source or self-hosted models, and the lowest US share. German AI labs and GPU-cloud firms sit here.
- Finance/Fintech. High US usage tempered by own risk and fraud models. Where third-party LLMs appear, Microsoft Azure OpenAI leads (EU residency, no-training terms).
- Manufacturing/Industrial. Build-heavy: vision and predictive-maintenance models are domain-specific and self-trained, so LLM vendors are largely irrelevant to the core product.
- Retail/Media/Marketing. The highest direct-OpenAI usage. Content and marketing firms wire to the cheapest capable API: pragmatic and price-led.
- EdTech/Other. Small builders defaulting to ChatGPT-class APIs; high US, high OpenAI.
- Legal/HR/Gov/CX. The loudest sovereignty messaging of any sector, yet most run a US proprietary model. This is the clearest gap between stated intent and technical reality.
- Energy/Mobility/Logistics. Optimisation models are often proprietary and large language models are peripheral. Low disclosure.
- Healthcare/Biotech. Build-not-buy is near-total. Medical-device rules (MDR/CE) push validated proprietary, on-prem models, and third-party LLMs barely appear.
3. The sovereignty gap
173 of the 514 German firms (~34%) actively market EU or German data residency, sovereignty or on-prem deployment. Of those, 72 run a US proprietary model anyway: committed to a sovereignty message, but not to a sovereign model layer. They concentrate in Legal/HR/Gov/CX and Finance, where compliance pressure is highest.
4. Takeaways for reporting
- Lead line: of German AI firms that disclose, about half depend on a US frontier model, and there is no German-hosted frontier alternative.
- Germany is typical of Europe on provider origin; its distinctiveness is mix (more Google and Anthropic), not a different sovereignty posture.
- The story splits by industry: regulated build-it sectors (health, industrial) are largely self-built, while software-layer sectors (legal/HR, finance, edtech) run mostly on US APIs.
- The sharpest finding is the say-do gap: the sectors that market sovereignty loudest are the most dependent on US models.
Method
Publicly identifiable German AI firms; provider taken from each company’s own website (privacy policy, sub-processor list, trust centre, product and AI pages). Self-disclosed, overlapping sets, point-in-time June 2026.
Compiled June 2026. The percentages cover the 221 German firms that disclose a provider.
Frequently asked questions
Which AI model provider do German AI firms use most?
OpenAI, in some form (direct or via Azure), reaches ~41% of the 221 disclosing firms. Google/Vertex follows at ~20% and Anthropic at ~18%, both above the European average, which makes Germany unusually tilted toward the challenger US labs.
How many German AI firms use a European model?
Mistral, the only EU-native frontier lab, appears at about 1 in 10 disclosing firms (~10%), slightly above the European mean of ~8% but still marginal. No German-hosted frontier model is in routine third-party use.
Do these percentages describe all German AI companies?
No. 514 publicly identifiable German AI firms were examined and 221 disclose a provider; all percentages use those 221 as the denominator. The correct phrasing is 'of German AI companies that publicly disclose a provider, ~X% use Y'.
Which German sectors build their own AI models?
Healthcare and biotech (~85% own models) and manufacturing (~71%), where medical-device rules and domain-specific vision models push validated, self-trained systems. Software-layer sectors such as legal/HR, finance and edtech mostly run on US APIs.
Continue reading
-
Research
Which AI model providers Dutch AI vendors actually use
A study of 240 Dutch AI vendors: OpenAI dominates by a wide margin, Anthropic is the default second model, and EU-owned Mistral is barely visible. On the sovereignty gap that follows.
Read article -
Product
Introducing GreenPT Code: coding models built for agents
Five open-weight AI coding models behind one OpenAI-compatible API, hosted on renewable European energy. Connect your coding agent in minutes.
Read article -
Sustainability
Artificial intelligence and sustainability: a practical guide
AI can cut emissions and waste, but it can also add its own footprint. A practical framework for using AI where it measurably reduces impact, with traceable assumptions.
Read article