Will one LLM rule them all?

A few years ago, the story of enterprise AI sounded almost suspiciously clean: a handful of frontier labs would build extremely powerful foundation models, companies would connect to them through APIs, everyone would add AI to their products, processes, dashboards, search bars, internal tools, customer service flows, and probably a few places where absolutely nobody had asked for it, and the winners would be those who had access to the biggest models, the largest compute clusters, and the most impressive benchmark screenshots.

It was a very attractive story, especially if you were raising billions of dollars, because it suggested that the model itself would become the moat: not the workflow, not the product experience, not the painful operational work of making AI useful inside a real company, but the model, sitting somewhere behind an API, quietly becoming the scarce layer everyone else had to rent.

For a while, this felt believable, because the gap between frontier models and everything else was large enough to make most alternatives look unserious, and most companies had neither the talent, the infrastructure, nor the appetite to do anything more ambitious than plug into whichever provider had the best demo that month. If you worked in enterprise AI during that period, you probably remember the mood: a mix of genuine excitement, executive urgency, vendor theatre, and many conversations that started with “Can we build something like ChatGPT, but for our company?” and ended with someone discovering document permissions.

But open-weight models have been getting better, and not in a cute hobby-project way, but in a way that is starting to change the economic and architectural assumptions behind enterprise AI. If good-enough models can be downloaded, adapted, hosted, evaluated, swapped, and combined with internal data and decent engineering, then the idea that a few closed models will permanently own the entire intelligence layer becomes much harder to defend, or at least much harder to treat as the only serious future.

This does not mean closed frontier models are suddenly irrelevant, because they are clearly not, and in many cases they are still the best option when quality, reasoning ability, multimodal capability, or speed of experimentation matters more than cost or control. But it does mean that companies should be much more careful about designing their AI strategy as if one provider, one model family, one cloud ecosystem, or one pricing model will remain the obvious answer forever.

In real enterprise environments, most use cases do not need magic; they need something much more boring and therefore much harder to sell on a keynote stage. They need systems that answer from the right context, respect access control, behave consistently, can be evaluated against real business outcomes, protect sensitive data, fit into existing architecture, survive procurement and security reviews, and still make financial sense once the innovation budget disappears and the project has to live in the normal IT landscape, where enthusiasm goes to meet maintenance costs.

That is why I think the serious enterprise AI question is not “Which model should we use?”, but “How do we build systems that can move when the model landscape changes?” Use frontier APIs where their quality clearly creates value, use open-weight models where they are good enough, cheaper, easier to govern, or better suited for sensitive workloads, and build evaluation pipelines that allow teams to compare options based on evidence rather than hype, fear, vendor dinners, or a leaderboard screenshot that has already aged badly by the time it reaches the steering committee.