Building our AI platform in-house: €40 million and 18 months
Tech/AI
In eighteen months the model is a commodity: the moat is data and distribution.
In the scenario Recovered 12+ months of market lead and avoided spending on a part destined to become a commodity.
The protagonist & the context
The CTO of a B2B software company with an established client base, tasked with deciding whether to build AI in-house or rely on third-party providers. The board has already put «native AI» in the pitch deck for the next funding round.
A vertical software company with ten thousand enterprise clients, strong on process data in a specific domain. It has mid-level engineering resources and an already crowded product roadmap. The «AI project» started as an internal initiative by a small research group and has taken on a life of its own.
Background
The field of large language models and general-purpose AI systems has transformed from a barrier to entry into shared infrastructure in the space of twenty-four months. The main API providers (models, embeddings, retrieval) have reduced costs by an order of magnitude in two years and continue to iterate every three to six months. Companies that invested in proprietary model development in 2022–2023 found themselves competing against freely available foundations in 2024. In this context, real competitive advantage has shifted to those who accumulate exclusive data and build product integrations that are hard to replicate — not those who train the largest model.
The dilemma
The internal narrative is compelling: «The model is our advantage, we must own it.» But the right question is not «who owns the model?» — it is «what does the client pay for?». Client interviews show that perceived value lies in process data, integration speed, and vertical support — not in the underlying algorithm. Building the model in-house requires eighteen months and forty million euros; buying it requires three months and operating margin. In the fifteen-month gap, the market moves, competitors ship, and the technology you are building may already be available as an API at a fraction of the cost.
Exhibits
Estimated time to get an AI capability into production under each strategy — the gap is the window in which competitors can consolidate (illustrative data).
Share of enterprise clients citing each factor as their «main reason to renew» — the AI model itself is nearly irrelevant compared with data and integration (illustrative data).
The contradictor's analysis
- The advantage is in the model, not in the data or the ability to sell.
- The team delivers in 18 months.
- The technology won't shift under our feet in the meantime.
In a field that changes every three months, 18 months of in-house development means launching something already old. The model is becoming a commodity, something everyone has: the real defensible advantage (the «moat») is proprietary data and sales channels. Building the model burns your precious time on the wrong part.
- Will the same capability be buyable «off the shelf» (via API) within a few months?
- Timing compared: buy in 3 months versus build in 18.
- What do clients pay for: the model, or the product and the data around it?
- If in 12 months that technology belongs to everyone, what justifies 18 months building it?
- How much does the market we lose by waiting cost us?
Provenance: vendors' public roadmaps · client interviews · time-to-launch comparisons · red-team base.
Resolution & consequences
Methodological note
Composite teaching case. Figures and details are illustrative; it is not a client outcome or a performance claim. Built, in the method of the Harvard Business Review, from recurring decision patterns in the sector, merged and anonymized to protect confidentiality. The «provenance» notes describe the type of evidence the engine cites with traceability in production. No Δ-CSI score is shown: composite cases have no frozen, reproducible scorecard.