Chart of the Day · SEP 19, 2026

Vercel AI SDK Downloads vs. the Lab SDKs

Vercel's AI SDK grew 6.5× in under a year, but the lab SDKs grew 8.7×. The layer that makes models swappable did not capture the growth.

Mode VC ·

The signal

The Vercel AI SDK fell from 31 to 23 weekly downloads per 100 downloads of the OpenAI, Anthropic and Google SDKs, after a peak of 40.

Weekly npm downloads of the ai package per 100 combined downloads of openai, @anthropic-ai/sdk and @google/genai. Complete Monday–Sunday weeks, 6 Oct 2025 – 30 Aug 2026.

Sources: npm registry downloads API · Official launch posts (launch feed, captured 7 Oct 2026)MODE VC

ai is the core package of Vercel's AI SDK; model providers plug in through adapter packages such as @ai-sdk/openai. Downloads count installs, including CI and packages that depend on these SDKs, so lab SDK counts include other tools built on top of them. Weeks with an npm gap day (the API returns 0) are dropped for all packages; the latest complete week is 24 Aug.

Show the data (43 rows)
Week ofaiLab SDKs combinedai per 100 lab SDK
6 Oct 20253.6M11.7M31.1
13 Oct 20253.7M12.2M30.8
20 Oct 20254.0M11.7M34.1
27 Oct 20254.2M11.6M36.4
3 Nov 20254.6M12.2M37.5
10 Nov 20254.8M12.3M39.2
17 Nov 20255.1M13.3M38.5
24 Nov 20254.9M12.1M40.4
1 Dec 20255.5M14.4M38.0
8 Dec 20255.2M14.5M35.9
15 Dec 20255.3M14.4M36.5
22 Dec 20253.1M9.4M32.4
29 Dec 20252.8M9.6M29.6
5 Jan 20265.6M15.7M35.5
12 Jan 20266.2M16.7M37.0
19 Jan 20266.4M18.2M35.1
26 Jan 20267.0M21.8M32.3
2 Feb 20267.6M22.8M33.2
9 Feb 20267.8M23.9M32.6
16 Feb 20268.0M23.4M34.0
23 Feb 20268.7M28.8M30.2
2 Mar 20269.9M31.4M31.5
9 Mar 202610.1M35.4M28.5
16 Mar 20269.8M34.9M28.1
23 Mar 202610.4M35.9M29.1
30 Mar 202610.1M37.1M27.2
6 Apr 202610.6M40.1M26.5
13 Apr 202611.5M42.9M26.9
20 Apr 202611.9M45.3M26.2
27 Apr 202612.1M50.7M23.9
4 May 202612.6M48.0M26.2
11 May 202613.2M50.4M26.3
18 May 202613.7M54.0M25.5
25 May 202613.0M61.9M21.0
8 Jun 202615.1M65.2M23.2
15 Jun 202615.7M68.8M22.9
29 Jun 202615.1M63.7M23.8
13 Jul 202617.7M72.7M24.4
20 Jul 202618.8M74.7M25.1
27 Jul 202619.8M80.2M24.7
3 Aug 202620.6M79.3M25.9
17 Aug 202622.4M89.6M25.0
24 Aug 202623.6M101.8M23.2

The Mode read

Our launch feed lists 10 new models from 8 labs in the first 19 days of September. A layer above the models, such as Vercel's AI SDK, promises that a team can try each one by changing a line of config. If models really are swappable, that layer should be taking a growing share of developer installs. Is it?

Not on npm. The ai package grew 6.5×, from 3.6M to 23.6M weekly downloads. The three lab SDKs together grew 8.7×. Per 100 lab SDK downloads, ai went from 31 to a peak of 40 in the week of 24 Nov 2025, then down to 23. Swapping does happen inside the layer: the Anthropic adapter went from 61% to 100% of the OpenAI adapter.

So the abstraction works, and it still did not take the market. Most new installs went straight to a lab. A neutral layer makes the model a commodity, which is our thesis, but making something a commodity does not mean you own the customer. The switch is useful. The moat is whatever the developer cannot move as easily: their deployment, their data, their evals, their bill.

Making models swappable is a feature. Owning what cannot be swapped is the business.

What we would back: Developer platforms that pair a model-neutral layer with something sticky: hosting, data, evals or billing.