China and the U.S. are in a race to create the first grid-scale nuclear fusion energy. After decades of U.S. leadership, China is catching up by spending twice as much and building projects at record speed.
Often called the holy grail of clean energy, nuclear fusion creates four times more energy per kilogram of fuel than traditional nuclear fission and four million times more than burning coal, with no greenhouse gasses or long-term radioactive waste. If all goes to plan, it will be at least a $1 trillion market by 2050, according to Ignition Research.
There’s just one big problem.
“The only working fusion power plants right now in the universe are stars,” said Dennis Whyte, professor of nuclear science and engineering at Massachusetts Institute of Technology.
The U.S. was first to large-scale use of fusion with a hydrogen bomb test in 1952. In the seven decades since, scientists around the world have been struggling to harness fusion reactions for power generation.
Fusion reactions occur when hydrogen atoms reach extreme enough temperatures that they fuse together, forming a super-heated gas called plasma. The mass shed during the process can, in theory, be turned into huge amounts of energy, but the plasma is hard to control. One popular method uses powerful magnets to suspend and control the plasma inside a tokamak, which is a metal donut-shaped device. Another uses high-energy lasers, pointed at a peppercorn-sized pellet of fuel, rapidly compressing and imploding it.
That’s how the U.S. pulled off the historic first fusion ignition, producing net positive energy at the Lawrence Livermore National Ignition Facility, or NIF, in 2022.
Here, the preamplifier module increases the laser energy as it heads toward the target chamber at the National Ignition Facitility.
Photo courtesy Damien Jemison at Lawrence Livermore National Laboratory
Since then, private investment in U.S. fusion startups has soared to more than $8 billion, up from $1.2 billion in 2021, according to the Fusion Industry Association. Of the FIA’s 40 member companies, 25 of them are based in the U.S.
Traditional nuclear power, created from fission instead of fusion, has seen a big uptick in investment as Big Tech looks for ways to fill the ever-increasing power needs of AI data centers. Amazon, Google and Meta have signed a pledge to help triple nuclear energy worldwide by 2050.
“If you care about AI, if you care about energy leadership … you have to make investments into fusion,” FIA CEO Andrew Holland said. “This is something that if the United States doesn’t lead on, then China will.”
Despite breaking ground on its first reactor nearly four decades after the U.S. pioneered the tech, China’s now building far more fission power plants than any other country.
China entered the fusionrace in the early 2000s, about 50 years after the U.S., when it joined more than 30 nations to collaborate on the International Thermonuclear Experimental Reactor fusion megaproject in France. But ITER has since hit major delays.
The race is on between individual nations, but the U.S. private sector remains in the lead. Of the $8 billion in global private fusion investment, $6 billion is in the U.S., according to the FIA.
Commonwealth Fusion Systems, a startup born out of MIT, has raised the most money, nearly $2 billion from the likes of Bill Gates, Jeff Bezos and Google.
Washington-based Helion has raised $1 billion from investors like Open AI’s Sam Altman and a highly ambitious deal with Microsoft to deliver fusion power to the grid by 2028. Google-backed TAE Technologies has raised $1.2 billion.
“Whoever has essentially abundant limitless energy … can impact everything you think of,” said Michl Binderbauer, CEO of TAE Technologies. “That is a scary thought if that’s in the wrong hands.”
When it comes to public funding, China is way ahead.
Beijing is putting a reported $1.5 billion annually toward the effort while U.S. federal dollars for fusion have averaged about $800 million annually the last few years, according to the Energy Department’s Office of Fusion Energy Sciences.
President Donald Trump ramped up support for nuclear, including fusion, during his first term, and that continued under former President Joe Biden. It’s unclear what fusion funding will look like in Trump’s second term, amid massive federal downsizing.
U.S. senators and fusion experts published a report in February calling for $10 billionof federal funds to help keep the U.S. from losing its lead.
But the U.S. may already have lost the lead when it comes to reactor size. Generally, the bigger the footprint, the more efficiently a reactor can heat and confine the plasma, increasing the chances for net positive energy.
A satellite image from January 11, 2025, shows a massive nuclear project in Mianyang, China, that appears to include four laser bays pointing at a containment dome roughly the size of a football field, about twice as big as the U.S. National Ignition Fusion Facility.
Planet Labs PBC
A series of satellite images provided to CNBC by Planet Labs shows the rapid building in 2024 of a giant new laser-fusion site in China. The containment dome where the fusion reaction will occur is roughly twice the size of NIF, the U.S. laser-fusion project, CNA Corporation’s Decker Eveleth said. The China site is likely a fusion-fission hybrid, FIA’s Holland said.
