Why AI startups get acquired: 14 deals from 2025 and 2026

How we built the list

We started from the acquisitions recorded while maintaining SOTA's catalog, listed in our guide to AI products that were renamed, acquired or shut down. We kept deals announced between May 2025 and July 2026 and checked each one against the buyer's or target's announcement, a regulatory filing or a major news outlet. Deal sizes, funding and adoption figures are as stated in those sources; where we could not find a figure we say it was not disclosed rather than estimate it. GitHub stars were read on October 3, 2026.

Fourteen deals is a small sample, and two of them, Nvidia and Groq and Google's part of the Windsurf deal, were licensing agreements with hires rather than full acquisitions. Read the patterns below as observations, not rules. The full table with sources is available as a CSV.

Most buyers were filling a gap in their own platform

Eight of the fourteen deals added a capability that sat right next to the buyer's core product. ClickHouse, a database company, bought Langfuse, whose observability platform already stored its traces in ClickHouse, and LibreChat, a chat interface, to offer what it calls an open-source agentic data stack. Elastic bought Jina AI to own the embedding and reranking models behind its search. Cloudflare bought Replicate for its catalog of more than 50,000 models, which it is bringing to Workers AI.

Software platforms did the same for AI features their customers asked for. Workday bought Flowise to give customers and partners an agent builder. OpenAI bought Promptfoo to add security testing to Frontier, its enterprise agent platform. Tricentis bought Tabnine for its enterprise code context engine to feed its testing agents, and Anaconda bought Kilo Code to cover the layer where developers write code with agents.

In each case the target was already used by the buyer's customers or built on the buyer's technology, so the buyer could see the demand before paying for it.

The targets had adoption a buyer could measure

Almost every target could point to usage numbers. Langfuse counted 19 of the Fortune 50 and 63 of the Fortune 500 among its customers and 26 million SDK installs a month. Promptfoo said more than 25% of the Fortune 500 relied on it, with 130,000 monthly active developers, from a team of 23 people. Kilo Code reached more than 3 million developers within 16 months. Flowise had more than 42,000 GitHub stars, Rows had been used by more than 2.2 million people, and Shopify ran LibreChat with more than 30 internal MCP servers.

Enterprise names mattered as much as raw user counts. Several announcements led with Fortune 500 adoption or named customers, which tells a buyer that the product has already passed security reviews at large companies.

Open source was common, and mostly survived the deal

Six targets had an open-source core: Flowise, LibreChat, Langfuse, Promptfoo, Continue and Kilo Code, and ProducerAI began as the open-source Riffusion project. Open source built the developer adoption that buyers could measure, and it did not stop a sale: permissive MIT or Apache 2.0 licences, sometimes with separate enterprise directories, were the norm.

After the deals, Langfuse, LibreChat, Promptfoo and Flowise all stated that they would stay open source. Continue was the exception: the product was shut down and its Apache 2.0 code was left to the community.

Small teams and modest funding were enough

Several targets were young and lightly funded. Langfuse had raised a $4 million seed round, Continue about $5 million and Kilo Code an $8 million seed. Promptfoo had raised an $18.4 million Series A and employed 23 people. Kilo Code was about 16 months old when Anaconda bought it, and Langfuse and Continue about three years.

Prices were rarely public. Only Jina AI's appears in a filing: Elastic reported a total purchase consideration of $44.6 million for a company that had raised $39 million. The Nvidia and Groq licence was reported at about $20 billion. None of the other buyers disclosed terms.

Talent deals look different

Four deals were mainly about people. Google paid $2.4 billion to license Windsurf's technology and hire its CEO and some researchers, and Cognition then bought the rest of the company, its product and its 350-plus enterprise customers. Nvidia licensed Groq's inference technology and hired its founder while Groq continued under a new CEO. Cursor's purchase of Continue brought a co-founder into Cursor and ended the product, and Google moved the ProducerAI team into Google Labs and relaunched the product as Google Flow Music.

These deals reward a team that is strong in a field the buyer wants to enter, but they are less predictable for users of the product.

What happened to the products

Nine of the fourteen products kept running under their own names, at least so far. Two were renamed: Windsurf became Devin Desktop and ProducerAI became Google Flow Music. Tabnine's technology was folded into Tricentis's testing platform, and tabnine.com now points there. Two were shut down: Rows closed on May 31, 2026 after its team joined Superhuman's Coda, and Continue closed with a deadline for users to export their data.

If you depend on a product that is acquired, check the announcement for commitments about pricing, data export and open-source licences, and keep your own export in a standard format.

What founders can take from this

Build next to a platform, not just on top of one. The clearest deals bought a product that the buyer's customers already used alongside the buyer's own: traces stored in ClickHouse, models served next to Cloudflare, agents for Workday customers.

Make your adoption easy to verify. Monthly active developers, SDK installs, named enterprise customers and Fortune 500 penetration appeared again and again in the announcements.

Open source helped more than it hurt. It created the bottom-up adoption buyers looked for, and permissive licences did not prevent a sale.

A small, focused team can be enough. Several targets sold within two to three years on single-digit millions of funding.

Decide early what you would want for your users. Most products kept running, but two were shut down, so commitments on data export and licences are worth negotiating.

Sources & method

Official documentation was checked on 2026-10-03. Selection advice is our assessment. Interfaces and requirements may change; check the linked source for the version you intend to use.

• SOTA dataset: AI acquisitions 2025-2026 (CSV)
• ClickHouse raises $400M and acquires Langfuse
• Langfuse self-hosting architecture (ClickHouse)
• ClickHouse welcomes LibreChat
• Elastic and Jina AI join forces
• Elastic Form 10-Q (Jina AI purchase consideration)
• Replicate is joining Cloudflare
• Workday acquires Flowise
• Promptfoo is joining OpenAI
• OpenAI to buy Promptfoo (CNBC)
• Anaconda acquires Kilo Code
• Kilo raises $8M (CNBC)
• Tricentis acquires Tabnine
• Cursor acquires Continue (The New Stack)
• Cognition to buy Windsurf (CNBC)
• Nvidia and Groq deal (CNBC)
• Rows will be joining Superhuman
• Google acquires ProducerAI (Music Business Worldwide)
• Litmaps acquires ResearchRabbit (Scoop)

How SOTA reviews projects · Suggest a correction · Download the evaluation checklist