AI startups made in Germany — Specialisation instead of computing power

‘Made in Germany‘ has long stood for mechanical engineering and German craftsmanship, yet in the global AI race that label suddenly seems to count for little. While US corporations compete for dominance with increasingly sophisticated AI models, Germany is still searching for its role. What could Germany’s distinctive strength be, and what might ‘Made in Germany’ stand for in the age of AI? Drawing on ten interviews with German AI startups, this article examines how AI is creating economic value in Germany in practice. The findings show that hardly any of the startups are trying to compete with OpenAI or Google. Instead, their founders rely on specialisation, deep industry knowledge and exclusive data. What does this mean for Germany’s position in the international AI race?

How can Germany’s AI startups grow and benefit from the technology when the United States and China seem to be pulling ever further ahead? The US AI market is booming, while Europe appears to be falling behind.

The market valuations of dominant US companies in the field of artificial intelligence have reached gold-rush proportions (Shay & Davenport, 2026), and investment in these companies accounted for one-third of total investment in US markets in the third quarter of 2025 (Kalyani, 2026). The AI market now accounts for just over a third of US economic growth, exceeding the contribution of the digital sector at the height of the internet boom in 2000 (Rubinton & Patro, 2026). Germany, however, appears to have benefited little from this development so far. Its economy is struggling, and the latest AI Index Report from Stanford (Sha et al., 2026) shows that Europe is far behind in the development of powerful new AI models.

Where is the productivity boost?

Some analysts see the AI products of major US companies as a potential source of the productivity gains Germany needs to generate renewed economic growth. So far, however, even modest productivity gains have largely failed to materialise in established companies (Stelter, 2025). Instead, there is growing recognition that AI-driven productivity gains are likely to emerge gradually over many years. This points to a new productivity paradox: although AI is a powerful technology, many of the complementary innovations needed to use it profitably are still lacking. Enthusiasm for powerful AI models is also tempered by the recognition that, while they perform many complex tasks very well, they repeatedly make errors that undermine productivity (Dell’Acqua et al., 2023) and sometimes fail even at simple tasks such as reading a clock (Saxena et al., 2025).

Rather than inspiring optimism, the challenges surrounding AI are currently creating considerable uncertainty in Germany. There are clear indications that AI use is associated with an erosion of users’ skills (de-skilling) (Rinta-Kahila et al., 2023), threatens jobs in some markets (Hui et al., 2023), and exacerbates existing pay inequalities or contributes to shifts in power to the detriment of employees (De Micheli et al., 2025).

AI models are becoming a standardised resource

More encouragingly, the cost of using AI models has fallen rapidly over time, while providers charge very high prices for increasingly small gains in computing power (Demirer et al., 2025). Open-weight models are also catching up quickly with proprietary foundation models, such as those offered by OpenAI, reaching comparable levels of performance only a few months later (artificialanalysis.ai, 2026). Organisations can run open-weight models themselves on reasonably powerful computers, freeing them from the pricing models of dominant providers. They can also retain full control over how the data they process is stored and used (Abonamah et al., 2021). Marketplaces such as OpenRouter allow users to switch quickly between providers and compare the prices, latency and benchmark results of different AI models.

Six types of AI startup

To distinguish between and assess fundamentally different value creation models among AI companies, Shay and Davenport (2026) draw on existing approaches to present a taxonomy of AI startup types. These types differ in their value propositions and face distinct competitive challenges:

• AI Originators work on new foundation models and are typically highly capital-intensive.
• AI Explorers prepare the technology for future generations of AI models, for example working on quantum computing or agentic AI.
• AI Infrastructure Builders provide the software- (e.g. vector databases) and hardware-side (e.g. AI chips) infrastructure for AI companies.
• AI Enhancers use existing, broadly applicable AI models and build their applications on top of them — for example, generating content for marketing purposes.
• AI Optimizers do not reflect their use of AI directly in their value proposition, but instead use it to optimise internal processes in traditional businesses, such as logistics planning.
• AI Experimenters describe the most conservative segment, in which companies have not integrated AI as a strategic resource and instead experiment with its use in various areas.

