Chinese open-source AI is starting to win over U.S. businesses

Google DeepMind is losing its grip on elite AI talent, new data shows
DeepSeek seeks $7.4 billion at a $74 billion valuation.
OpenAI data-center head leaves amid broader exodus.
And OpenAI releases more details on the Hugging Face hack.

Enterprises are starting to look beyond America for their AI needs.

As Chinese open-weight AI models increasingly close the gap with their closed-source U.S. counterparts, more enterprises are warming to the idea of using Chinese alternatives. It’s easy to see why: open-weight models offer companies more opportunities to fine-tune AI models, are much cheaper, and generally give enterprises more control, including more assurance that their data isn’t being used to train potentially competing products.

While Anthropic and OpenAI still dominate how American businesses buy AI, new spending data suggests that at least a minor shift is underway.

According to Ramp’s latest AI Index—which tracks token and subscription spend across its customer base—the share of businesses paying for model serving platforms, which give companies access to open source and Chinese-developed models, rose to 6.1% of total AI-spending businesses in July, up from 4.5% in January 2026.

The shift may show increased enterprise interest in open-weight models such as Moonshot’s Kimi K3, which made waves on its release for being both an unusually large open-weight model but also one that showed coding and agentic performance close to leading proprietary systems.

Z.AI, another Chinese AI lab, this week confirmed that it built Ox Alpha, a previously anonymous model that had been performing well on AI benchmarks and gained enthusiastic reviews from AI developers over the weekend. The company has renamed it GLM-5.3-Flash and said it will charge $0.15 per million input tokens and $0.50 per million output tokens—another aggressively priced Chinese offering in a market where DeepSeek and Moonshot have already put pressure on rivals’ pricing. Z.ai is also thought to close to releasing a larger version of GLM-5.3 that many believe will rival some of the best models from Anthropic and OpenAI in cyber capabilities, potentially a watershed event that many fear will usher in a new era of AI-powered cyber attacks for which industry is woefully unprepared.

China is now clearly ahead in the open-model race. For example, Hugging Face found that, in almost every month of 2026, the largest and most capable open model came from a Chinese lab, while the U.S.’s most notable recent challengers have come from Thinking Machines Lab, and, more recently, Meta. But those American releases have generally not matched Kimi K3’s scale or developer pull.

Alex Brunicki, a partner at Backed VC, told me he’s already seeing a shift in how enterprises are approaching open source models.

“We’ve seen a lot of companies…developing industry‑specific foundation models using open source models that are then fine tuned on very particular data sets,” he said. “They’re not necessarily using the frontier models for all of the work that they’re doing. They’re actually using these open source models which are free to use.”

At least two established organizations have recently publicly switched to open-source for some areas of their business. Thomson Reuters said this week it has built an in-house model, called Thomson-1, based on Snowdon, a system the company developed by adapting Alibaba’s open-source Qwen model.

The model will handle document-review tasks that previously ran on Claude. As part of the announcement, CTO Joel Hron said companies do not need ever-larger, more expensive models to get useful results, and that starting from a strong open foundation and specializing it deeply can produce capable AI at lower cost.

Harvey, the legal tech firm backed by OpenAI, Sequoia, and Andreessen Horowitz, also recently announced that its new model, Harvey Tenet, was post-trained on top of Moonshot AI’s open-weight Kimi K3 and that it outperformed both its base model and U.S. frontier systems, including Fable 5 and GPT-5.6 Sol, on complex legal agentic tasks. Harvey had previously built its product by customizing closed models from Anthropic, OpenAI, and Google.

Ramp’s lead economist, Ara Kharazian, wrote that recent growth of open-source has not yet dented spending on OpenAI or Anthropic directly. New AI buyers are still choosing the established American labs; for example, Anthropic gained the most ground among businesses in July, rising 1.1 percentage points to 43.5% market share, and OpenAI climbed just 0.23 points to 39.7%.

However, Anthropic’s Fable 5, thought to be the most advanced model on the market—so much so that the U.S. government briefly suspended foreign access to it for national security purposes— accounted for just 6% of tokens businesses purchase from Anthropic and 11.4% of dollars spent on Anthropic models overall, despite being priced at roughly $10 per million tokens, twice the cost of OpenAI’s GPT-5.6 Sol.

