Open-weight models handled 30.9% of the tokens sent to Databricks-hosted models on its Unity Gateway in the week of 2 November, up from 7.3% in the week of 20 July, according to the Databricks AI Index, which Databricks published on 16 November. Proprietary models had the other 69.1%.
Within the open-weight group, DeepSeek went from 0% of tokens in the week of 20 July to 28.2% in the week of 2 November, by our sums of the page’s model figures. Our number: the open-weight share rose 23.6 points, about 4.2x, in 15 weeks, by our arithmetic.
Business Pill · Embedded dependency explained
An evergreen Business Engineer explainer on embedded dependency, what sits inside the things a company buys. It teaches the general idea only and does not discuss the Databricks AI Index.
The key insight: As we read the Databricks AI Index, the open-weight share climbed fast and then held: it rose from 7.3% to 33.2% of tokens to Databricks-hosted models between the weeks of 20 July and 14 September, and has stayed between 29% and 32.6% since, at 30.9% in the week of 2 November.
What the Index Counts
The index page counts open-weight as “GLM, Kimi, DeepSeek, Qwen, Llama, GPT-OSS and Gemma; the rest is proprietary (Claude, GPT, Gemini and Grok)”. This chart measures “Input plus output tokens to Databricks-hosted models on Unity Gateway”. Databricks co-founder Patrick Wendell wrote on X that the index data “covers >10,000 companies”.
Wendell wrote that “OSS models are quickly increasing in token share amongst enterprises, growing from nearly zero to 30% in the 5 month period”. The page’s chart starts at 7.3% in the week of 20 July and peaked at 33.2% in the week of 14 September; since then it has stayed between 29% and 32.6%.

What Was Posted on X
Patrick Wendell, a Databricks co-founder, wrote in the third post of his 16 November thread that open-source models grew from nearly zero to 30% of enterprise token share in five months, and that, anecdotally, companies use them mostly to cut costs on high-volume work. The index chart this piece reports starts at 7.3% in the week of 20 July and reads 30.9% in the week of 2 November.
Who Gets the Open-Weight Tokens
Zhipu’s GLM models had 60.4% of open-weight tokens in the week of 2 November, by our sums, down from 92.3% in the week of 20 July. GLM 5.3 Flash alone had 33.3% and GLM 5.3 had 21.6%. DeepSeek’s DS V4.1 Flash had 28%, and Moonshot’s Kimi K3 had 7.9%, the page shows.
Spend splits differently. At list prices Moonshot took 23.9% of open-weight spend against 8.8% of tokens, about 2.7x its token share, while DeepSeek took 9.1% of spend against 28.2% of tokens, by our arithmetic. Zhipu took 63.1% of spend. Moonshot’s spend share was 51.1% in the week of 14 September.

Why Enterprises Use Them
Wendell wrote: “Anecdotally, we see those models used mostly to reduce costs on high volume workloads (large batch workloads, coding, etc).” The page lists a Databricks post of 4 November, “The Enterprise Case for Open-Weight Models on Databricks”, described as weighing “cost, data control and governance”.
On the page’s coding benchmark, built from Databricks’ own engineering tasks, GLM 5.3 Flash scored 89.1 at $0.178 a task and DeepSeek V4.1 Flash 1017 scored 93.1 at $1.021. Claude Opus 5.5 scored 89.1 at $1.616 a task, the page shows.
This is the second piece in our set on the index; the first covers Anthropic’s 70% of closed-model spend. Our earlier pieces covered Menlo Ventures’ enterprise AI spend figures and a Berkeley paper on agent harness cost.
The Structural Read
As we read the provider split, open-weight volume and open-weight spend reward different labs: DeepSeek had 28.2% of open-weight tokens but 9.1% of spend by our sums, while Moonshot had 8.8% of tokens and 23.9% of spend.
As we read Wendell’s thread, the case for these models is cost on high-volume work, and he calls that observation anecdotal; the index itself shows token shares, not savings.
Patrick Wendell, Databricks co-founder, on X
“Anecdotally, we see those models used mostly to reduce costs on high volume workloads (large batch workloads, coding, etc).”
Three Implications
FOR BUYERS On the page’s coding benchmark GLM 5.3 Flash scored 89.1 at $0.178 a task, the same score as Claude Opus 5.5 at $1.616.
FOR MODEL MAKERS Zhipu, DeepSeek and Moonshot together had 97.4% of open-weight tokens in the week of 2 November, by our sums.
RELATED Our first piece on the index covers the proprietary side, where Anthropic has 70.1% of spend by our sums.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.
The framework’s rule is: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, the index shows enterprises treating part of the model layer as the abundant one, routing about three in ten hosted tokens to open-weight models, while 69.1% still goes to proprietary ones. The framework predates the index and does not discuss it.
What Is Not Established
Wendell labels the cost-cutting reason anecdotal. The open-weight share covers tokens to Databricks-hosted models, while the provider charts also count traffic routed to external providers, so the two scopes differ. Spend figures are at list prices, which the page says are not what Databricks bills.
We read the index page in full, including the values behind each chart, and Wendell’s six-post thread. We did not contact Databricks, Zhipu, DeepSeek or Moonshot for this piece.
The Bottom Line
Open-weight models took 30.9% of tokens to Databricks-hosted models on Unity Gateway in the week of 2 November, up from 7.3% in the week of 20 July. DeepSeek rose from 0% to 28.2% of open-weight tokens, by our sums.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. All figures come from the Databricks AI Index page at developers.databricks.com/ai-index, which we read in full on 16 November 2026, including the exact values its charts plot, and from Databricks co-founder Patrick Wendell’s six-post thread on X of the same day. The page marks these charts as updated on 2 November 2026. The open-weight share covers tokens to Databricks-hosted models on Unity Gateway; spend shares are at providers’ list prices, not Databricks’ billing.
Provider totals, point changes and multiples are our arithmetic from the page’s model-level values. Databricks, Zhipu, DeepSeek and Moonshot were not contacted for this piece.
Sources: Databricks AI Index (page updated 2 Nov 2026, published 16 Nov 2026) · Patrick Wendell on X, 16 Nov 2026 (post 1 of 6) · Patrick Wendell on X, 16 Nov 2026 (post 3 of 6)









