Which AI model provider do Dutch AI vendors use most?

OpenAI, by a wide margin. Of the 240 vendors, 69% use OpenAI directly, rising to 79% when Azure OpenAI is included. Anthropic (Claude) comes second at 24%, almost always as a secondary or optional model alongside OpenAI.

How many Dutch AI vendors use an EU-owned model?

Almost none. Mistral is the only EU-owned frontier model in the entire study and appears with just 5% of vendors. Every other commercial frontier model in use is American.

Which AI model providers Dutch AI vendors actually use

Q: What is the EU sovereignty gap?

The difference between stated intent and actual choice. 62% of vendors communicate EU hosting or digital sovereignty as a value, while only around 5% actually run an EU-owned model. The remaining sovereignty claims rest on American models in an EU region, which addresses data residency but not jurisdiction.

Summary · The Dutch applied-AI market runs predominantly on American frontier models, with OpenAI as the leader. Among vendors, 69% use OpenAI (79% including Azure OpenAI), Anthropic comes second at 24%, and EU-owned Mistral appears with just 5%. The key message: 62% of vendors signal EU hosting or sovereignty, while only around 5% actually run an EU-owned model.

Prepared for GreenPT market research · 16 June 2026 · sample of 240 vendors.

Executive summary

The Dutch applied-AI market runs predominantly on American frontier models, with OpenAI as the leader:

OpenAI is used by 69%; the OpenAI family (including Azure OpenAI) reaches 79%.
Anthropic (Claude) comes second at 24%, almost always as a secondary or optional model alongside OpenAI.
Any Microsoft route (Azure OpenAI or Copilot) covers 42%, the standard for large enterprises and system integrators.
Mistral, the only EU-owned frontier model in the study, appears with just 5%.
Open source and self-hosted models are used by 11%, a real but minority segment, driven by sovereignty.

62% of vendors signal EU hosting or sovereignty, while only around 5% actually run an EU-owned model. That gap between stated intent and the reality of American models is the market opportunity.

The central finding in one image: well over three fifths of vendors communicate EU sovereignty as a value, but only one in twenty actually runs an EU-owned model. Everything else rests on American models.

1. Goal and methodology

1.1 Goal

To quantify which AI model providers Dutch AI vendors use in their products and services, and to express the result in percentages, with particular attention to the split between American frontier models and EU-owned or open-source models, which is relevant for an EU-hosted LLM provider.

1.2 Building the sample

The starting point was a compiled market map of Dutch and internationally active AI players: global system integrators, Dutch pure-play data and AI vendors, GenAI-native builders, a long tail of automation and agent builders, a harvest from professional networks and a batch from public company directories. After de-duplication this yielded 240 unique vendors.

1.3 Provider research

Provider usage was established vendor by vendor through web research, in this order of evidence: (1) their own technology, stack, “how it works”, subprocessor, DPA, trust or security pages; (2) case studies, engineering blogs and documentation; (3) cloud partner badges (Microsoft, Google Cloud, AWS, which imply Azure OpenAI, Vertex and Bedrock respectively); (4) job postings that name exact tools; (5) press, professional networks and directories. Nine parallel research rounds covered the full sample; the results were merged with the provider signals already on record.

Every provider assignment carries a confidence level: Confirmed (named on the vendor’s own source), Signalled (partner badge, job posting or external source), or Inferred (product type strongly implies it, used sparingly). Where nothing could be established, the vendor was recorded as Not disclosed rather than guessed.

1.4 The two denominators (read this before the figures)

Because most vendors do not publish their stack, a single percentage would be misleading. Two complementary views are reported throughout:

”% of vendors with a known stack” (n = 240), the share among vendors whose model usage could actually be established. This is the clean comparison and the headline figure.
”% of all vendors” (n = 720), which counts non-disclosure as unknown. These are therefore conservative lower bounds, not upper bounds.