That model, by comparison, makes up 25% of OpenAI’s tokens and 23% of its spend. Kharazian argues Fable 5 has effectively found the market’s ceiling, and businesses are not willing to pay a premium for the best model on the market when a cheaper one is good enough.

Taken together, all this may suggest American frontier labs have found the limit on what customers will pay for the newest, most expensive model, and open-source may just be stepping in to fill that gap.

With that, here’s more AI news.

[email protected]
@beafreyanolan

FORTUNE ON AI

OpenAI, independent firms publish reports on rogue AI attack on Hugging Face. Here are the main takeaways—and what OpenAI still hasn’t disclosed. — by Emily Forlini

Anthropic wants investors to believe its ‘total addressable market’ is worth $30 trillion—nearly the size of the entire US economy — by Beatrice Nolan

Google DeepMind is losing its grip on elite AI talent, new data shows — by Beatrice Nolan

Nvidia doubles quarterly revenue to $96 billion and crushes Wall Street growth targets as CEO Jensen Huang says demand is accelerating — by Amanda Gerut

Nvidia agrees to buy Hugging Face for $12.9 billion, reports — by Beatrice Nolan

AI IN THE NEWS

DeepSeek is seeking to raise 50 billion yuan, or about $7.4 billion, in a second funding round that would value the Chinese AI startup at 500 billion yuan ($74 billion), according to a report in The Information. The company generated 475 million yuan, or $70.7 million, in revenue in the first seven months of 2026—roughly 10 times its full-year revenue in 2025—but still recorded a 715 million yuan net loss. Its API business had an 82.9% gross margin, while overall gross margin was 44.6%. DeepSeek has also hired investment banks to prepare for a potential Shanghai IPO next year, the report said.

AI researcher Barret Zoph announced on social media he was joining Google DeepMind, where he will work on reinforcement learning and post-training. Reinforcement learning and post-training have become central to the push to make frontier models better at reasoning, coding, and carrying out multistep tasks. Zoph will be the vice president of research at the company. He began his AI career in Google Brain’s residency program, and on X described the move as a return to the company. Zoph was a long-standing OpenAI employee, who then left to co-founder of Thinking Machines with former OpenAI CTO Mira Murati. He then abruptly defected from there back to OpenAI earlier in the year following a dispute with Murati. Read more in The Wall Street Journal.

Moonshot AI is in talks with Microsoft, Amazon, and Google on revenue-sharing agreements that would allow the U.S. cloud providers to host its Kimi K3 model, Reuters reported. The Chinese startup is seeking up to 30% of revenue generated by K3-related services on Azure, AWS, and Google Cloud, according to people familiar with the discussions. The talks are still in their early stages, and there is no guarantee of agreements. Revenue splits, data access, and auditing token usage are among the unresolved issues, per the report.

Chris Malone, OpenAI’s head of data centers, has left the company. Malone joined OpenAI in 2025 to help oversee its data-center buildout, including work connected to the company’s Stargate infrastructure plans. An OpenAI spokesperson said the company had recently reorganized its infrastructure organization “to support the scale and pace of our work.” The departure comes amid a broader exodus of senior OpenAI executives; in the last few months, the company has lost COO Brad Lightcap, who announced he’s leaving after eight years to start a new venture; Fidji Simo, who oversaw much of OpenAI’s core business and stepped down after taking a medical leave; chief revenue officer Denise Dresser, who exited after less than a year in the role; and former chief product officer Kevin Weil, who left in April. Read more in the Wall Street Journal.

EYE ON AI NUMBERS

700

That’s roughly how many AI agents participated in the seven-day cyberattack on Hugging Face involving OpenAI models, according to an independent review by the Model Evaluation and Threat Research organization and Redwood Research. Around 1,200 agents overall exchanged more than 70,000 messages and files through an unauthorized message board, coordinating ways to game a hacking evaluation. Some agents reportedly pursued dead-end approaches to generate information that could help the broader group.

OpenAI said the incident was a “warning shot,” showing that capable agents can bypass technical controls, collaborate through unapproved channels, and take dangerous actions without a human directing them. Read more in Fortune.