AI NewsModels & agentsAnnouncement
Databricks: enterprise single-vendor AI use fell from 64 to 28 percent
Databricks published its first AI Index. It shows enterprise AI use shifting from one vendor to many. Open-weight models took four times more token traffic in four months.

Image: Databricks
Why it mattersA team planning for one AI vendor is planning for how enterprises worked in July. The real pattern now is many models in one product, swapped as prices move.
A buyer picks one AI model for a product and builds around it. Then the next cheaper or smarter model comes out. The team has to decide whether to switch. Most of the discussion about that decision runs on guesses, because nobody publishes what other companies are actually doing. Databricks has now published a number.
The company released its first AI Index on 16 November. It draws on token traffic from more than 10,000 companies that use its Unity Gateway. The data runs from 21 September to 7 November 2026. Databricks says it shows two things. Enterprise AI is moving from one vendor to many. And open-weight models are taking a bigger share than they were four months before.
One-vendor use fell by more than half
In July, 64 percent of enterprises in the data used only one AI provider, Databricks says. By November, that figure fell to 28 percent. So almost three in four enterprises now use at least two providers. The company's own gateway makes multi-provider use easier, which is worth noting when reading these numbers. But the direction is clear, and Databricks publishes it in that caveat.
Open-weight share rose from 7 to 31 percent
Open-weight models took 7.3 percent of enterprise token traffic in July, Databricks says. In November, that figure is 30.9 percent. That is more than a fourfold rise in four months. An open-weight model is one whose weights are published. A team can run it on its own hardware. Or it can run through a provider other than the one that trained it.
One new model can take a third of token share in six weeks
When Claude Opus 5.5 launched, it captured 37.9 percent of proprietary-model tokens within six weeks, Databricks says. A product that was built around a specific model a quarter ago is likely running a different one now. This is one data point from one gateway. It is what Databricks has seen in its own data. It is not the whole market.
The cost spread between users is 281 times
The median user on Databricks spends 2.49 US dollars a week on AI, the company says. The user at the 99th percentile spends 699.74 US dollars a week. That is a spread of 281 times, and the median has itself doubled from 1.10 US dollars six weeks before. The index names one cost-quality pair: GPT-6 Luna and Claude Opus 5.5 both reach about 89 percent task success on Databricks' internal coding benchmark. Luna costs 0.05 US dollars a task on average, Databricks says. Opus costs 1.62, which is more than 30 times the Luna figure.
What a team building an AI product should take from this
A product that assumes one provider for the next year is planning for July. The data is Databricks' own and tilts toward Databricks customers. Even so, the pattern of swapping models is visible in its own numbers. A product that can swap a model without rewriting its calling code will save work. The year already has one case of a 30x cost difference at the same accuracy.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


