China is focusing on large language models (LLMs) in the artificial intelligence space.
Blackdovfx | Istock | Getty Images
China’s attempts to dominate the world of artificial intelligence could be paying off, with industry insiders and technology analysts telling CNBC that Chinese AI models are already hugely popular and are keeping pace with — and even surpassing — those from the U.S. in terms of performance.
AI has become the latest battleground between the U.S. and China, with both sides considering it a strategic technology. Washington continues to restrict China’s access to leading-edge chips designed to help power artificial intelligence amid fears that the technology could threaten U.S. national security.
It’s led China to pursue its own approach to boosting the appeal and performance of its AI models, including relying on open-sourcing technology and developing its own super-fast software and chips.
China is creating popular LLMs
Like some of the leading U.S. firms in the space, Chinese AI firms are developing so-called large language models, or LLMs, which are trained on huge amounts of data and underpin applications such as chatbots.
Unlike OpenAI’s models which power the hugely popular ChatGPT, however, many of these Chinese companies are developing open-source, or open-weight, LLMs which developers can download and build on top of for free and without stringent licensing requirements from the inventor.
On Hugging Face, a repository of LLMs, Chinese LLMs are the most downloaded, according to Tiezhen Wang, a machine learning engineer at the company. Qwen, a family of AI models created by Chinese e-commerce giant Alibaba, is the most popular on Hugging Face, he said.
“Qwen is rapidly gaining popularity due to its outstanding performance on competitive benchmarks,” Wang told CNBC by email.
He added that Qwen has a “highly favorable licensing model” which means it can be used by companies without the need for “extensive legal reviews.”
Qwen comes in various sizes, or parameters, as they’re known in the world of LLMs. Large parameter models are more powerful but have higher computational costs, while smaller ones are cheaper to run.
“Regardless of the size you choose, Qwen is likely to be one of the best-performing models available right now,” Wang added.
DeepSeek, a start-up, also made waves recently with a model called DeepSeek-R1. DeepSeek said last month that its R1 model competes with OpenAI’s o1 — a model designed for reasoning or solving more complex tasks.
These companies claim that their models can compete with other open-source offerings like Meta‘s Llama, as well as closed LLMs such as those from OpenAI, across various functions.
“In the last year, we’ve seen the rise of open source Chinese contributions to AI with really strong performance, low cost to serve and high throughput,” Grace Isford, a partner at Lux Capital, told CNBC by email.
China pushes open source to go global
Open sourcing a technology serves a number of purposes, including driving innovation as more developers have access to it, as well as building a community around a product.
It is not only Chinese firms that have launched open-source LLMs. Facebook parent Meta, as well as European start-up Mistral, also have open-source versions of AI models.
But with the technology industry caught in the crosshairs of the geopolitical battle between Washington and Beijing, open-source LLMs give Chinese firms another advantage: enabling their models to be used globally.
“Chinese companies would like to see their models used outside of China, so this is definitively a way for companies to become global players in the AI space,” Paul Triolo, a partner at global advisory firm DGA Group, told CNBC by email.
While the focus is on AI models right now, there is also debate over what applications will be built on top of them — and who will dominate this global internet landscape going forward.
“If you assume these frontier base AI models are table stakes, it’s about what these models are used for, like accelerating frontier science and engineering technology,” Lux Capital’s Isford said.
Today’s AI models have been compared to operating systems, such as Microsoft’s Windows, Google‘s Android and Apple‘s iOS, with the potential to dominate a market, like these companies do on mobile and PCs.
If true, this makes the stakes for building a dominant LLM higher.
“They [Chinese companies] perceive LLMs as the center of future tech ecosystems,” Xin Sun, senior lecturer in Chinese and East Asian business at King’s College London, told CNBC by email.
“Their future business models will rely on developers joining their ecosystems, developing new applications based on the LLMs, and attracting users and data from which profits can be generated subsequently through various means, including but far beyond directing users to use their cloud services,” Sun added.
Chip restrictions cast doubt over China’s AI future
AI models are trained on vast amounts of data, requiring huge amounts of computing power. Currently, Nvidia is the leading designer of the chips required for this, known as graphics processing units (GPUs).
