Izak-Jan van den Nieuwendijk Portfolio Manager Public Cloud
28 August 2026

The development of GenAI: from experiment to infrastructure

GenAI has shifted in two years from “which chatbot do we choose?” to a strategic infrastructure question. Yet many organisations remain stuck on that first question: which model and which chat solution is the right choice? Understandable, but it is the least durable decision in the entire landscape. Models change within months. The question that holds its value for years is how to configure models, knowledge sources, authorisations, and logging so that components remain interchangeable. An AI strategy that focuses solely on today’s strongest vendor misses the point. The architecture that still offers room for adjustment three years from now is the important starting point.

This article addresses five themes that together define the new playing field: the shift from tool to infrastructure, the consolidation and fragmentation in the model layer, the split cost market, the rise of open-weight models as a basis for data sovereignty, and the regulation that now actively shapes the market.

1. Infrastructure rather than tool

GenAI is no longer a standalone tool. The components themselves, models, platforms, and tooling, remain procurement items. But the coherence between them has become infrastructure. This does not mean every organisation must design and build that infrastructure independently; that work fits well with a vendor or partner. What cannot be outsourced are the underlying decisions. Which models suit which tasks? Which knowledge sources connect to which processes? How are authorisations and logging arranged so that components are replaceable without rebuilding the whole? Organisations that do not answer these questions themselves receive the answers from their vendor, and only notice this when they attempt to switch for the first time.

In practice, this decision converges on a single point: the gateway. This is the internal access layer behind which multiple models sit, where per task it is determined which model is called, where consumption and costs are made visible, and where logging and filtering are managed in one place. Whoever holds this layer independently is able to switch models without affecting the applications. Whoever places this layer with a vendor also transfers the ability to switch. In this market, models are the shortest-lived component; everything built around them must survive those changes.

This alters the nature of vendor selection, though what matters differs by role. At the technology layers, the focus shifts away from who delivers the strongest performance today towards which architecture still offers flexibility three years from now. Replaceability is tangible here: the prompts, the evaluation sets (fixed test questions used to measure the quality of responses), the accumulated knowledge layer, and the logging are transferable assets, provided they are technically and contractually exportable. The practical question to put to a vendor is therefore not whether the platform is open, but what exactly transfers upon departure, in which format, and within what timeframe.

For implementation and management partners, a different criterion applies: the ability to drive adoption. Adoption is only partially purchasable, because oversight, ownership, and the assessment of outcomes remain the work of the organisation itself. The partner that reinforces this rather than takes it over is the partner worth seeking.

2. The model layer consolidates; the layer above it fragments

The model layer is consolidating around a limited number of major providers: OpenAI, Anthropic, Google, Meta, Mistral, and now also Chinese parties such as DeepSeek and Qwen. The refresh rate there is measured in months, not years. Above this layer, the landscape fragments. Platforms, intermediate layers connecting models to sources and tools, and software for making internal documents searchable by AI appear faster than they prove themselves. The realistic assumption is that a portion of the tooling chosen today will no longer exist in three years, or will have been acquired.

This is not an argument for waiting, but for deliberately choosing replaceable components and widely adopted interfaces. An encouraging sign is that these interfaces are beginning to standardise: the Model Context Protocol has rapidly become the de facto standard for connecting tools and data sources to models. Adhering to such standards is now more important than the choice of framework itself. Stability lies not in choosing a single vendor, but in how the architecture is constructed. That is an architectural principle, not a procurement decision.

3. The cost market is split in two

Two narratives exist about AI costs, and both are true. Model usage is billed in tokens, small fragments of text; the more a model reads and writes, the higher the cost. At the lower end of the market, prices are falling sharply: performance that cost tens of dollars per million tokens a few years ago is now available for cents. At the upper end, the opposite is happening: the latest generation of top models is becoming more expensive, because advanced reasoning capability remains scarce. Simple work becomes nearly free; top-tier work rises in price. The gap between the two determines which architectural choices are sound.