“A fusion-fission hybrid essentially is like replicating a bomb, but as a power plant. It would never work, never fly in a place like the United States, where you have a regulatory regime that determines safety,” Holland said. “But in a regime like China, where it doesn’t matter what the people who live next door say, if the government says we want to do it, we’re going to do it.”
China’s existing national tokamak project, EAST, has been setting records, volleying with France’s project WEST in the last couple months for the longest ever containment of plasma inside a reactor, although that’s a less monumental milestone than net positive energy.
Another huge state-funded Chinese project, CRAFT, is set to reach completion this year. The $700 million 100-acre fusion campus in eastern China will also have a new tokamak called BEST that is expected to be finished in 2027.
China’s CRAFT appears to follow a U.S. plan published by hundreds of scientists in 2020, Holland said.
“Congress has not done anything to spend the money to put this into action,” he said. “We published this thing, and the Chinese then went and built it.”
U.S. fusion startup Helion told CNBC some Chinese projects are copying its patented designs, too.
“China, specifically, we’re seeing investment from the state agencies to invest in companies to then replicate U.S. companies’ designs,” said David Kirtley, founder and CEO of Helion.
In the early 2000s, budget cuts to domestic fusion research forced U.S. universities to halt work on new machines and send researchers to learn on other country’s machines, including China’s.
“Instead of building new ones, we went to China and helped them build theirs, thinking, ‘Oh, that’d be great. They’ll have the facility. We’ll be really smart,'” said Bob Mumgaard, co-founder and CEO of Commonwealth Fusion Systems. “Well, that was a big mistake.”
China now has more fusion patents than any other country, and 10 times the number of doctorates in fusion science and engineering as the U.S., according to a report from Nikkei Asia.
“There’s a finite labor pool in the West that all the companies compete for,” Binderbauer said. “That is a fundamental constraint.”
Commonwealth Fusion Systems SPARC tokamak being assembled in December 2024 in Devens, Massachusetts, is scheduled to use superconducting magnets to reach fusion ignition in 2027.
Commonwealth Fusion Systems
Besides manpower, fusion projects need a huge amount of materials, such as high power magnets, specific metals, capacitors and power semiconductors. Helion’s Kirtley said the timeline of the company’s latest prototype, Polaris, was set entirely by the availability of semiconductors.
China is making moves to corner the supply chain for many of these materials, in a similar play to how it came to dominate solar and EV batteries.
“China is investing ten times the rate that the United States is in advanced material development,” Kirtley said. “That’s something we have got to change.”
Shanghai-based fusion company Energy Singularity told CNBC in a statement that it “undoubtedly” benefits from China’s “efficient supply chain.” In June, Energy Singularity said it successfully created plasma in record time, just two years after beginning the design of its tokamak.
That’s still a far cry from reaching grid-scale, commercial fusion power. Helion aims to be first with a goal of 2028. Commonwealth has announced the site in Virginia where it plans to bring the first fusion power plant, ARC, online in the early 2030s.
“Even though the first ones might be in the U.S., I don’t think we should take comfort in that,” said MIT’s Whyte. “The finish line is actually a mature fusion industry that’s producing products for use around the world, including in AI centers.”
Chinese tech giant Baidu has released two new free-to-use artificial intelligence models as it vies to regain its leading position in the country’s fiercely competitive AI space.
The Baidu models launched Sunday included the company’s first reasoning-focused model, and come ahead of plans to move toward an open-source strategy.
However, experts told CNBC that while the release of the models is a positive development for Baidu, they also highlight how it is playing catch up as its Ernie bot — one of China’s earliest versions of a ChatGPT-like chatbot — struggles to gain widespread adoption.
“The new models make Baidu more competitive since the company has been lagging behind in a reasoning model release,” Lian Jye Su, chief analyst at Omdia, told CNBC.
A reasoning model is a large language model that breaks down tasks into smaller pieces and considers multiple approaches before generating a response. It is designed to process complex problems in a similar way to humans.
Chinese startup DeepSeek upended the global AI race and transformed China’s ecosystem in January when it released its R1 reasoning model, which rivaled American competitors despite costing a fraction of the price.
Baidu has said its new ERNIE X1 reasoning model “delivers performance on par with DeepSeek R1 at only half the price,” and has “stronger understanding, planning, reflection, and evolution capabilities.” CNBC has not been able to independently verify this claim.
According to Wei Sun, principal analyst of artificial intelligence at Counterpoint Research, Baidu’s future competitiveness could hinge on whether its new models deliver on the promised performance and cost advantages.
“Baidu is clearly in catch-up mode, largely due to its slow innovation pace and underestimating rapid shifts in market dynamics,” Sun said.
Experts list a number of reasons for Baidu’s struggles and slow rate of innovation.
“Baidu fell behind when they tried to build proprietary models and compete for funding for AI,” Ray Wang, principal analyst and founder of Constellation Research, told CNBC. He added that the company has also suffered from recent government crackdowns and was distracted by “regulatory nonsense.”
