Why no AI provider holds the lead, and what to do about it

Three years ago the decision looked simple: OpenAI had the best model and half the enterprise spend. Today another provider has it, and in eighteen months it could be a third. This article brings together the published data on how leadership in AI models has rotated, what companies are doing about it, and what it means for a Spanish company of 100 to 1,000 employees about to buy a platform for the next several years.

Leadership has changed hands twice in three years

Menlo Ventures measures every year which providers capture enterprise spend on language models in the United States. The trend is clear:

Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise, a survey of 495 enterprise AI decision-makers in the United States, November 2025. It's US data; there's no Spanish equivalent.

There isn't much to interpret. The provider that dominated in 2023 now has just over a quarter of the spend, and the one that was marginal is now first. Nobody knows what the 2027 table will look like, and that's exactly the point: any company that locks itself to one provider today is betting the table won't turn again.

Companies already use several models

The market's answer has been not to choose. According to ICONIQ, companies use 3.3 different models on average (July 2026), up from 2.8 a year earlier. An a16z survey of CIOs (January 2026) puts at 81% the share of companies using three or more model families, up from 68% the year before. And LangChain finds that more than three-quarters of teams putting agents into production use several models and assign them by complexity, cost and latency (December 2025).

The reason isn't ideological. One model writes better in Spanish, another reasons better over long contracts, a third is ten times cheaper for high-volume repetitive queries. And the ranking shifts with every new version: what's best for a task today may not be in three months.

Sources: ICONIQ, 2026 State of AI Report, July 2026; a16z, CIO survey, January 30, 2026; LangChain, State of Agent Engineering, n=1,340, December 2025.

Switching models is easy in theory and rare in practice

Here's the nuance that turns the above into a purchasing problem. Menlo measures that only 11% of companies switched model providers in the last year, despite quality and price moving a lot. And a16z documents why: switching cost rises once the work is agentic. An agent that runs a multi-step process is tuned for a specific model; changing the model forces a review of every step and everything that depends on it.

The consequence is uncomfortable: the more value AI produces at a company, the more expensive it becomes to stop depending on the provider it was built with. That's the definition of lock-in. And the time to avoid it is before you build, not after.

What this means for a company about to buy

Three things, in order of importance.

1. Choosing the model must be configuration, not a project. If switching the model an agent uses requires rebuilding it, the company isn't really choosing. Ask any provider how much it costs to switch the model on an agent already in use.
2. Agents and knowledge have to belong to the company, not the model. Instructions, connected sources, permissions and history have to outlive the model they were built with. It's the only way the market's rotation doesn't force you to start from zero.
3. Independence is insurance whose price keeps rising. Today it's cheap: most companies haven't built anything yet that depends on one model. In two years, with dozens of agents in use, it will be expensive. Like any insurance, you buy it before the loss, not after.

The counterargument, so you have it

There's a serious position on the other side, and it's worth knowing. Ben Thompson (Stratechery, March 2026) argues that profit flees the modular parts of a value chain toward the integrated ones, and that agent differentiation will sit in the integration between the model and the environment around it — meaning the labs themselves will be the point of integration, and an independent layer will be the part that gets commoditized. It's a respectable argument. Our answer is that for a 200-person Spanish company, the integration that matters isn't the one between a model and its lab, it's the one between AI and the company's own systems, processes and permissions, and no lab is going to build that custom for them. You can judge for yourself which reading fits your case better.

What we don't claim

To be precise about what the data actually says. We don't claim open-weight models are going to replace closed ones in the enterprise: the share of workloads on open-weight models fell from 19% in 2024 to 11% by late 2025, per the same Menlo survey, and a mid-sized Spanish company isn't going to host models on its own. Nor do we claim frontier prices keep falling without limit: what gets cheaper is yesterday's capability, and each moment's most advanced models have gotten more expensive. What we do claim is more modest and more solid: no provider holds the lead long enough to make marrying it worthwhile, and switching costs rise over time.

How Azimut handles it

Azimut is model-independent by design: its own layer with one module per provider, with no aggregator in between, on top of which the user chooses the model per task from OpenAI, Anthropic, Google and Mistral. Agents, connected sources, permissions and usage belong to the company and don't depend on the model they were built with. Switching a model is configuration. It's the insurance this article talks about, included in the price of the seat.

Frequently asked questions

Isn't it better to wait for a provider to win?

Three years of data show no sign that's about to happen. Waiting has a cost: in the meantime, each team buys on its own and what it builds doesn't get shared.

Doesn't using several models complicate things for employees?

No, if the platform does it for them. The user chooses the agent; whoever created the agent configured the model. For most of the workforce, the model choice is invisible.

What happens to the model an agent uses if the provider changes its price?

The agent's model gets changed in configuration, without rebuilding the agent. The usage panel shows the effect on spend.