AI providers are in an odd position. Their core product is becoming a commodity, and yet they're locked in one of the most expensive races in the history of technology. Both things are true at the same time, and the way providers are responding says a lot about where the real moat in enterprise AI is forming.
The commoditization paradox
In some respects, models have become easy to build. Open weights models are widely available, training techniques are well documented, and capable models arrive from new labs every few months. For many everyday tasks, several models are good enough, and the differences between them are hard to notice.
At the same time, the frontier keeps moving. The top models still pull ahead on benchmarks, and each new release resets expectations. Being the best matters, but being the best doesn't last.
That creates a paradox for providers. They have to spend heavily to stay at the frontier, knowing that any lead is temporary and that the floor beneath them keeps rising. If performance alone won't keep customers, something else has to. For enterprise customers, that something is the organization itself.
Lock-in at the C-level
The enterprise strategy is straightforward: win the CTO. If the CTO mandates one platform for the whole organization, the provider doesn't need to win every engineer, every team, or every use case. It wins them all at once.
Anthropic and Claude are the clearest example. The goal isn't just to have the best coding model this quarter. It's to have the entire organization running on Claude, with workflows, tooling, training, and habits built around it. Once that alignment is in place, the company tends to stay, even if a rival eventually releases a better model.
To be fair, Anthropic has often earned its position. Claude has posted top benchmark scores, particularly for coding, for much of the recent past. But that won't always be true. No provider stays on top forever, and a smart provider plans for the day it isn't.
Model dexterity is the ideal
From the company's point of view, the better position is model dexterity: the ability to switch models contextually as new ones arrive. One provider might lead in coding, another in reasoning, and an open weights model might be good enough and far cheaper for routine work.
Benchmark leadership rotates at a breakneck pace. A model that leads in the spring can be overtaken by summer. Companies that can move with that rotation get the best tool for each job. Companies that can't are stuck with whatever their platform offers until the next contract renewal.
AI has become tribal
There's also a cultural side to this. AI has become surprisingly tribal. People have favorites, and they defend them.
An individual might personally prefer Grok or ChatGPT because they believe it's the best product. Then they go to work and use Claude, because that's what the organization is aligned on. Their personal judgment and their professional tools point in different directions.
That alignment has real benefits. It creates consistency, simplifies security reviews, and makes it easier to share prompts, workflows, and best practices across teams. But it comes at a cost to flexibility. The organization gains coherence and gives up some ability to use the best tool available.
Token commitments as a buffer
Providers understand that they won't always win on performance alone, so they build a buffer. Token commitments and annual spend agreements are part of that buffer. They create a financial reason to stay put even when a competitor looks attractive.
But the contract isn't the real protection. The real protection is the stickiness that comes from a top-down mandate. When an entire organization has standardized on one platform, the cost of switching isn't just a line in a contract. It's retraining people, rebuilding workflows, re-running security reviews, and convincing leadership that the change is worth the disruption. The commitment locks in the spend. The mandate locks in the organization.
Top-down versus bottom-up
There's a real difference between a CTO saying "everyone must use this platform" and a bottom-up approach where people use the models and tools they prefer, as long as those tools meet organizational standards.
In a bottom-up model, the organization sets the guardrails rather than picking the vendor. Any tool is acceptable if it meets requirements such as SOC 2 compliance and zero data retention. Within those guardrails, teams choose what works best for them.
Bottom-up is harder to manage. It requires clear standards, good visibility, and consistent enforcement. But it's the better direction, because it lets the organization benefit from whichever model is best at any given moment instead of whichever model won the last procurement cycle.
An exaggerated example
To make the risk concrete, consider a deliberately exaggerated scenario. Imagine everyone agrees that Claude is the superior model for coding, but a company mandate requires engineers to use Google Gemini because that's the provider the organization is aligned with. The engineers know there's a better tool available. They just aren't allowed to use it.
In practice, most organizations that standardize on a coding model choose Claude, because it has had the strongest model much of the time. So the mandate and the best choice often line up. But the Gemini example shows what happens when they don't. A provider-aligned mandate can lock a company into a subpar outcome, and the people closest to the work are the first to feel it.
The real enterprise moat
In the enterprise, the moat isn't the model. It's the organizational mandate. Providers know their models will be matched and eventually surpassed, so they invest in becoming the default for the entire company, reinforced by spend commitments and top-down decisions.
That strategy makes sense for the provider. The open question is whether it makes sense for the company. The tension is between a provider's interest in lock-in and an organization's interest in always using the best tool for the job. The companies that come out ahead will be the ones that set clear standards, stay model-dexterous, and make sure their AI strategy is driven by what works best rather than by who they signed with.
No comments yet