AI Momentum 104.8 +2.6%GitHub Pulse 5171 -5.9%Capital Flow $206.5B +5650.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Enterprise Adoption

Only 28% of companies on Databricks' AI service now stick to one AI supplier, down from 64% in July

Corporate buyers are spreading their AI work across more suppliers, and switching quickly. That is the main finding of a new index from Databricks, the data software company valued at $190 billion in August. It published the figures on November 16.

The index draws on anonymized usage from more than 10,000 companies that run their AI requests through Unity Gateway. That is a Databricks service that sends each request to whichever AI model a company picks, from OpenAI, Anthropic, Google or others. It also tracks budgets and spending caps.

In July, about 64% of those companies used a single AI model supplier. By November 2, only 28% did. Over the same stretch, open-weight models went from 7.3% of the AI work flowing through the service to 30.9%. Open-weight models are ones whose makers publish them, so companies can run them on their own or rented computers without paying a lab for each use.

For buyers, the message is that being tied to one AI supplier may matter less than many assume. For investors in the AI labs, it is a warning that market share can move within weeks.

Share moves in weeks, not years

The swings in the data are large. In the six weeks after Anthropic released Claude Opus 5.5, that one model went from 7.2% to 37.9% of usage of paid, closed models. Anthropic's overall share rose from 57.6% to 68.8%, though Databricks does not say what that share is measured against.

Loyalty did not protect Anthropic's other products. Its Claude Fable 5.1 model fell from 5.5% of closed-model usage to 0.6% in the six weeks after its launch. Databricks says OpenAI's share also grew meaningfully, led by its GPT-6 series models, but it gave no exact figure.

Venture investor Tomasz Tunguz, writing about the same data, reads it as the two leading labs splitting the market by price. In his analysis, Anthropic takes about 70% of spending on closed models and dominates the most expensive tier. OpenAI wins the high-volume, low-cost tier after two rounds of cuts that lowered the price of its low-cost Luna model by about 90% in total.

Tunguz also notes that the open-weight share, after peaking at about 33% in September, stalled at 31% in the same week OpenAI made its latest price cut. He treats the timing as a sign that price cuts can slow the shift to open models, though the two events coinciding does not prove one caused the other.

Cheaper per word is not cheaper per job

Databricks also ran coding tasks drawn from its own software work, checked against hidden tests. Two models, OpenAI's GPT-6 Luna and Anthropic's Claude Opus 5.5, each completed 89.1% of tasks. Luna averaged about $0.05 per task. Opus 5.5 averaged $1.62, roughly 32 times as much.

The most capable model tested, Claude Fable 5.1, completed 96.0% of tasks at $7.47 each. A model called Muse Wonder scored 94.9% at $1.53. The last 1.1 points of quality cost nearly five times as much. Two Chinese open-weight models, DeepSeek V4.1 and GLM 5.3, tied for third at 93.1%.

Low prices per word can also mislead. Google's Gemini 3.8 Flash is sold cheaply, but it took about 140 back-and-forth steps per task and cost $11.97 per job, among the highest in the test. Databricks' own summary: "lower token prices don't guarantee lower costs."

Spending is also highly uneven inside companies. The median user's AI spending more than doubled in six weeks, to $2.49 a week. Users at the 99th percentile spent $699.74 a week, about 281 times the median. For finance chiefs, that points to a small group of heavy users accounting for an outsized share of the bill.

Who is counting, and what they sell

The numbers deserve caution. Databricks sells the routing service it measured, along with hosting for open-weight models. A story of buyers mixing suppliers and moving work to cheaper models supports its sales pitch, and its blog post closes by pointing readers to the gateway's cost-management tools.

The sample is also narrow. It covers only companies that use Unity Gateway, and the post describes that base loosely, as both "thousands" of enterprises and more than 10,000 companies. Tunguz notes the traffic leans toward data and coding work, not the whole corporate market.

Spending figures are estimated from published list prices, not the discounted rates large buyers often negotiate. The single-supplier measure counts any supplier used since July 20, not suppliers used at the same time, according to a reading of the method by Superpower Daily. It also counts each lab's closed and open-weight models as one supplier, for example Google's Gemini and Gemma.

The coding tests come from Databricks' private codebase, and no outside party has checked them. Databricks itself says the results "may vary across organizations." It plans to update the index weekly.

Even with those limits, the data points the same way throughout: more suppliers, more price pressure, and fast switching when a better or cheaper model ships. For anyone valuing an AI lab on today's market share, the data suggests a lead can disappear within weeks.