Most of the leading AI companies are training their systems on Nvidia’s most high-performance chips — but not in China.
Over the past year or so, the U.S. has ramped up export restrictions on advanced semiconductor and chipmaking equipment to China. It means Nvidia‘s leading-edge chips cannot be exported to the country and the company has had to create sanction-compliant semiconductors to export.
Despite, these curbs, however, Chinese firms have still managed to launch advanced AI models.
“Major Chinese technology platforms currently have sufficient access to computing power to continue to improve models. This is because they have stockpiled large numbers of Nvidia GPUs and are also leveraging domestic GPUs from Huawei and other firms,” DGA Group’s Triolo said.
Indeed, Chinese companies have been boosting efforts to create viable alternatives to Nvidia. Huawei has been one of the leading players in pursuit of this goal in China, while firms like Baidu and Alibaba have also been investing in semiconductor design.
“However, the gap in terms of advanced hardware compute will become greater over time, particularly next year as Nvidia rolls out its Blackwell-based systems that are restricted for export to China,” Triolo said.
Lux Capital’s Isford flagged that China has been “systematically investing and growing their whole domestic AI infrastructure stack outside of Nvidia with high-performance AI chips from companies like Baidu.”
“Whether or not Nvidia chips are banned in China will not prevent China from investing and building their own infrastructure to build and train AI models,” she added.
Elon Musk listens as reporters ask U.S. President Donald Trump and South Africa President Cyril Ramaphosa questions during a press availability in the Oval Office at the White House on May 21, 2025 in Washington, DC.
Chip Somodevilla | Getty Images
Tesla shares gained about 5% on Tuesday after CEO Elon Musk over the weekend reiterated his intent to home in on his businesses ahead of the latest SpaceX rocket launch.
The billionaire wrote in a post to his social media platform X that he needs to be “super focused” on X, artificial intelligence company xAI and Tesla as they launch “critical technologies” on the heels of a temporary outage.
“As evidenced by the uptime issues this week, major operational improvements need to be made,” he wrote, adding that he would return to “spending 24/7” at work. “The failover redundancy should have worked, but did not.”
An outage over the weekend briefly shuttered the social media platform formerly known as Twitter for thousands of users, according to DownDetector. Earlier in the week, the platform suffered a data center outage. X has suffered a series of outages since Musk purchased the platform in 2022.
Read more CNBC tech news
Musk has previously indicated plans to step away from his political work and prioritize his businesses.
During Tesla’s April earnings call he said that he would “significantly” reduce his time running President Donald Trump‘s Department of Government Efficiency.
In the last election cycle, Musk devoted time and billions of dollars to political causes and toward electing Trump in 2024. However, a story over the weekend from the Washington Post, citing sources familiar with the matter, said that Musk has grown disillusioned with politics and wants to return to managing his businesses.
Last week, Musk said in an interview at the Qatar Economic Forum that he planned to spend “a lot less” on campaign donations going forward.
The comments from Musk precede SpaceX’s Starship rocket Tuesday evening. Pressure is on for the company after two Starship rockets exploded in January and March.
Ahead of the launch, Musk announced an all hands livestream on X at 1 p.m.
Tesla is still facing fallout from Musk’s political foray, with protests at showrooms and other brand damage.
In April, Tesla sold 7,261 cars in Europe, down 49% from last year, according to the European Automobile Manufacturers’ Association.
National Economic Council Director Kevin Hassett said Tuesday that the Trump administration does not want to “harm Apple” with tariffs.
“Everybody is trying to make it seem like it’s a catastrophe if there’s a tiny little tariff on them right now, to try to negotiate down the tariffs,” Hassett told CNBC’s “Squawk Box” on Tuesday. “In the end, we’ll see what happens, we’ll see what the update is, but we don’t want to harm Apple.”
Hassett’s comments come after President Donald Trump said in a social media post that Apple will have to pay a tariff of 25% or more for iPhones made outside the U.S. Apple has historically manufactured its products in foreign countries including China, India and Vietnam.
