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

The development of GenAI: from experiment to infrastructure

In just two years, GenAI has shifted from “which chatbot should we choose?” to a strategic infrastructure question. Yet many organisations are still stuck on that first question: which model and which chat solution should we go with? Understandable, but it’s the least durable decision in the entire landscape. Models change within months. The question that holds up for years is how you structure models, knowledge sources, authorisations, and logging so that individual components remain replaceable. An AI strategy that focuses solely on today’s strongest vendor misses the point. The architecture that still allows room for adaptation three years from now is the true starting point.

In this article, we cover six themes that together define the new playing field: the shift from tool to infrastructure, the consolidation and fragmentation within the model layer, the split-cost market, the rise of open-weight models as a foundation for data sovereignty, the regulations that are now actively shaping the market, and the factor that carries the most weight in practice.

1. Infrastructure rather than tool

GenAI is no longer a standalone tool. The components themselves (models, platforms, and tooling) remain a matter of procurement. But the way they connect has become infrastructure. That doesn’t mean every organisation needs to design and build that infrastructure itself; that work can perfectly well be handed to a vendor or partner. What cannot be outsourced are the decisions beneath it.

Those decisions concern how models are linked to tasks, how knowledge sources are tied to processes, and how authorisations and logging are arranged. The benchmark is always the same: can a component be removed without having to rebuild the whole? Those who don’t answer these questions themselves will have them answered by their vendor, and only notice when they try to switch for the first time.

In practice, that decision comes down to one point: the gateway. This is the internal access layer behind which multiple models sit, where it is determined per task which model is called upon, where usage and costs are made visible, and where logging and filtering are managed in one place. Whoever owns that layer can switch models without touching the applications. Whoever places that layer with a vendor also transfers the ability to switch. Models are the shortest-lived component in this market; everything built around them must survive those changes.

This shifts the nature of vendor selection, though what matters differs by role. At the technology layer, it is less about who delivers the strongest performance today and more about which architecture still offers room to manoeuvre three years from now. Replaceability can be made concrete: prompts, evaluation sets (fixed test questions used to measure the quality of responses), the accumulated knowledge layer, and logging are transferable assets, provided they are technically and contractually exportable. The useful question to ask a vendor is therefore not whether the platform is open, but precisely what comes with you when you leave, in which format, and within what timeframe.

For implementation and management partners, a different criterion applies: the ability to drive adoption. Adoption can only be partly purchased, since the direction, ownership, and assessment of outcomes remain the work of the organisation itself. The partner who strengthens that rather than takes it over is the partner you’re looking for.

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 Chinese players such as DeepSeek and Qwen. Above that, the landscape is fragmenting. Platforms, intermediate layers connecting models to sources and tools, and software for making your own documents AI-searchable are emerging faster than they can prove themselves. We take it as a given that some of the tooling being chosen today will no longer exist in three years, or will have been acquired.

That’s not an argument for waiting, but it is an argument for deliberately choosing replaceable components and standard interfaces. An encouraging sign is that those interfaces are beginning to standardise. The Model Context Protocol has rapidly grown into the de facto standard for connecting tools and data sources to models, and is now supported by all major model providers. Sticking to such standards matters more than the choice of framework itself. Stability comes not from choosing one party, but from the way the architecture is constructed. That is an architectural principle, not a procurement decision.

3. The cost market is split in two

There are two stories about AI costs, and both are true. Model usage is billed in tokens, small pieces of text; the more a model reads and writes, the higher the bill. At the bottom of the market, prices are falling sharply: performance that cost tens of dollars per million tokens a few years ago is now available for a few cents. At the top, the opposite is happening, because advanced reasoning capability remains scarce and the latest generation of top models is becoming more expensive. Simple work is becoming nearly free; premium work is rising in price. The gap between the two determines which architectural choices make sense.

Then there is consumption. Total costs are rising for most organisations, because usage 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 that gets stuck in a loop is a cost incident, not a system failure. Hard limits per task are therefore a basic control measure. The expectation that AI will automatically become cheaper is correct per unit, but not on the total bill.

The most useful unit of measurement, in our view, is not the price per million tokens but the cost per completed task, set against the time it would take to do it manually. Organisations that work this out early are able 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 can therefore be run in your own environment. On complex reasoning tasks, they still lag behind the closed top models. But for the majority of everyday work such as summarising, rewriting, searching your own documents, and structured information retrieval, the practical difference has become small.

For Dutch-language use, there is a nuance that is often overlooked. The bottleneck is rarely in text generation, but more often in embeddings. This is the technique that converts documents into a form the AI can search. It is precisely in legal, financial, and administrative language use where quality differences are most significant.

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;
  • management of the encryption keys;
  • the nationality of the managing personnel;
  • ownership of the technology;
  • the jurisdiction under which the vendor operates.

Parties that call themselves sovereign score very differently across those dimensions. It is worth asking the question per dimension.

The honest flip side is that sovereignty moves 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, along with the associated expertise and costs. That calls for a deliberate trade-off rather than a default answer.

5. Regulation: from precondition to reality

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

The most significant change this year is the Digital Omnibus on AI, which came into force at the end of July 2026. This is a European revision act that adjusts parts of the AI Act. The reason was practical: implementation stalled because the standards, guidelines, and supervisory authorities needed for compliance simply didn’t exist yet. 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 devices, the date is August 2028.

However, there is no across-the-board relaxation. Transparency obligations have not been postponed and have applied since August 2026, additional prohibited applications have been introduced, and European supervision has been strengthened. In the Netherlands, supervision is taking shape through the AI Act Implementation Act, with a central role for the Dutch Data Protection Authority.

The significance of the delay is therefore smaller than the coverage suggests. The requirements have been shifted, not dropped, and the underlying infrastructure (registration, logging, evaluation, and demonstrable human oversight) takes more time to put in place than the remaining period might suggest. Organisations now thinking about their AI architecture would do well to treat compliance as an ongoing part of the design. Those who build it in from the start will avoid costly adjustments later.

6. The real brake is not in the technology

One observation colours our market view most strongly: the brake is rarely in the technology. In virtually every organisation, the state of information management, the accuracy of authorisations, data quality, and employee adoption are more decisive for outcomes than the choice of model. An organisation with that foundation in order can work with almost any architecture. An organisation without it won’t solve the problem with any model.

This makes the order of operations important. We see in practice that organisations that first clean up their authorisation model and document management achieve results with AI faster than organisations that start with a more advanced model. GenAI is developing at a pace that demands strategic thinking. Not just about technology, but about governance, costs, and architecture. The organisations that take this seriously now are building a foundation that won’t get in the way tomorrow.

Getting started with a manageable AI environment

We are happy to think alongside you in translating strategy into a future-proof and manageable AI environment. Depending on where you currently stand, there are three logical starting points:

  • Want to first understand what GenAI concretely delivers in your processes? The AI Kickstart produces a working proof of concept on a use case from your own organisation in five days, including security and compliance setup and strategic advice on the next steps.
  • Want to enable employees to work safely with the latest models? Secure Chat is a fully managed, vendor-independent AI platform in an isolated European environment, with Microsoft 365 integration, audit logging, and the option for local models.
  • Want to keep the gateway and the underlying AI platform in your own hands? With Managed Foundry, we manage the Azure AI Foundry infrastructure, including automated compliance enforcement, customer-managed encryption keys, and private endpoints within the EEA.


Get in touch if you’d like to discuss the architectural choices that suit your situation.

Want to know what Solvinity can do for you?

Contact us and find out today.

Also read

More articles

Meer