1.5 How to read the percentages (why they add up to more than 100%)

These shares are overlapping sets, not pie slices. Each row answers an independent yes/no question, namely “does this vendor use this provider?”, and most vendors use several providers at once. One vendor is therefore counted in every row that applies. As a result the columns deliberately add up to well over 100%. A vendor like WeAreBrain or SAMEN AI appears at the same time under OpenAI and Anthropic and Google and Mistral. The percentages are therefore not mutually exclusive and must never be summed.

“OpenAI family” is the only de-duplicated figure. Because OpenAI (direct API) and Azure OpenAI deliver the same GPT models, some vendors use both. Simply adding the two rows would double-count them, so the “OpenAI family” line counts each vendor only once:

OpenAI access route	Vendors	% of 240
OpenAI direct only	123	51%
Azure OpenAI only	24	10%
Both OpenAI direct and Azure	42	18%
OpenAI family (any route, de-duplicated)	189	79%

So 69% (direct) plus 28% (Azure) is not 97% of vendors. It is 189 unique vendors, or 79%, because 42 vendors sit in both rows. The same overlap logic applies to every provider in the ranking.

The OpenAI family de-duplicated: 123 vendors use only the direct API, 42 use both direct and Azure, and 24 use Azure only. Together 189 unique vendors (79%), not 97%. The remaining 51 vendors with a known stack ("51 other") use no OpenAI route at all.

2. Provider ranking

The headline result. The middle column, % of the 240 vendors with a known stack, is the key figure.

How to read this table: the percentages are overlapping sets, not a pie. A vendor that uses several providers is counted in every relevant row, so the columns add up to more than 100% by definition and must not be summed (see §1.5). “OpenAI family” below is the only de-duplicated union figure. ”% of all vendors” is shown only where it is meaningful.

Provider / tool	Vendors	% (n=240)	Origin
OpenAI	165	69%	US
Azure OpenAI	66	28%	US (MS)
Anthropic (Claude)	57	24%	US
Microsoft Copilot	57	24%	US
Google / Vertex	45	19%	US
AWS Bedrock	21	9%	US
Mistral	12	5%	EU 🇪🇺
IBM watsonx	15	6%	US
Open source / self-hosted	27	11%	Self-host
n8n (orchestration)	15	6%	EU self-host
Make.com (orchestration)	12	5%	EU/US SaaS

Note: rows add up to more than 100% because many vendors use several providers; ”% of all” is shown only where meaningful.

The OpenAI family dominates at 79%, followed at a distance by the Microsoft routes (42%) and Anthropic (24%). The only EU-owned and self-hosted options (green) sit at the bottom: open source at 11%, Mistral at 5%.

3. What the figures mean

3.1 OpenAI is the standard, by a wide margin

OpenAI is the de facto default layer of the Dutch applied-AI stack: 69% of vendors with a known stack, rising to 79% once Azure-hosted OpenAI is included. It spans every layer: global system integrators, Dutch consultancies and the GenAI-native long tail. Confirmed examples: 10xStudio, AI Advies Bureau, AI Opener, AI-ris, Accenture Netherlands, Aigency, Aizy, Appfront, Avanade, BigData Republic.

3.2 Anthropic is the default second model, rarely the primary one

Claude appears with 24% of vendors with a known stack, but almost always as an optional or secondary model behind OpenAI within a multi-model setup, not as the house standard. Confirmed examples: 10xStudio, AI Advies Bureau, AI Opener, Accenture Netherlands, Aizy, Blits.ai, CleverTech, Codelevate, Conversed.ai.

3.3 Microsoft is the gateway for large organisations

Azure OpenAI and Copilot combined means 42% of vendors with a known stack reach OpenAI models through Microsoft, the dominant route for system integrators, the Big Four and Microsoft partners, typically presented with the EU Data Boundary as an answer to data residency (not jurisdiction).

3.4 The EU-owned gap: Mistral is barely visible

Mistral is the only EU-owned frontier model in the entire study, and it appears with just 12 vendors (5%). Every other commercial frontier model in use is American. This is the central structural fact of the market.