Shay and Davenport (2026) use this typology to examine potential competitive advantages and ways in which each category can differentiate itself. They also consider how organisations can align their AI activities with their respective roles. AI Originators, Explorers and Infrastructure Builders create value through technological innovation closely related to AI models. AI Enhancers, by contrast, primarily use existing technology but combine it with “proprietary data, domain focus and user experience” (Shay & Davenport, 2026, p. 3). AI Optimizers and Experimenters are not AI-centric startups; they occupy more peripheral positions in the AI landscape.

How German AI startups create value

We conducted ten interviews with founders of young AI startups founded between 2020 and 2025. The startups employed between 8 and 60 people and were selected to cover as broad a range of use cases as possible. Our goal was to examine their use of AI without presupposing what we would find, given that the technology shapes core elements of their business models. None of the startups we studied is seeking to develop its own foundation model or compete directly with the major US providers. Instead, the founders draw on existing AI technologies and combine them with in-depth knowledge of specific industries, processes and problems. The innovation lies not in the model itself, but in how it is applied and adapted to its context.

From the clinic to the classroom

This is particularly evident in the healthcare sector. Two startups use AI to improve medical treatment processes. One makes individual biological rhythms visible and derives recommendations for the optimal timing of medical treatments. Another combines clinical services with software and AI solutions to improve treatment processes in a specialised area of medicine. The focus is on involving patients more closely in treatment and decision-making processes and creating transparency about the course and prospects of success of treatments.

This pattern is also evident in education. One startup pursues the vision of an AI-supported learning companion for school students. The focus is not on developing new language models, but on how learning content, learning behaviour and individual needs can be combined to make learning more personalised and effective.

Detecting errors, uncovering fraud, automating marketing

Similar developments can be observed in industry. Two startups use AI for automated quality control in manufacturing processes, helping companies detect errors and deviations at an early stage. One develops AI-based solutions for visual quality inspection in industry. Using synthetic data, it creates tailored datasets and inspection algorithms for a range of testing tasks.

Three further startups offer solutions in e-commerce, fraud detection and process optimisation. One brings personalised advice from bricks-and-mortar retail into the digital realm, enabling AI-driven product recommendations for online shops. Another helps insurers in particular detect manipulated or AI-generated images in claims. As generative AI makes it ever easier to produce deceptively realistic content, reliably verifying its authenticity is becoming an increasing challenge. Again, the innovation lies not primarily in the AI technology itself, but in its application to a concrete problem in the insurance industry: the early detection of fraud attempts using digital evidence. A further startup uses AI to make internal processes in the trade of refurbished second-hand products more efficient. Unlike most of the startups mentioned so far, AI is not at the centre of its value proposition, but instead primarily supports the scaling of the business model.

While the startups mentioned so far improve existing decision-making or inspection processes, others focus on partially automating human labour. One startup, for example, develops AI-based voice agents that handle telephone customer enquiries, arrange appointments or qualify leads. Another uses AI to automate large parts of digital marketing, creating personalised campaigns and landing pages at scale. Despite their differences, these startups share a common logic: rather than advancing general AI technologies themselves, they use them to solve real, highly context-dependent problems with the help of AI.

Where the startups fit in

A comparison of the ten founding companies interviewed shows that the majority fall into the Enhancer category. Several startups explicitly state in the interviews that they work in a model-agnostic manner, building applications on top of generic AI models, as described by Shay and Davenport (2026). Only one startup falls into the “Optimizer” group. The value of its AI use stems from efficiency gains; the business model would, in theory, also be possible without AI, but requires the efficiency it provides to remain economically viable.

Several startups fall outside Shay and Davenport’s (2026) typology. They use their own AI models but do not offer these models on the market themselves. This sets them apart from the strongly technology-centred AI Originators and AI Explorers. Instead, they combine their own AI model architectures and training with domain-specific data to create new services. To extend the typology, one might introduce the term AI Combinators for this group.

