De Infrastructuurrekening: Waarom de Echte AI-Kosten Pas Nu Zichtbaar Worden

The pilots are over. The proofs-of-concept have been approved, the board presentations have landed, and somewhere between the tenth and the hundredth use case, something shifted: AI stopped being a line item in the innovation budget and became infrastructure. That transition carries a price tag that most European boardrooms budgeted for incorrectly — not because the numbers were hidden, but because the pilots were deliberately small enough to make the real economics invisible. Now that scale is arriving, so is the invoice.

This is not a story about AI being expensive. It is a story about AI costs being structurally misunderstood at the moment when scaling decisions are being locked in. The organisations that correct that misunderstanding now will make rational capital allocation decisions. Those that do not will discover the gap in 2027, when reversing course is both technically and contractually painful.

The Q1 2026 earnings season provided the first unambiguous signal that enterprise AI has crossed from experimental to infrastructural demand. Amazon Web Services reported GPU capacity constraints across multiple regions, with reserved instance lead times extending beyond what enterprise procurement cycles can comfortably absorb. Microsoft Azure flagged equivalent pressure in Europe specifically, accelerating datacenter investment in the Netherlands and Poland to close a gap that is already affecting enterprise customers. Google Cloud acknowledged constrained TPU cluster availability without offering a timeline for relief. Three hyperscalers, one earnings cycle, the same message: demand has outrun supply, and the queue is not short.

The capacity signal matters because it reframes the cost conversation. When infrastructure is constrained, pricing power shifts to suppliers. Reserved instance discounts — the mechanism by which large enterprises have historically negotiated cloud costs down by 30 to 40 percent — become less negotiable when the underlying capacity is genuinely scarce. Dutch and broader European enterprises entering multi-year AI infrastructure commitments in this environment are doing so with less leverage than they would have had twelve months ago, and likely less than they will have in twenty-four months when new datacenter capacity comes online. Timing, in other words, is a cost variable.

The deeper problem is structural: the standard AI contract — seat licences, per-API-call pricing — captures only the visible fraction of total cost. For any organisation running retrieval-augmented generation at scale, the real cost architecture has at least four additional layers. Token pricing at inference compounds rapidly; a legal or financial services operation processing thousands of documents daily through a RAG pipeline can accumulate token costs of tens of thousands of euros per month on top of the platform licence — costs that were absent from pilot-phase calculations because document volumes were artificially constrained. Vector database storage for large document archives carries its own managed infrastructure cost that consistently surprises finance teams at first invoice. Fine-tuning and periodic retraining on domain-specific data — the step that moves a generic model toward genuine competitive differentiation — runs on GPU hours priced significantly above inference, and the training cycle does not end at deployment. And observability tooling, the monitoring infrastructure required to maintain output quality and safety in production, represents a recurring operational cost that appears in no pilot budget and is non-negotiable in any serious production system.

Business Implications

For CFOs finalising H2 2026 budgets, the practical correction is straightforward: AI cost modelling must include five components, not one. Licences and API consumption are the starting point. Inference infrastructure, vector storage, and database costs are the second layer. Labour — ML engineers with production system experience, a profile commanding a 20 to 30 percent market premium in the Netherlands — is the third. Observability and quality monitoring tooling is the fourth. The fifth, consistently omitted, is a redevelopment reserve: models age, business requirements shift, and the cost of keeping a production AI system current is not zero. For mid-sized Dutch organisations scaling seriously, total annual AI costs will run two to four times the initially contracted licence value. Any multi-year financial plan that does not reflect this range is not a plan — it is an aspiration.

For CTOs and CIOs, the more urgent issue is lock-in. RAG infrastructure, fine-tuned model weights, and vector databases built on a single cloud platform create migration costs that were not part of the original build decision. Organisations that treated cloud platform selection as a procurement question — evaluated on per-unit pricing — are discovering it was an architectural decision with multi-year financial consequences. The technical cost of migrating a production RAG system to a competing platform is substantial; the contractual cost, in a constrained-capacity environment where hyperscalers hold leverage, is higher still. IT procurement in 2026 requires architectural and commercial expertise that did not exist as a required competency two years ago. Boards that have not yet elevated this capability inside their technology leadership are exposed.

The winners in this environment are organisations that built cost modelling discipline into their AI governance from the pilot phase — and the vendors, consultants, and FinOps specialists who can help those that did not catch up quickly.

ZeroForce Perspective

The Zero Human Company thesis rests on a specific assumption: that AI-driven automation generates returns sufficient to justify continuous reinvestment in capability. That assumption breaks if the cost structure of AI at scale is misunderstood at the point of commitment. What the current infrastructure squeeze reveals is that the path to a genuinely autonomous enterprise runs directly through financial literacy about AI economics — not enthusiasm about AI capabilities. The organisations that will reach meaningful automation at scale are not necessarily those with the most sophisticated models. They are those whose CFOs understand token pricing, whose CTOs price lock-in into platform decisions, and whose boards treat AI infrastructure as a capital allocation question rather than a technology experiment. The boardroom conversation has changed. The budget category needs to change with it.