Stripe's OpenRouter deal was one of three open-weight acquisitions in a month

TechCrunch reported on 28 August that open-weight model companies have become the most active acquisition targets in the sector. Three deals in short order:

• Nvidia, reportedly around $13 billion for Hugging Face, unconfirmed
• Nvidia, $6 billion with Poolside, an open-weight model builder, with most of its employees joining Nvidia
• Stripe, more than $7 billion for OpenRouter, two weeks earlier

The adoption numbers

Compared with that spending, the usage figures are small. TechCrunch cites a survey of spending data by Ramp finding 6 percent of companies use open-weight models, and a measurement by Jellyfish putting it at 2 percent of software engineers.

Nik Albarran, AI product lead at Jellyfish, told TechCrunch where the usage is concentrated: companies with repeated inference workloads, such as customer service chat. High volume, high repetition, and an open model can be tuned to answer cheaply.

For coding and agent work he says the most capable proprietary models usually still win, because the requests vary more, the reasoning demands are higher, and the proprietary labs offer easier access and in some cases a token subsidy.

His view on what changes that: "There are not many companies where that is the case yet … [but] if the prices continue to go up from the frontier labs, more and more companies will be forced to at least consider it. When your AI-driven workflows are much more mature, that's when it makes sense to invest in self-hosting models."

Why the buyers are buying

Nvidia's reason comes from its market position. Its largest customers are building their own inference chips, including OpenAI's Jalapeño, announced this week. If model builders are making chips, owning part of the model business is a way to protect itself. Nvidia already releases its own open Nemotron family, and TechCrunch notes few developers have adopted it, which is a reason to buy the platform developers already use rather than keep building.

Stripe's reasoning is about how AI usage is measured and paid for. Patrick Collison, in a statement quoted by TechCrunch: "Tokens are the central currency for companies building with AI, and it's clear that the real-world economic potential will depend on making good use of scarce compute resources."

Lin Qiao, chief executive of Fireworks, which routes and hosts open-weight models for corporate users, says her company processes 40 trillion tokens a day, more than either the Gemini or OpenAI APIs. She expects specialisation to win: "Every single app company should consider hiring an in-house researcher. They can use their product and product data to build their own model."

The condition that makes open weights worth a look

This does not mean you should switch to open weights. At 2 percent of engineers, the honest view is that most teams have no reason to yet.

The useful part is knowing the condition under which it becomes worth considering, and Albarran states it plainly: a mature, repetitive, high-volume workflow, plus rising prices from the leading labs.

Both conditions are worth tracking on your own numbers rather than on a headline: what the leading labs charge you this quarter, and what the workflow you would move actually costs to run. The other thing worth watching is who ends up owning the tools a switch would run through.

Source

• Open-weight AI companies are the Valley's hottest acquisition targets, TechCrunch, 28 August 2026