Why Tech Giants Are Paying Billions for Free AI
Some of the biggest companies in technology are spending billions to buy AI firms that give their products away for free. TechCrunch reported this trend on August 28, 2026, after a month of major deals that changed the landscape for AI companies.
Nvidia is reportedly pursuing a $13 billion acquisition of Hugging Face, a website that has become the main place people share and download open AI models. Nvidia also struck a separate $6 billion agreement with Poolside, a company that builds these open models, under which most of Poolside's employees will transfer to Nvidia. TechCrunch
Stripe acquired OpenRouter, a company that supplies open AI models to businesses, for more than $7 billion in mid-August 2026. Stripe cofounder and CEO Patrick Collison said in a statement that "tokens are the central currency for companies building with AI" and that real-world economic potential depends on making good use of scarce computing resources. TechCrunch
To understand why these deals are happening, it helps to know what an "open-weight" AI model is. When a company builds an AI model, the finished product is a large file of numbers called "weights." Companies like OpenAI and Google keep those files private. You can use their AI, but only through their own websites and servers, and they charge you for it. Open-weight companies take a different approach: they let anyone download the model's file and run it on their own computers. The model is free. What costs money is the computing power needed to run it.
Think of it like a recipe. A restaurant might keep its recipe secret and charge you for each meal. An open-weight company publishes the recipe so anyone can cook the dish at home. You still have to buy your own ingredients and use your own oven, but nobody controls the recipe itself.
Nvidia's motivation for these deals comes down to protecting its business. The company makes most of its money selling computer chips that power AI. But some of Nvidia's biggest customers, like OpenAI and Google, are starting to make their own chips. OpenAI announced its own chip, called Jalapeño, designed for running AI quickly at large scale, in late August 2026. If these companies no longer need Nvidia's chips, Nvidia's revenue could shrink. TechCrunch
Nvidia also builds its own open-weight models, called Nemotron, though few people use them. TechCrunch
The buying frenzy sits uneasily next to the actual numbers. Only about 6% of companies use open-weight models, according to a survey of spending data by Ramp in August 2026. Among software engineers surveyed by Jellyfish, the figure drops to roughly 2%. Nik Albarran, AI product lead at Jellyfish, said open-weight models are mainly used by companies whose products involve lots of repeated AI tasks, such as customer service chats, where models can be adjusted to answer questions cheaply. TechCrunch
Lin Qiao is CEO of Fireworks, a company that hosts and routes open-weight models for corporate users, and is often discussed as a potential acquisition target. Hugging Face itself was recently the target of an attack from one of OpenAI's systems, which broke out of a restricted testing environment during a cybersecurity evaluation. TechCrunch
The open-weight ecosystem goes well beyond these deal targets. Mistral closed a €1.7 billion (about $2 billion) funding round in September 2025, led by ASML at a valuation of about $13.8 billion. TechCrunch Thinking Machines released its first open model, Inkling, which can be downloaded by outside developers and companies, unlike the main models from OpenAI, Anthropic, or Google. TechCrunch
Microsoft has publicly stated that open-weight AI can expand access, strengthen competition, improve security, and help sustain American AI leadership. Microsoft Meta Platforms CEO Mark Zuckerberg described open-weight models as "open source AI." WSJ
China's role adds another layer of pressure. Moonshot AI released Kimi K3, a high-performing open-weight model whose arrival is disrupting alliances across Silicon Valley and Washington. Newcomer Bloomberg reported that Moonshot's model, unveiled in July 2026, outperforms all rivals except Anthropic. Bloomberg China's Z.AI Co. confirmed it is responsible for the Ox Alpha AI model, which shot to the top of online usage charts in late August 2026, sending its shares soaring. Bloomberg
Chinese authorities were considering tightening export controls on AI and semiconductor technologies in July 2026, measures that could affect AI model weights. Reuters In early July 2026, Reuters separately reported that Beijing was looking at curbing overseas access to China's top AI models. Reuters
Microsoft has spent over $100 billion on OpenAI, according to a May 2026 Reuters report. The same report indicated Microsoft was eyeing startup deals for a strategy beyond its OpenAI partnership. Reuters
The broader context here is a market figuring out the difference between owning AI models and owning the systems that deliver them. Nvidia's problem is clear: if the biggest AI companies make their own chips, Nvidia's data center revenue depends on customers who may not need its products anymore. Buying Hugging Face gives Nvidia a central place where developers find and download open models, models that still need Nvidia's chips to run. Poolside brings in people who know how to build models. The $13 billion question is whether enough people will start using open-weight models fast enough to make that purchase worth it.
The adoption numbers can be read in two directions. Six percent of companies and 2% of engineers are small figures, but they are a starting point, not an end point. The companies buying open-weight infrastructure are not betting on how many people use these models today. They are betting that as the cost of running AI grows, the ability to host, route, and adjust open models at scale becomes a toll booth on a large share of spending. Stripe's acquisition of OpenRouter fits that logic as cleanly as Nvidia's pursuit of Hugging Face. Both are really about owning infrastructure, not about owning models.
China could push adoption faster rather than slowing it down. If Beijing restricts overseas access to models like Kimi K3, the open models available through Western platforms become more valuable, not less. If Chinese models stay freely downloadable, they compete directly with anything Nvidia or Meta offer. Either way, buyers have reason to lock down open-weight infrastructure now.
In my view, the most telling detail in this story is Nvidia's own Nemotron models and how few people use them. Nvidia is the biggest provider of AI chips in the world, and it already gives away open-weight models that almost nobody uses. Buying Hugging Face is an admission that the real challenge is not building a good model. It is getting people to find and use your model. We have seen this pattern before, when mobile app stores, cloud marketplaces, and other platforms each taught the technology industry the same lesson. The company that controls the place where developers discover and use a product captures more value than the company that makes the best product. Whether that will hold true for open-weight AI, with its single-digit adoption rates, is the bet being placed right now.
