Four Ways to Build an AI Venture: Freedom-to-Operate and the Odds of Scaling in Europe
Artificial intelligence now attracts a larger share of European venture capital than ever: in 2025, 31% of all funding raised in Europe went to AI and machine learning companies (Atomico, State of European Tech 2025). Capital, however, is not success. Of the European companies that raised a seed round in 2010–2013, only 1.2% went on to reach a valuation above $1 billion, while 77% ended out of business or self-sustaining without further funding or an exit (Atomico, Dealroom).
Behind the label “AI startup” sit very different businesses. Some build products on models owned by others; some create new models from mathematical theory. What separates them is not the algorithm, but how much of the technology they exploit they actually control. That difference shapes which risks they face and where they are most likely to fail.
This article draws on more than two decades of experience in the deep tech ecosystem, as founder of a university AI spin-off through to its exit and as advisor to AI and deep tech ventures across Europe, combined with official and widely respected sources. Part I classifies AI ventures on a gradient of freedom to operate. Part II follows 100 ventures of each type from founding to scale, milestone by milestone.
Part I: Four types of AI ventures ranked by freedom-to-operate
Not all AI ventures are the same business. Two companies can both be called AI startups while one builds a product on someone else’s model and the other creates a new model from mathematical theory. What separates them is how much of the technology they exploit they actually control, and on whom that control depends.
The framework below combines five references: i) WIPO’s definition of freedom to operate provides the central concept; ii) the EU AI Act separates providers of general-purpose AI models from the downstream providers who build on them; iii) the OECD Framework for the Classification of AI Systems classifies AI by its data and its model; iv) pitchBook’s AI market segments reflect how investors split the market; and v) the EU Technology Readiness Level scale measures how mature the technology is when the venture starts.
Freedom-to-operate and whom it depends on
WIPO defines freedom-to-operate as being free to use a product or service as planned without incurring legal liability. It also treats it as a business matter: a product development and risk management strategy (WIPO).
For AI ventures we take the business view and assess five dimensions. For each one, the question is who holds the control when the venture does not:
- Technology: who owns the core model and the IP behind it. When the venture does not, control sits with the model provider or with the patent and IP holders.
- Cost: who sets the price of the critical inputs. These are API usage, cloud capacity and compute, set by model providers, cloud providers and chip suppliers.
- Roadmap: who decides how the core technology evolves. When the venture does not, it must adapt to someone else’s release cycle.
- Regulatory role: which obligations the EU AI Act places on the venture. These are supervised by the AI Office and national authorities, alongside any sector regulator.
- Market access: whether the venture reaches customers directly or through an intermediary that controls distribution.
The more of these dimensions a venture controls, the higher its freedom-to-operate. The price of that freedom is time and capital: the more a venture owns, the later it meets the market.
Four types along the gradient
A venture’s type is set by what it owns and creates, not by the AI technique it uses. The same technique, such as combining neural and symbolic methods, can appear in any type. It belongs to Type 4 only when the venture invents the method itself.
Type 1: Applications built on third-party models
These ventures build products on a model they access from a third-party, typically through an API, and add value through the user experience and the workflow around it. Because the core technology already works, they can reach a usable product very quickly.
- Technology: borrowed. The core model belongs to the provider.
- Cost: set by the model provider’s pricing.
- Roadmap: follows the provider’s releases, which may also add the venture’s own feature to the base product.
- Regulatory role: downstream provider under the AI Act, integrating someone else’s model into its own system.
- Market access: direct and fast, but the same route is open to anyone using the same model, including the model provider itself.
- Depends on: the model provider above all, then the cloud provider behind it.
- Starting maturity: high on the TRL scale from day one.
- Freedom to operate: very low.
Type 2: Vertical AI with proprietary data and adapted models
These ventures also start from existing models, but add what the model provider does not have: proprietary sector data, models adapted to a specific domain, and regulatory know-how for one industry. They fall within the vertical applications segment, which led AI venture deals in 2025 with 4.609 transactions against 1.821 for horizontal platforms (PitchBook).
- Technology: partly owned: data and domain adaptation are theirs, the base model is not.
- Cost: still largely set by model and cloud providers, though own or open models can reduce that exposure.
- Roadmap: shared between the venture’s domain priorities and the base model’s evolution.