Usage compounds this. Total costs are rising at most organisations, because consumption is increasing and because AI agents, applications that independently execute multiple steps in sequence, consume many times more tokens per task than a single chat query. An agent caught in a loop is a cost incident, not a system failure; hard limits per task are therefore a management measure, not a luxury. The expectation that AI automatically becomes cheaper holds per unit, but not for the total bill.

The most useful unit of measurement is not the price per million tokens, but the cost per completed task, set against the time that task would take manually. Organisations that calculate this early are positioned to make targeted decisions about which applications are cost-effective and which are not. Without that insight, the AI budget grows with enthusiasm rather than with the business case.

4. Open-weight models and data sovereignty

The development that has changed most in character over the past two years is that of open-weight models: models that are freely available and therefore run within a private environment. On complex reasoning tasks they still trail the closed top models. But for the majority of day-to-day work, such as summarising, rewriting, searching through internal documents, and structured information retrieval, the difference has become small in practice. For Dutch-language use, a nuance is frequently overlooked: the bottleneck rarely lies in text generation, but more often in embeddings, the technique that converts documents into a form the AI uses to search through them. Particularly in legal, financial, and administrative language, the quality differences there are considerable.

This makes hosting in a Dutch or European environment a realistic option rather than a principled exception. But sovereignty is not a quality mark; it is a sliding scale with dimensions that move independently of one another: the location of the data, key management, the nationality of the managing personnel, ownership of the technology, and the jurisdiction under which the vendor operates. Parties that describe themselves as sovereign score very differently across these dimensions. It is worthwhile to examine the question per dimension.

The honest counterside is that sovereignty shifts work rather than eliminates it: availability, model updates, GPU capacity, and safety filters then fall to the organisation itself, or to the management partner fulfilling that role, with the associated knowledge and costs. That calls for a deliberate assessment, not a standard answer.

5. Regulation: from precondition to reality

Regulation is now a market-shaping force, not an afterthought, though the pace has been adjusted this year. The GDPR remains the leading framework. The AI Act has been introduced in stages since 2024: the prohibition on certain AI practices and the obligation to promote AI literacy among staff have applied since February 2025; the rules for providers of large, general-purpose AI models since August 2025.

The most important change this year is the Digital Omnibus on AI, which entered into force at the end of July 2026. This is a European revision act that amends parts of the AI Act. The motivation was practical: implementation stalled because the standards, guidelines, and supervisory bodies against which organisations were to demonstrate compliance simply did not yet exist. The Omnibus therefore grants more time. The obligations for high-risk systems, such as AI used in recruitment and selection or in decisions affecting individuals, now apply from 2 December 2027. For AI embedded in regulated products, such as medical equipment, that date is August 2028. This is not, however, a broad relaxation: the transparency obligations have not been delayed and apply from August 2026, additional prohibited applications have been introduced, and European oversight has been strengthened. In the Netherlands, supervision is taking shape through the Uitvoeringswet AI-verordening (AI Act Implementation Act), with a central role for the Dutch Data Protection Authority (Autoriteit Persoonsgegevens).

The significance of the delay is therefore smaller than the reporting around it suggests. The requirements have shifted, not been removed, and the underlying arrangements, registration, logging, evaluation, and demonstrable human oversight, take more time to put in place than the remaining period implies. Organisations currently thinking about their AI architecture would do well not to treat compliance as a final destination, but as an ongoing part of the design. Addressing this from the outset prevents costly adjustments later.

More information

One observation shapes the market view most: the constraint rarely lies in the technology. At virtually every organisation, the state of information management, the accuracy of authorisations, data quality, and staff adoption are more decisive for the outcome than the choice of model. An organisation with that foundation in order works effectively with almost any architecture. An organisation without it resolves nothing with any model.

GenAI is developing at a pace that compels organisations to think strategically, not only about technology, but also about governance, costs, and architecture. Organisations that take this seriously now are building a foundation that does not become an obstacle tomorrow.

Solvinity is available to support the translation of strategy into a future-proof and manageable AI environment. Get in touch to discuss what is achievable for your organisation.

Want to know what Solvinity can do for you?

Contact us and find out today.

Also read

More articles

Meer