The model labs are in almost every stack now. Most lists do not show it.

309 of the 756 lists name no model provider at all. But 263 of those companies use at least one vendor whose own list does. Follow the chain one step and the labs go from about half the market to nearly all of it. The same holds for the two largest inference vendors, shown last.

The carriers are ordinary infrastructure. Companies that do not list Anthropic reach it most often through Cloudflare (135 companies), Datadog (109), Sentry (104), Twilio (94) and Zendesk (80). For OpenAI it is Cloudflare, Slack, Sentry, Datadog and Salesforce.

Two steps means the vendor has written a lab into its own data terms. Your security review approved Sentry. It did not review the model provider on Sentry's own list.

Four deals since March changed who stands behind a vendor on these lists. A subprocessor list names the product. It does not name the new parent.

SpaceX buys Cursor

Signed in June, reported closed in August, about $60B. Among the 218 companies, lists that name xAI or Cursor went from 6 to 17. Glean, Procore, Railway, Depot, Samsara, Ahrefs and Avoca now name Cursor as a subprocessor.

Stripe buys OpenRouter

Announced August 19. Stripe is on 39% of lists. It has agreed to buy the router in front of hundreds of models, and it already owns Metronome, which meters usage.

Salesforce buys Fin (Intercom)

Signed June 15, $3.6B. Salesforce, Slack, Tableau and Heroku together are on 44% of lists. With Intercom that becomes 51%.

Dynatrace buys Arize

$915M. In the same months Dynatrace listed a model provider, Anthropic, for the first time.

Cursor's own list after the deal is the most telling. It added two GPU clouds (CoreWeave and Nebius) and two more model sources (Google Gemini and Meta). It dropped Mistral, Turbopuffer, MongoDB and GitHub. It still lists Anthropic and OpenAI. A company owned by one lab names its parent's three rivals as subprocessors.

The labs publish lists too. OpenAI lists AWS, Azure, Google Cloud, Oracle Cloud, CoreWeave and Cerebras. Anthropic lists AWS, Google Cloud and Azure, plus Palantir, Brave, Parallel, ElevenLabs and Turbopuffer.

There are 86 pairs where each company names the other as a subprocessor. Some of the loops: Anthropic and Cloudflare, Anthropic and ElevenLabs, Anthropic and Sentry, OpenAI and Salesforce, OpenAI and CoreWeave, Vercel and SpaceXAI, Fireworks AI and Linear.

One small move worth watching. OpenAI's March list had Intercom. Its November list does not, and Accenture is new on it. Anthropic's November list has Intercom. I do not know the reason. The months overlap with Salesforce agreeing to buy Intercom.

Across the 218 companies, 55 list more foundation model providers than in March and 3 list fewer. No company removed OpenAI or Anthropic. Companies with two or more providers went from 33% to 44%.

Anthropic grew twice as fast as OpenAI (32% against 16%). Gemini more than tripled. It was added by Cursor, Perplexity, Figma, Clay, Zed, Postman, Snyk and Wiz. There were six removals in the whole group, all in the second tier: Cohere (Asana, Notion), Mistral (Cursor), Amazon Nova (Box, Weaviate) and Meta Llama (Shopify). The big three grow. The middle stands still.

Many companies reach models through their cloud and list only the cloud. Reading the stated purpose of each entry: 81 of the 279 companies that list Azure say it is for AI or language models. So do 122 of the 506 that list Google Cloud and 66 of the 621 that list AWS. This is a keyword match on the purpose text, so treat it as a floor. Azure is the usual route to OpenAI models, so a count of OpenAI users taken from vendor names alone is too low.

24 companies listed a model provider for the first time. Most are not AI startups. 16 chose Anthropic, 15 chose OpenAI, and 9 chose both at once.

Dynatrace, New Relic, Darktrace, Wiz, Netlify, Neon, Airbyte, Starburst, Qdrant, Brex, Procore, Tailscale, Iterable, Workato, Gumroad, Weights & Biases.

Companies with a model-serving vendor went from 22% to 33%. In March the most-listed serving vendor was Groq. In November it is Baseten. Baseten and Fireworks AI each gained 14 companies and lost none.

The adopters are large products. Baseten: Datadog, HubSpot, MongoDB, Zapier, Glean, Sierra. Fireworks AI: Shopify, Asana, Datadog, Linear, Zapier. Modal: Notion, Linear, Glean, Sierra, Hex.

ByteDance is on 9 lists, Kling on 9 and MiniMax on 7. The companies are creative tools: Canva, Figma, ElevenLabs, HeyGen, Synthesia, Luma AI, Runway, Hedra, InVideo. Chinese text models (DeepSeek, Moonshot) appear only on routers and low-code tools that let the customer pick the model: ngrok, Taskade, Glide, n8n, Browser Use.

The 218 companies used 599 distinct vendors in March and 725 in November. In the layers I checked, the most-listed vendor lost share in most. Data pipelines are the exception.

New leaders by number of lists: auth went from Auth0 to WorkOS, analytics from Segment to PostHog, model serving from Groq to Baseten. In support, Pylon gained 12 companies and lost none. Zendesk was dropped by Ada, CircleCI, Dynatrace, Netlify, New Relic and OpenRouter.

Companies with a separate embeddings vendor fell from 6% to 3%. Jina AI went from 12 lists to 3. It is gone from the lists of Databricks, MongoDB, Shopify, Weaviate and Contentful. This rests on a small number of removals, so it is the weakest finding here. My read is that retrieval moved inside the model providers and the main database, but the lists do not prove that.