“I have long ago informed Tim Cook of Apple that I expect their iPhone’s that will be sold in the United States of America will be manufactured and built in the United States, not India, or anyplace else,” Trump wrote in the post. “If that is not the case, a Tariff of at least 25% must be paid by Apple to the U.S. Thank your for your attention to this matter!”
By some estimates, a U.S.-made iPhone could cost as much as $3,500.
Read more CNBC tech news
“If you think that Apple has a factory some place that’s got a set number of iPhones that it produces and it needs to sell them no matter what, then Apple will bear those tariffs, not consumers, because it’s an elastic supply,” Hassett said.
Hasset’s comments continue the administration’s push to pressure companies to shoulder the cost burden of Trump’s tariffs, instead of raising prices for consumers.
Earlier this month, Trump told retail giant Walmart to “EAT THE TARIFFS” after the company warned it would have to pass those added costs on.
Shares of Apple were up more than 1% Tuesday.
Apple did not immediately respond to CNBC’s request for comment.
Dr. Priti Patel, CMIO at John Muir Health, uses Ambience before starting a patient encounter.
Courtesy of Ambience Healthcare
Artificial intelligence startup Ambience Healthcare on Tuesday announced a new medical coding model that outperforms doctors by 27%.
Ambience uses AI to draft clinical notes in real-time as doctors consensually record their visits with patients. The company used tools from OpenAI to build the new model.
The startup is part of a fiercely competitive market that has taken off as health-care executives search for solutions to help reduce staff burnout and daunting administrative workloads.
The company’s new model can listen to patient encounters and identify ICD-10 codes, which are internationally standardized classifications for different diseases and conditions. There are about 70,000 ICD-10 codes that are regularly updated and used to facilitate billing and other reporting processes in health care.
Ambience said its new ICD-10 model can reduce billing mistakes and help clinicians and professional coders work more efficiently. The model notched a “27% relative improvement over physician benchmarks,” according to a release on Tuesday.
“We’re not replacing doctors or coders,” Brendan Fortuner, Ambience’s head of engineering, told CNBC in an interview.“What we’re doing is we’re liberating them from administration, and we’re fixing mistakes that help make health care better, safer, more cost-effective.”
More CNBC health coverage
Documenting ICD-10 codes has traditionally been a labor-intensive task in health care, but it’s a crucial way to track outcomes, mortalities and morbidities in a standardized way, said Dr. Will Morris, the chief medical officer of Ambience.
“If you think about it from a data perspective, it’s how you can compare and contrast clinician A to B, or health system A to B,” Morris said in an interview. “It’s the cornerstone for quality.”
Ambience’s technology is used at more than 40 health-care organizations, like Cleveland Clinic and UCSF Health. It has raised more than $100 million, according to PitchBook, from investors including Kleiner Perkins, Andreessen Horowitz and the OpenAI Startup Fund.
The company is reportedly seeking fresh capital at a valuation of over $1 billion, according to a report from The Information. Ambience declined to comment on the report.
Ambience trained its new AI model using OpenAI’s reinforcement fine-tuning technology. This technology allows companies to tune OpenAI’s best reasoning models for very specific domains, like health care.
To validate the model, Ambience tested it against a “gold panel” set of labels, the company said. The labels were established by a group of expert clinicians who evaluated complex clinical cases and came to an agreement on what the right codes were.
Ambience’s AI platform for compliant documentation, CDI, and coding.
Courtesy of Ambience Healthcare
The company then recruited 18 different board-certified doctors and compared their performance on ICD-10 coding accuracy to the model’s performance. That comparison showed the Ambience technology performed 27% better than the physician baseline.
“It shows for the first time that an AI system can actually surpass clinician experts at a very, very important administrative task, especially in coding,” Fortuner said.
Ambience already has similar capabilities available for other medical codes like Current Procedural Terminology (CPT) codes, and Fortuner said it’s exploring how to tackle other areas like prior authorizations, utilization management and clinical trial matching.
The company’s new ICD-10 model will roll out to customers over the summer.
“Getting it right at the point of care is a fundamental change,” Morris said.