Schematic: the market for frontier models in the Netherlands is almost entirely American. Mistral is the only EU-owned option and holds a narrow margin.

3.5 Open source and self-hosting is the real sovereignty choice

Open source or self-hosted models are used by 11% of all vendors, the genuine data-sovereignty segment, concentrated among vendors that serve government, healthcare and regulated clients. Examples: AI Advies Bureau, AI Horizon, AI-Recruiting, AIFAIS, Appfront, CAIDEL, CleverTech, Data Science Lab (DSL), DecodifAI (Nodevate), FenxLabs.ai, GLBNXT, Goodspeed, LearnWise AI, Louwman AI Automation. These vendors align most naturally with an EU-hosted offering.

13% of vendors run their own models for computer vision, perception, fraud detection or language processing, and are therefore not in the LLM provider market at all. Examples: AI-InfraSolutions, Aiir Innovations, Birds.ai, BlueGen.ai, BrainCreators, Crux Digits, Fizyr, Fourthline, Friss, Gambit Cyber, GeoPhy, Geronimo.AI, HULO.ai, Healthplus.ai. They are excluded from the LLM provider analysis, but count towards sizing the addressable market.

3.7 The multi-provider and gateway pattern

The most advanced builders are explicitly model-agnostic and route across several providers, often through a gateway. That is exactly the abstraction an EU-hosted model slots into. Examples with three or more frontier providers: Accenture Netherlands, Tata Consultancy Services (TCS), Xebia (including GoDataDriven and Xccelerated), Green Creatives, WeAreBrain, orq.ai, AI Opener, Blits.ai, SAMEN AI.

4. Implications for GreenPT

Displacement, not greenfield. Roughly 8 in 10 vendors with a known stack already run OpenAI-family models. This is about replacing an existing choice, not first adoption, and it is easiest where a gateway already abstracts the model.
Sell on the gap between intent and reality. 62% of vendors communicate EU hosting or sovereignty as a value, while only around 5% use an EU-owned model. That contradiction is both the qualifying question and the heart of the pitch.
Warmest segment: open source and self-hosters (11%) plus the government, healthcare and regulated group. They already chose control over convenience; an EU-hosted managed model removes their self-hosting burden.
Rebut “Azure EU region = sovereign”. The most common counterargument is Microsoft’s EU Data Boundary. The message has to draw the distinction between data residency and jurisdiction (US CLOUD Act) sharply, because 42% reach OpenAI through Microsoft.
Mistral is the only EU competitor, and it is barely present. There is real room for a first mover with a Dutch or EU-hosted alternative.

5. Limitations and reliability

Non-disclosure. Not all vendors publish their model choice. The percentages cover the vendors whose stack could be established and assume those are representative, plausible but unproven.
Mixed confidence. The headline figures bundle confirmed, signalled and inferred assignments. A stricter confirmed-only reading gives a fully defensible view; it lowers absolute shares but preserves OpenAI’s dominance.
Implementation agencies build on client choice. Many implementation partners use whatever the client mandates, so their “provider” is situational rather than fixed.
Snapshot. Compiled in June 2026; the GenAI-native long tail changes fast, with many vendors founded in 2024 and 2025.
Consistency check. The extended sample (n=240) closely tracks an earlier verified count of 54 vendors (OpenAI 61% to 69%, Anthropic 28% to 24%, Mistral 6% to 5%), which increases confidence in direction and magnitude.

6. Two reading views of the data

The analysis is built from two complementary views, both summarised above:

Aggregated provider count. The ranking, both denominators, the core aggregates and the accompanying bar charts (§2). This is the bird’s-eye view of the market.
Per-vendor overview. All 240 vendors with their assigned provider, confidence level, EU or hosting signal and source. Vendors without any signal are marked separately so they do not muddy the denominators.

Together these views make it possible to follow both the market-wide distribution and the underpinning for each vendor.

Compiled 16 June 2026. Figures derived from the study described in §1. The percentages cover the vendors whose model choice could be established.

Which AI model providers Dutch AI vendors actually use

Executive summary