Databricks data shows open-weight AI models now handle 31% of enterprise workloads as cost-cutting accelerates
The Gist
- Open-weight AI models jumped from 7.3% to 30.9% of enterprise token traffic in 4 months
- Chinese models (Zhipu, DeepSeek, Moonshot) dominate open-weight usage on US infrastructure
- Enterprises use open-weight for high-volume tasks like classification, not frontier model replacements
Key Quotes
If your product makes a model call on every record, every ticket, or every email, a third of your enterprise buyers' traffic now runs on models that cost a fraction of frontier pricing.
The most capable model is not the most used. Once a model is good enough at a lower price, enterprises move to it.
Key Insights
- Open-weight models now handle 31% of enterprise AI workloads, up from 7.3% in July, as companies shift high-volume tasks to lower-cost options.
- Enterprises are rapidly diversifying model usage, with average unique models per customer rising from 4.2 to 7.1 since January, treating models like cloud regions.
- Anthropic's share of proprietary model usage grew from 57.6% to 68.8%, largely due to the launch of Opus 5.5, which captured 37.9% of proprietary traffic in six weeks.
- The most capable model isn't always the most used; enterprises prioritize cost efficiency once a model is 'good enough' for their needs.
- Databricks' internal benchmark shows that cost per completed task, not cost per token, is the critical metric for pricing AI products, with a 32x cost difference observed for the same score.
- Median enterprise AI spend per user more than doubled from $1.10 to $2.49 in six weeks, with P99 spend reaching $699.74 per week, highlighting extreme variability in usage.
Actionable Takeaways
- Evaluate shifting high-volume AI workloads (e.g., classification, extraction) to open-weight models to reduce costs.
- Benchmark AI models based on cost per completed task, not cost per token, to optimize gross margins.
- Monitor enterprise adoption trends of open-weight and proprietary models to align pricing and product strategy.
- Prepare for extreme variability in AI spend per user, with P99 users costing 281x the median.
Data Points
- 31% (Share of enterprise AI token traffic handled by open-weight models in November, up from 7.3% in July.)
- 64% to 28% (Drop in enterprises relying on a single model provider from July to November.)
- 4.2 to 7.1 (Increase in average unique models per customer since January.)
- 68.8% (Anthropic's share of proprietary model usage, up from 57.6%.)
- $2.49 (Median enterprise AI spend per user, up from $1.10 over six weeks.)
- $699.74 (P99 weekly spend per user on AI tokens.)
RevBots.ai View:
GTM teams must audit their AI stack COGS immediately as competitors rebuilding on open-weight models will undercut pricing by 3-5x.
Full Story:
SaaStr →
Join The RevBots ARMy
The insider daily for Autonomous Revenue Masters.