Jaya Gupta of Foundation Capital argued in an X article titled Agent Cloud, published on 17 November, that AI agents now choose how work gets done while it runs, and that “Once the application holds the state, every model becomes interchangeable”.
Our own data shows buyers already spreading across suppliers: 71.1% of companies active on Databricks’ Unity Gateway in the week of 2 November had used models from two or more providers since 20 July, up from 36% in the week of 20 July, according to the Databricks AI Index. The thesis is hers; the counts are from the sources we cite.
Business Pill · Model Routing
An evergreen Business Engineer explainer on model routing, sending each task to the right AI model. It teaches the general idea only and does not discuss Gupta’s article.
The key insight: As we read it, Gupta’s thesis and our datasets meet at one point: buyers already hold more than one supplier. 71.1% of active Databricks gateway companies had used two or more model providers since 20 July, while the switching she describes happens inside a task, which none of our datasets measures.
What Gupta Argues
Gupta, whose X profile says she is at Foundation Capital, summed it up in an X post: “The boundary between what is “application” and “infrastructure” is merging”. In the article she writes: “Cloud computing let companies rent the means of computation. Agents let them continuously redesign the means of production.”
Her mechanism is substitution inside a task. She writes that “Every input now has a substitute the agent can reach mid-task”, listing “a small model for easy cases”, “local inference where privacy matters more than capability” and “a sandbox that runs the code instead of a long chain of reasoning”.
The consequence she draws for suppliers: “They stop being priced on what they sell and start being priced on what they let the customer stop buying”. She adds that “This doesn’t make them commodities”, because “A frontier capability may stay indispensable”.

What Was Posted on X
Posted by Jaya Gupta of Foundation Capital at 04:58 UTC on 17 November, this post gives a three-point summary of her X article Agent Cloud: the boundary between application and infrastructure is merging, agents let companies continuously redesign the means of production, and every input now has a substitute an agent can reach mid-task. The piece quotes the article and sets her thesis against our own Databricks, Gamma and OpenRouter data.
The Numbers She Cites
Gupta cites one of Cursor’s experiments in which the same task cost $10,565 “with a frontier model in every seat” and $1,339 “with a frontier planner directing cheap workers”. That is 87.3% less, by our arithmetic.
She cites Glean’s own benchmark against Claude Cowork, in which keeping state in sandbox files cut tokens per task from 4.4M to 1.3M and cost from $2.98 to $0.58, about 80.5% less on cost by our arithmetic. She also writes that public software traded at a median 6.2 times revenue at the end of 2024 and 3.3 times by March 2026.

What Our Own Data Shows
In our Databricks AI Index piece, 47.9% of gateway companies active in the week of 2 November had used three or more providers, and the share with four or more rose from 4.4% to 31.4%. Those figures count providers used since 20 July, not switches inside one task.
One buyer shows how fast the mix can move. In Gamma’s data, per our Gamma Model Index piece, OpenAI’s share of its trailing seven-day AI spend rose from 11.5% to 43.7% between the weeks ending 7 September and 6 November, while Anthropic’s fell from 63.2% to 36.0%. Gamma says the shift followed A/B tests on several key jobs; Anthropic kept 78% of the planning job.
On the supply side, one open-weight model can be bought from many hosts. OpenRouter lists 50 endpoints from 39 providers for DeepSeek’s V4.1 Flash 1017, and DeepSeek’s own servers carried 12.4% of that model’s OpenRouter tokens over the latest day, per our DeepSeek piece.
The Structural Read
As we read Gamma’s index, one heavy buyer moved its bill between two labs in about two months: OpenAI’s share rose from 11.5% to 43.7% while Anthropic kept 78% of the planning job, a per-job choice rather than a wholesale switch.
As we read the OpenRouter table, the supply side is already split for open-weight models: 39 providers list DeepSeek’s V4.1 Flash 1017, and DeepSeek’s own servers carried 12.4% of its OpenRouter tokens over the day.
Jaya Gupta, Foundation Capital, Agent Cloud (X article, 17 November 2026)
“Each one is a credible alternative to buying.”
Three Implications
FOR MODEL LABS Gupta writes that a lab can help customers send less through its models only until doing so cuts its own revenue.
FOR APPLICATION COMPANIES She cites Intercom moving its Fin AI agent from seats to $0.99 per resolved conversation.
FOR SOFTWARE VALUATIONS She writes that public software traded at a median 6.2 times revenue at the end of 2024 and 3.3 times by March 2026.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The framework says: “Foundation models become relatively commoditized as orchestration layers capture differentiation.”
As we read Gupta’s article, her agent cloud is that orchestration layer taken one step further: the layer that decides, per task, whether to buy a model call at all. That reading is ours; the analysis does not discuss Gupta or Foundation Capital.
What Is Not Established
The Cursor and Glean figures are as Gupta cites them; we have not checked them against Cursor’s or Glean’s own publications, and Glean’s is its own benchmark. Gupta does not say whether Foundation Capital has investments in the companies she names.
Our datasets do not measure mid-task switching. Databricks counts gateway customers only, Gamma is a single company, and OpenRouter shows where one model’s traffic is served. We did not contact Gupta or Foundation Capital for this piece.
The Bottom Line
Gupta argues that agents let applications swap models, compute and code mid-task, so suppliers are priced on what they let customers stop buying. In the data we track, 71.1% of Databricks gateway companies active in the week of 2 November had used two or more model providers since 20 July, up from 36% in July.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Gupta’s words are quoted from her X article Agent Cloud and her X post of 17 November. The Cursor, Glean and software-multiple figures are as she cites them and were not independently verified. Databricks, Gamma and OpenRouter figures are from our earlier pieces, linked in the text, which cite the Databricks AI Index, the Gamma Model Index and OpenRouter’s model pages. The 87.3% and 80.5% figures are our arithmetic. We did not contact Gupta or Foundation Capital for this piece.
Sources: Jaya Gupta on X (17 Nov 2026) · Jaya Gupta, Agent Cloud, X article (17 Nov 2026) · Databricks AI Index (16 Nov 2026)









