What an AI deployment leaves you with

The first sign of the problem is rarely technical. It is a number on an invoice that keeps on getting bigger, a model that’s been deprecated causing operational disruption, or a change in acceptable use terms that made a business’s use case no longer compliant with the vendor’s policies. 

Over 88% of businesses now use generative AI in at least one function. Yet, should they stop paying, change providers, or lose access to a model version, much of what has been built would be lost.

Over the past few months, it’s AI bills that have become the story with large, well-run companies discovering that the cost of heavy AI use has outstripped the value it returns. 

Some would argue that the issue is how teams are using these tools: too much experimentation, too little discipline. But what if it was the lack of model ownership? 

Just like the difference between renting a house and owning one, the problem is not the monthly payment, it’s what you’re left with when the payments stop. 

A house you’re renting isn’t yours, nor is the model you’re using.

Most people intuitively understand what renting in real estate terms entails. You pay every month and the property is maintained. On the other hand, the asset appreciates in someone else's name, if you stop paying you’re out, and if the landlord changes the terms or decides to paint the property in neon pink, you accept or leave.

AI, in its current dominant form, works the same way and it helps to split the system into two to see how.

The base model is a commodity. It’s powerful and improving fast, but also broadly interchangeable, just like a rented property. 

The learning layer is the mechanism that turns data into domain expertise. In our real estate analogy, that would be equivalent to structural improvements made to a property, like a loft conversion or kitchen extension, that will meanigfully increase the value of a property. 

While a tenant would be unlikely to fund a loft conversion on a property they don’t own, in today's standard AI deployments, investing heavily to build a business on top of a rented AI model is considered standard practice.

Have you ever seen your rent go down?

Renting has a cost shape that generally goes upward and, with AI, it’s now stopped being a quiet line item. 

Uber's chief technology officer reported that the company had burned through its entire 2026 AI coding budget in roughly four months. Around the same time, Microsoft began pulling engineers in one division off the AI coding assistant they had adopted, moving them to cheaper internal tooling. Goldman Sachs has since put a figure on the trajectory, estimating that the move to AI agents could drive a 24-fold rise in token demand by 2030.

The economics of AI have become complex and volatile. As AI moves from pilot to production, and adoption spreads across teams, bills rise while that spending never turns the model into an asset. 

What if you woke up one morning and the house you live in was now neon pink?

Model providers, like landlords, set the terms and it’s up to the renter to accept them or not. 

In practical terms, this can translate into model deprecation with businesses needing to re-test, re-tune and re-validate everything. It can also translate into cessation of service in the event that the provider revisits its acceptable-use policy, and a use case that was compliant before is deemed no longer acceptable. 

The best illustration of the latter is not a real estate analogy but when the U.S. government issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national. Unable to verify nationality in real time, Anthropic disabled both models for every user worldwide, including its own US customers who the order hadn't targeted at all. Access vanished overnight through no decision users had made. 

While this case refers to a government decision rather than a provider's, the key takeaway remains that when users rely on a model they don't own, access can be pulled by a third party overnight. 

The illusion of ownership. 

There are two misconceptions that are worth addressing because both give an unjustified sense of ownership. 

First is self-hosting the model. There are good reasons to self-host a model: data privacy, lower breach exposure and overall better control over security. But ownership isn’t one of them. Businesses may control where it runs, but not what it is, and that leaves the dependency exactly where it was.

The second is going open. While open-source models admittedly do offer flexibility and freedom from lock-in, an open model changes the infrastructure without touching the learning layer. 

In both cases, the test is the same: is the institutional knowledge these systems accumulate an asset that belongs to the corporation, or is it part of a relationship it cannot control? 

The landlord club.

The case for buying a property is obvious to most. While renting can be easier and more convenient, the money spent does not build equity.

The same logic applies to AI models. Owning a model and the mechanics that train it means that the large investments businesses make in AI go towards an asset they own. This asset learns from their data, is a size that is fit for purpose, calibrated to their use case, and therefore consumes only what it needs to. And more importantly, it is one that they fully control and that can’t be taken away. 

The economics follow directly. A domain-specific model runs leaner, can learn live in production with the right technology and, as such, remains current without retraining cycles. The result is training and inference cost down.

‍