- Regulatory role: downstream provider. A venture that modifies a model substantially enough can become the provider of a new model and take on those obligations.
- Market access: mostly direct, but conditioned by data partners and by sector regulatory clearance.
- Depends on: data holders and partners, sector regulators, and the providers of the base model.
- Starting maturity: mid TRL; the base technology works, the domain performance must still be validated.
- Freedom to operate: low to medium, rising with the depth of its data and integration.
Type 3: AI platforms and tooling on existing infrastructure
These ventures own one layer of the AI technology stack, such as MLOps and model-evaluation tooling, which is embedded in other AI systems.
- Technology: owned for their layer; the infrastructure underneath is not.
- Cost: set largely by cloud and compute providers.
- Roadmap: their own, but it must stay compatible with the main model and cloud ecosystems.
- Regulatory role: generally outside the model provider obligations; their customers’ AI Act needs often shape demand.
- Market access: conditioned by compatibility with the main cloud and model ecosystems, whose integrations often act as the gateway; once embedded, hard to displace.
- Depends on: hyperscalers and model ecosystems, which can also bundle the same function into their own platforms.
- Starting maturity: mid to low TRL.
- Freedom to operate: medium.
Type 4: New foundational AI built from theoretical mathematics
These ventures create the core technology itself: new model architectures or learning methods derived from theory, with the IP owned by the venture. They fit the deep tech definition of novel scientific or engineering breakthroughs turned into products (Dealroom, Lakestar, Walden Catalyst).
- Technology: owned: the science, the IP and the models.
- Cost: driven by access to large-scale compute and research talent.
- Roadmap: fully their own.
- Regulatory role: provider of a general-purpose AI model when such a model is placed on the EU market, with obligations on technical documentation, copyright policy and a public summary of training content (European Commission).
- Market access: often indirect at first, relying on licensees or partners to carry the technology to market until the venture has its own product.
- Depends on: compute providers, long-term funders (public and private), and, for spin-offs, the research institution that holds the original IP.
- Starting maturity: the lowest, at basic principles observed or experimental proof of concept on the EU TRL scale.
- Freedom to operate: high.
Scope: the AI software layer only
This analysis covers AI software only. Ventures whose core asset is hardware are out of scope, because software and hardware scale in fundamentally different ways. Software can serve each new customer at almost no extra cost, while hardware scales through manufacturing, inventory and supply chains. PitchBook tracks these ventures in separate segments, semiconductors and autonomous machines.
The rule of thumb: low freedom buys speed to market and exposes the venture to a competitor it cannot control. High freedom buys defensibility and exposes it to the long, capital-hungry road from lab to market.
Part II: From founding to scale, how many make it and where they fail
Knowing what a venture owns tells us what risks it carries. It does not tell us how far it gets. Part II follows the four types along the road from founding to sustainable, scalable market success, and asks two questions: at which milestones do AI ventures most often stall, and how many of them overcome each one? To make the four types comparable, we use a base-100 analysis: we imagine 100 European ventures of each type starting at the same point and count how many are still moving after each milestone, within a common observation window of about ten years. The window is the same for all types, but the time each type needs is not. In my experience, applications built on third-party models can scale in four to seven years, vertical AI and platforms in six to ten, while new foundational AI often needs ten to fifteen. Measured at year ten, some Type 4 ventures counted as not yet successful are therefore still on their way.
No single dataset tracks AI ventures by type through every milestone, so the funnels are an illustrative model. It is anchored in official and widely used European sources: Eurostat business survival statistics, Dealroom’s cohort studies with Atomico and LocalGlobe, the European Investment Bank, the Draghi report, the European Innovation Council, CB Insights and MIT’s NANDA initiative. The adjustments between types come from my own experience and investigation of AI and deep tech ventures.
The standard itinerary of an AI venture
Whatever an AI venture owns, the road to scale passes through the same sequence of proofs. First the technology has to work outside the lab, and then someone has to pay for it. That payment has to repeat with many customers who keep using the product. The revenue behind it has to come from a product or recurring subscription rather than from grants, research projects or tailor-made services. Then capital has to be raised on the strength of that traction, and a go-to-market engine has to be built that can sell beyond the first market and the founders’ own network. Only after all of that, typically between 5 and 15 years after founding depending on the type of venture, can it be called a scalable market success.
