Don’t build your AI product on rented ground

English · Nederlands

An old paper road map on a dark desk, its roads turning into glowing blue network lines that run into a laptop, with a padlock and a property tag attached to the map.

If your AI product runs on someone else’s model, you are in the position app builders were in when they ran on someone else’s map. The foundation is expensive, strategic and owned by a few companies, some of which also compete with you. In 2011 I wrote about what the Nokia and Microsoft deal meant for mobile maps. Reading it again in 2026, most of my predictions missed, but the question underneath has become the central one in AI.

What I saw then

In February 2011 Nokia and Microsoft announced their smartphone partnership, and NAVTEQ, Nokia’s map business, was named as a key asset. My argument: digital maps are not free. Building a global map database costs so much that a handful of companies supplied almost everyone, and Google was working hard to become less dependent on them.

I expected three things. Maps would get much cheaper for Microsoft. Free turn-by-turn navigation would come to Windows phones. And Microsoft and Nokia would offer developers maps and points-of-interest data as a neutral platform, because apps like foursquare were stuck on the maps of Google, a competitor. Most of that did not happen the way I thought. The point about ownership did.

Two fenced plots with For Rent signs: on the left an old Nokia phone showing a map, under the Nokia logo; on the right a laptop with an AI chat, under the Microsoft logo. Landlord signs above them read OpenAI and Anthropic.
Landlords then and now: maps in 2011, foundation models today.

What is the same now

Swap “map” for “foundation model” and the structure is familiar. Training a frontier model costs so much that a few companies supply most of the market. Thousands of products are built on top through APIs. And several suppliers also sell their own assistants, so the platform is also a competitor.

Two things the map market taught us are already repeating.

The owner of the asset sets the price, once you depend on it. In July 2018 Google overhauled its Maps pricing. Where developers had 25,000 free map loads per day, they now got a $200 monthly credit, worth about 28,000 loads per month, and the price per 1,000 extra loads went from $0.50 to $7 (Maps Marker). Many apps had been built on the assumption that maps were practically free. Model prices are falling today. That says nothing about what happens once a provider knows you cannot leave.

Access can change overnight. In June 2025, days after reports that OpenAI would buy the coding tool Windsurf, Anthropic cut Windsurf’s direct access to several Claude models with less than a week’s notice. Anthropic’s co-founder said it would be odd to sell Claude to OpenAI (The Decoder). It is the foursquare-on-Google-Maps problem from my post, played out in a week.

The biggest buyers draw the obvious conclusion. Microsoft, OpenAI’s closest partner, added Anthropic’s models to Microsoft 365 Copilot in September 2025, alongside OpenAI’s (Pure AI). If Microsoft wants a second source, so should you.

What is different

The analogy is useful, not exact.

There are more suppliers. In 2011 a handful of companies had a global map. Today several labs offer capable models, and open-weight models you can run yourself are good enough for many tasks. Multi-sourcing is a realistic option, which it barely was for maps.

Models are less interchangeable than map tiles. A street is a street in any database. Two models given the same prompt behave differently, so your switching costs hide in prompts, evaluations and the behaviour your customers have grown used to. A switch that looks like a week of work on paper can easily take a quarter.

And your own assets matter more. In maps, the base layer was most of the value. In AI products the workflow, the customer data and the system of record around the model often carry more value than the model itself. That is good news, if you invest in them.

What to do about it

  1. Map your dependency. For each AI feature, write down which model it uses, what happens if the price doubles, the terms change or access stops next month, and how much revenue is attached.
  2. Build for switching. Route model calls through one internal layer and keep an evaluation set for each critical feature, so you can test a different model in days rather than months.
  3. Keep a second source ready. For revenue-critical features, test and maintain at least one alternative provider, including an open-weight option for sensitive data or high volumes.
  4. Protect your own price. Do not pass a supplier’s rate card straight to customers. Price on your value unit, and keep contract room for supplier price changes.
  5. Invest in what the model provider does not own. Workflow, data, integrations and audit trail. That is what keeps customers with you when the model underneath changes.

Rent the engine, own the road

In 2011 I wrote that maps are expensive strategic assets, not free utilities. Foundation models are the same, at a much larger scale. Renting one is fine; most companies should. Just make sure that if the landlord raises the rent or changes the locks, your customers stay with you.

Sources