Surveying the state of America’s artificial intelligence landscape earlier this year, Misha Laskin was concerned.
Laskin, a theoretical physicist and machine learning engineer who helped create some of Google’s most powerful AI models, saw a growing embrace among American AI companies of free, customizable and increasingly powerful “open” AI models.
But most of these models were being made in China, and these systems were quickly gaining ground on their U.S. competitors.
“These models were not that far behind the frontier. In fact, they were surprisingly close to the frontier. The ones that are coming now,” Laskin said, pausing slightly, “well they’re palpably close to the frontier.”
Laskin founded a startup called Reflection AI, recently valued at $8 billion, to provide an open-source American alternative to these increasingly capable Chinese models that have gained traction in Silicon Valley.
“You’re starting to see glimpses of open-model companies actually driving the frontier of intelligence in China, and overall, the frontier of intelligence,” Laskin said.
Over the past year, a growing share of America’s hottest AI startups have turned to open Chinese AI models that increasingly rival, and sometimes replace, expensive U.S. systems as the foundation for American AI products.
NBC News spoke to over 15 AI startup founders, machine-learning engineers, industry experts and investors, who said that while models from American companies continue to set the pace of progress at the frontier of AI capabilities, many Chinese systems are cheaper to access, more customizable and have become sufficiently capable for many uses over the past year.
The growing embrace could pose a problem for the U.S. AI industry. Investors have staked tens of billions on OpenAI and Anthropic, wagering that leading American artificial intelligence companies will dominate the world’s AI market. But the increasing use of free Chinese models by American companies raises questions about how exceptional those models actually are — and whether America’s pursuit of closed models might be misguided altogether.
Michael Fine, head machine learning at Exa, an AI-focused search company valued at $700 million and supported by Silicon Valley mainstays like Lightspeed Venture Partners and Nvidia, said running Chinese models on Exa’s own hardware has proved to be significantly faster and less expensive than using bigger models, like OpenAI’s GPT-5 or Google’s Gemini, in many cases.
“What often happens is we’ll get a feature working with a closed model and realize it’s too expensive or too slow, and we ask, ‘What levers do we have to make this faster and cheaper?’”
“That usually means replacing the closed model with the equivalent open model and then running it on our own infrastructure,” Fine said.
Chinese models, like DeepSeek’s R1 and Alibaba’s Qwen, are free to use and considered “open-source” or “open-weight” because anyone can download, copy, modify and operate them. They differ from leading American systems like Anthropic’s Claude or OpenAI’s most popular GPT models, which are “closed,” or proprietary, and accessed through data centers and pipelines controlled by the big tech giants.
For years, American closed-source models from OpenAI and Anthropic vastly outperformed both American and Chinese open alternatives. Even well-resourced in-house efforts to use open-source models struggled: Bloomberg tried to create an internal tool, BloombergGPT, using open-source models trained on its expansive collection of financial news and documents, only to see it trail OpenAI’s closed models on financial knowledge.
Yet in the past year, Chinese companies like DeepSeek and Alibaba have made huge technological advancements. Their open-source products now closely approach or even match the performance of leading closed American models in many domains, according to metrics tracked by Artificial Analysis, an independent AI benchmarking company.
“The gap is really shrinking,” Lin Qiao, CEO of Fireworks AI and co-creator of PyTorch, the dominant framework for training AI models, said of the capability differences between American closed-source and Chinese open-source models.
Who controls the future?
American AI companies and the federal government have noticed the recent rise of Chinese models, and experts have even labeled America’s lack of powerful open-source models an “existential” threat to democracy.
While Meta’s high-profile Llama series of open-source models has historically led American open-source efforts, CEO Mark Zuckerberg has signaled Meta’s intention not to open-source all of its “superintelligence” AI models. The performance of Llama models has also stalled in recent years, one of the reasons why open-source users have shifted to better-performing Chinese open-source models.
Yet American open-source efforts may be gradually awakening, as American innovators attempt to boost American open-model competitiveness.
In July, the White House released an AI Action Plan that called for the federal government to “Encourage Open-Source and Open-Weight AI.”
In August, ChatGPT maker OpenAI released its first open-source model in five years. Announcing the model’s release, OpenAI cited the importance of American open-source models, writing that “broad access to these capable open-weights models created in the US helps expand democratic AI.”
And in late November, the Seattle-based Allen Institute released its newest open-source model called Olmo 3, designed to help users “build trustworthy features quickly, whether for research, education, or applications,” according to its launch announcement.
Lambert, of the Allen Institute, has also launched the “ATOM Project” — an acronym for “American Truly Open Models.” As the ATOM Project’s manifesto declares: “America has lost its lead in open models — both in performance and adoption — and is on pace to fall further behind.”
“If we want to be the preeminent nation in the AI era, we cannot cede such a critical piece of the ecosystem to any nation,” Lambert told NBC News via email.