The four types do not differ in the milestones they face but in where they start, how easily they clear each one, and in a few cases the order. A Type 1 venture meets the first milestone almost on day one, while a Type 4 venture may spend years and most of its capital there. A Type 4 venture often raises significant money before it has any product–market fit, which is why we count funding as a milestone only when it is raised on market traction. The milestones are therefore the same for all types, even though some are trivial for one type and decisive for another:
- Technology validated. The solution works in an operational environment, which on the EU TRL scale means a system prototype demonstrated in operational conditions (TRL 7). Achieving it means the technical risk is largely retired.
- Market validation. First paying customers or paid pilots confirm that a real problem is being solved. Achieving it means someone outside the founding team is willing to pay.
- Product–market fit. Demand repeats: new customers buy the same product, existing ones keep using it and act as references. Achieving it means the product sells without being reinvented for each client.
- Business model validated. Revenue comes mainly from a product, licence or SaaS model with healthy unit economics, not from grants, R&D projects or bespoke services. Achieving it means growth will not require proportional growth in headcount.
- Growth funding secured. A Series A or equivalent round is raised on the basis of traction, not only on the team or the science. Achieving it means professional investors are betting on the market, not only on the technology.
- Scalable go-to-market. A repeatable sales and partner engine works across several countries or segments, typically the stage of Series B. Achieving it means growth no longer depends on the founders selling.
- Scalable market success. Once the venture grows on recurring product revenue and can finance further growth from revenue or from capital markets. Achieving it means the venture is a scaleup, not a survivor.
A venture that stops before milestone 7 does not always close. Many keep operating on grants, projects or services. In this analysis they count as survivors, not as successes.
What the evidence says, milestone by milestone
The sources do not split their data by our four types, but together they show where the itinerary breaks:
- Survival baseline. Across all new EU businesses, only 45% of the enterprises born in 2013 were still active five years later (Eurostat). Staying alive is already a coin toss before any question of scale.
- Technology validated (milestone 1). Only about a third of the inventions patented by European universities are commercially exploited (Draghi report). Research-born technology, the origin of most Type 4 ventures, mostly never reaches the market.
- Market validation and product-market fit (milestones 2-3). In CB Insights’ analysis of 111 startup post-mortems, 35% cited no market need and 20% being outcompeted; 38% ran out of cash or failed to raise, which is usually the consequence of the first two (CB Insights). For AI specifically, MIT found that only 5% of custom enterprise AI tools reach production (MIT NANDA). Most AI pilots never become recurring business.
- Growth funding (milestone 5). In Europe, 19% of seed-funded startups raised a Series A within 36 months (Dealroom, Atomico, LocalGlobe). The EIC shows what support can achieve: its portfolio of about 700 startups has attracted over €12 billion of follow-on investment, and more than 70 companies are valued above €100 million (EIC).
- Scalable go-to-market (milestone 6). Only 10% of European deep tech companies convert from Series A to Series B, against 24% in North America (Global Deep Tech Report 2026, via Munich Startup). European scaleups also raise about half the capital of their San Francisco peers over ten years (EIB).
- Scalable market success (milestone 7). Dealroom followed 1,064 European companies that raised a seed round in 2010–2013 until late 2020, seven to ten years after their seed. 43% never raised again, 77% ended out of business or self-sustaining without further rounds or an exit, only 3% reached a sixth round, and 1.2% reached a valuation above $1 billion (Atomico, Dealroom).
Two conclusions follow: First, the steepest losses happen at the market milestones (2 to 4), not at the technology milestone, except for research-born ventures. Second, the long-run European cohort data confirm the order of magnitude of our end point: a few in a hundred reach durable scale.
The funnel of 100 ventures, type by type
Type 1: Applications built on third-party models. Almost all of them clear the technology milestone, since the model already works: 95 of 100. The fall comes at the market. About 55 win paying pilots, but only 22 reach product-market fit, because what they sell is easy to copy and the model provider can ship the same feature. Only 10 turn usage into a sustainable product business once inference costs are counted. 7 raise growth funding and 3 build a scalable go-to-market. About 2 reach scalable market success. Main failure point: product-market fit.