Innovation and value creation beyond Big Tech

This observation may have implications for Europe’s AI sector as a whole. The race to develop the most powerful AI models currently appears to be playing out primarily in the United States and China. Meanwhile, access to the capabilities of these models is increasingly becoming a widely available resource, much like electricity or cloud computing. For many startups in Germany, economic growth appears to come from innovation and entirely new offerings rather than efficiency gains in traditional processes. The decisive factor may therefore be less who develops the most powerful AI model than who can put it to the most effective use. In the case of AI Enhancers, this means building a distinctive business model through a high degree of specialisation in their underlying data. For AI Combinators, developing their own model architectures and training approaches, with or without proprietary data, can provide a strong competitive position.

Challenges associated with AI use, such as de-skilling, job losses and shifts in power, cannot be explained conclusively by the extended typology. Nevertheless, AI Optimizers appear particularly likely to be associated with job losses and de-skilling. Their use of AI centres on reducing costs by optimising work processes, which typically also reduces the need for human labour. AI Combinators, by contrast, need highly skilled people to continue developing their data and models. Their model-agnostic approaches also make AI Enhancers and AI Combinators less dependent on dominant providers and may thus mitigate current shifts in power.

Although the study is not based on a representative sample, it suggests that the prospects of German and European startups may depend less on keeping pace in the global race to develop ever-larger AI models. More important may be their ability to combine existing AI technologies with specialised industry knowledge, exclusive access to data, a deep understanding of customers and the practical ability to put these elements to work.

References

Abonamah, A. A., Tariq, M. U., & Shilbayeh, S. (2021). On the Commoditization of Artificial Intelligence. Frontiers in Psychology, 12, 696346. https://doi.org/10.3389/fpsyg.2021.696346 

artificialanalysis.ai. (2026). Open Weights: Frontier Language Model Intelligence By Country, Over Time. https://artificialanalysis.ai/trends

De Micheli, B., Dente, G., Faioli, M., Smilari, A., Omersa, E., Patras, S., Ravaglia, M., Vancauwenbergh, S., Barslund, M., Tobback, I., & European Parliament (Hrsg.). (2025). Digitalisation, artificial intelligence and algorithmic management in the workplace: Shaping the future of work: cost of non-Europe. European Parliament. https://doi.org/10.2861/0136788 

Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4573321 

Demirer, M., Fradkin, A., Tadelis, N., & Peng, S. (2025). The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs (Nr. W34608; S. w34608). National Bureau of Economic Research. https://doi.org/10.3386/w34608 

Hui, X., Reshef, O., & Zhou, L. (2023). The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4527336 

Kalyani, A. (2026, Mai 18). Is Optimism for Artificial Intelligence Boosting Investment? – San Francisco Fed. https://www.frbsf.org/research-and-insights/publications/economic-letter/2026/05/is-optimism-for-artificial-intelligence-boosting-investment/ 

Rinta-Kahila, T., Penttinen, E., Salovaara, A., Soliman, W., & Ruissalo, J. (2023). The Vicious Circles of Skill Erosion: A Case Study of Cognitive Automation. Journal of the Association for Information Systems, 24(5), 1378–1412. https://doi.org/10.17705/1jais.00829 

Rubinton, H., & Patro, B. A. (2026). Tracking AI’s Contribution to GDP Growth. https://www.stlouisfed.org/on-the-economy/2026/jan/tracking-ai-contribution-gdp-growth 

Saxena, R., Gema, A. P., & Minervini, P. (2025). Lost in time: Clock and calendar understanding challenges in multimodal LLMs. arXiv preprint arXiv:2502.05092. 

Sha, S., Fattorini, L., Raymond, P., Yolanda, G., Vanessa, P., Santarlasci, L., Juan, P., Nestor, M., Russ, A., & Erik, B. (2026). The AI Index 2026 Annual Report. https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf 

Shay, J. P., & Davenport, T. H. (2026, Februar 25). Six Types of AI Startups, Explained. MIT Sloan Management Review. https://sloanreview.mit.edu/article/six-types-of-ai-startups-explained/ 

Stelter, D. (2025, Januar 27). Deutschland droht den Anschluss zu verlieren. Handelsblatt, (18), 12.