Nvidia, Stripe Buy Open-Weight AI Firms

Nvidia is reportedly set to acquire Hugging Face for $13 billion, according to TechCrunch. This follows Nvidia's $6 billion agreement with Poolside and Stripe's acquisition of OpenRouter for more than $7 billion two weeks ago.

These massive deals target companies in the open-weight AI sector, where models and their underlying code are shared publicly. The spending reflects a strategic shift as major tech firms hedge their bets against the dominant frontier labs like OpenAI and Google.

For Nvidia, the chip-making giant, the acquisitions are a defensive play. The company seeks to reduce its dependence on deals with major hyperscalers and frontier labs, especially as those same labs develop their own inference chips. OpenAI, for instance, announced capabilities for its Jalapeño chip this week.

Nvidia already creates its own Nemotron family of open-weight models, but their adoption has been limited. Controlling Hugging Face, described as a "GitHub for the AI era," would give Nvidia direct access to a vast ecosystem of developers it can steer toward its hardware and software standards.

The Economics of Inference

The soaring cost of AI inference is a key driver. Companies are exploring cheaper alternatives, including models from Chinese firms like Moonshot, DeepSeek, and Alibaba. Current adoption remains small, however.

A survey of spending data by Ramp found only 6% of companies use open-weight models. Data from developer tools firm Jellyfish puts the figure at just 2% of software engineers.

Nik Albarran, AI product lead at Jellyfish, told TechCrunch these models are primarily used for high-volume, repetitive tasks like customer service chats. They can be tuned to answer questions cheaply. Stripe's CEO Patrick Collison echoed this, stating in a release that "the real-world economic potential will depend on making good use of scarce compute resources."

For more complex tasks like coding or agentic work, frontier models from proprietary labs often win out. They offer easier access and sometimes token subsidies. Albarran notes the main current appeal of open models is control and configurability, not cost savings.

"If the prices continue to go up from the frontier labs, more and more companies will be forced to at least consider it," Albarran said. He added that investing in self-hosting models makes the most sense when a company's AI-driven workflows are mature.

The Push for Specialized Models

Lin Qiao, CEO of open-weight models host Fireworks, argues for a future of highly specialized intelligence. Her company processes 40 trillion tokens daily, a volume she claims exceeds the API traffic of either Gemini or OpenAI.

"Every single app company should consider hiring an in-house researcher," Qiao told TechCrunch. "They can use their product and product data to build their own model. The future is actually specialized intelligence."

Fireworks, often discussed as a potential acquisition target itself, bets on model diversity. As large language models proliferate and improve, companies will find it easier to train models specifically for their needs.

The recent acquisition spree underscores how early the AI business still is. The dominance of a few frontier labs is not seen as inevitable. As tech giants look to secure their positions, the allure of open technology and the developers who build with it is proving powerful. The funding trends show capital is flooding into this alternative ecosystem.