Type 2: Vertical AI with proprietary data and adapted models. 80 validate their technology in the target domain, and 50 sign first paying customers. 28 reach product-market fit, the highest of the four types, because domain data and regulatory know-how are hard to copy. The decisive drop is the next one: only 15 move from tailor-made projects for each client to a repeatable product model. 10 raise growth funding and 6 scale across countries. About 5 reach scalable market success. Main failure point: the shift from projects to product.
Type 3: AI platforms and tooling. 75 have a working platform, but only 40 get paying adopters, because their value appears only once embedded in someone else’s system. 20 reach product-market fit and 11 validate a pricing model. 8 raise growth funding, and 4 stay defensible long enough to scale as cloud and model providers bundle similar functions. About 3 reach scalable market success. Main failure points: adoption, then defensibility.
Type 4: New foundational AI from theoretical maths. This is the only type that loses most of its cohort at the technology milestone: just 30 of 100 bring their method to an operational demonstration. 18 find a first paying use case or licensee, 10 reach product-market fit, and 6 build a product or licensing business that does not depend on grants and research contracts. 4 raise growth funding on traction rather than on the science, and 2 scale. About 1 reaches scalable market success. Main failure point: the climb from the lab to an operational demonstration, then the move from grants to market revenue.
The four funnels compared

Illustrative base-100 model anchored in Eurostat, Dealroom/Atomico, EIB, EIC, CB Insights and MIT NANDA data. M1 technology validated, M2 market validation, M3 product-market fit, M4 business model validated, M5 growth funding, M6 scalable go-to-market, M7 scalable market success
The funnels start from the European baselines above and are adjusted for each type using the segment evidence and my experience. The figures show orders of magnitude and the location of the main drops, not a measured census.
The type with the most freedom-to-operate loses most of its cohort before it ever meets a customer. The types with the least freedom meet the market quickly and lose most of their cohort there. Across all four, the end point is the same order of magnitude: between 1 and 5 in 100.
Conclusion: the path-to-market decides, not the technology
The four types of AI ventures carry very different risks. Some depend on someone else’s model, others own the science outright. They start at different points, and they fall at different milestones. Yet they end in the same place: within a ten-year observation window, only between 1 and 5 out of every 100 become scalable market successes. The rest close, or survive on grants, research projects and services without ever becoming a scalable product business.
The evidence points the same way for every type. Outside research-born ventures, the steepest losses happen at the market milestones: market validation, product-market fit and the move to a repeatable business model. Founders’ own post-mortems name no market need and competition far more often than technical failure, and most AI pilots in companies never reach production. Even for foundational AI, where the technology is the first hurdle, the ventures that make it are those that find a first paying use and leave grant dependency behind.
The lesson is that the technology or the solution is never what decides success. The path-to-market is. Freedom-to-operate determines which risks a venture faces. A deliberate path-to-market strategy determines whether it overcomes them: knowing which customers to target, which business model to build, which milestones to prove and in which order, and when to raise capital on traction rather than on promise.
- Know where your venture sits on the freedom-to-operate gradient, and which milestone is most likely to stop you.
- Plan the path-to-market from the start, with the same rigour as the technology roadmap.
This is the logic behind the Path2Market methodology. If you want to test where your AI venture stands, request a free 1-hour speed-coaching session.
Sources
- WIPO, Freedom to Operate tool description and Managing IP for access: freedom to operate
- European Commission, General-purpose AI models in the AI Act: questions and answers
- OECD, Framework for the Classification of AI Systems (Digital Economy Papers No. 323, 2022)
- PitchBook, AI market map: 500 leading VC-backed AI companies
- European Commission, Technology readiness levels, Horizon 2020 General Annexes
- Dealroom, Lakestar, Walden Catalyst, The 2026 European Deep Tech Report
- Eurostat, Key figures on European business 2021: business dynamics
- Atomico and Dealroom, State of European Tech 2020: Seed to $1B+
- Atomico, State of European Tech 2025, as reported by Euronews
- Dealroom, Atomico, LocalGlobe, The Journey to Series A in Europe
- Global Deep Tech Report 2026, as reported by Munich Startup
- European Investment Bank, The scaleup gap (2024)
- European Commission, The future of European competitiveness (Draghi report) (2024)
- European Innovation Council, EIC Impact Report 2025
- CB Insights, The top reasons startups fail
- MIT NANDA, The GenAI Divide: State of AI in Business 2025