CFOTO | Future Publishing | Getty Images
Proprietary models keep their source code and underlying architecture confidential, in contrast to models from the likes of DeepSeek, whose source code is made freely available on the open web for possible modification and redistribution.
“Using a closed-source approach means that [Baidu] was training its model from scratch whereas the open-source models were able to leverage certain parts that were communal to developers,” said Kai Wang, a senior equity analyst for Morningstar.
Baidu, however, said last month that it would make its next-generation AI model Ernie open-source from June 30, according to Reuters.
“Baidu has always been very supportive of its proprietary business model and was vocal against open source, but disruptors like DeepSeek have proven that open source models can be as competitive,” said Omdia’s Su.
He added that Baidu is “merely following the footstep” of its biggest competitors in China, namely Alibaba, DeepSeek, and Tencent, which have all now released open-source models.
During an earnings call late last year, Baidu CEO Robin Li said he expects generative AI to turn Baidu search into the “new killer app in the age of AI.”
“The secret to AI is data, chips, good math, and cheap energy … You have to have data to play. They have the data,” said Constellation Research’s Wang.
Correction: This article has been updated to state that Baidu’s new AI models are free to use.
ZHEJIANG, CHINA – MARCH 16 2023: A view of the logo of ERNIE Bot, an AI chatbot service developed by Chinese search engine Baidu, March 16, 2023.
Long Wei | Future Publishing | Getty Images
Shares of Chinese tech giant Baidu were trading up 10.7% in Asia on Tuesday, as investors appeared to react positively to the release of two new AI models over the weekend.
Baidu released two new artificial intelligence models on Sunday, including the latest version of its foundational “Ernie” model and a new reasoning model that it said rivaled DeepSeek’s R1 model. CNBC is not able to verify these claims.
A reasoning model is a large language model designed to process complex problems in a similar way to humans, breaking prompts down into smaller pieces and considering multiple approaches before generating responses.
According to Kai Wang, a senior equity analyst for Morningstar, the stock jump is likely a “delayed reaction” to the new models as Baidu vies to regain a leading position in China’s AI space.
“The stock also hasn’t gotten as much love as the other hyperscalers but still it’s a platform that stands to benefit from greater AI demand since enterprises will need someone to help them with hosting, scaling, and computing power,” he said.
A hyperscaler refers to a major cloud computing company that provides massive data centers for computing storage and demands.
Baidu said on Sunday that its ERNIE X1 reasoning model “delivers performance on par with DeepSeek R1 at only half the price,” and has “stronger understanding, planning, reflection, and evolution capabilities.”
Chinese AI start-up DeepSeek upended the AI industry in January when it released its R1 open-source reasoning model, which rivaled models of American competitors, despite claims it was produced at a fraction of the cost and with far less powerful chips.
DeepSeek quickly overtook Baidu in China’s AI race, despite the company being one of the first in the market to launch a ChatGPT-like chatbot with its Ernie Bot, according to Wei Sun, principal analyst of artificial intelligence at Counterpoint Research, who noted other tech giants like Alibaba and Bytedance have also pulled ahead.
“Baidu’s competitiveness hinges on whether its new models truly deliver on the promised performance and cost advantages,” Sun said, noting, however, that AI pricing, particularly in China’s market, is highly fluid.
Baidu’s latest models, similarly to DeepSeek’s R1, have been released as open-sourced, meaning the source code is freely available on the open web for possible modification and redistribution.
This represents a change from Baidu’s prior strategy of focusing on proprietary models.
“By open-sourcing its models, Baidu seeks to once again position its technology as an industry standard, strengthening its influence in the AI community and expand its market share,” said Sun.
Co-founder & CEO of Rippling Parker Conrad speaks onstage during the TechCrunch Disrupt conference in San Francisco on Oct. 20, 2022
Kimberly White | TechCrunc | Getty Images
Human resources software startup Rippling sued competitor Deel in federal district court on Monday, claiming that “Deel cultivated a spy” to orchestrate a trade-secret theft.
The employee met with Deel executives and passed internal Rippling records to a reporter, according to San Francisco-based Rippling’s complaint in the U.S. District Court for California’s Northern District.
Rippling claimed in the filing Deel violated the 1970 Racketeer Influenced and Corrupt Organizations Act and misappropriated trade secrets.
The two startups are among the most world’s most valuable. Investors valued Rippling at $13.5 billion in a funding round announced last year, while Deel told media outlets in 2023 that it was worth $12 billion. Deel ranked No. 28 on CNBC’s 2024 Disruptor 50 list.
“Weeks after Rippling is accused of violating sanctions law in Russia and seeding falsehoods about Deel, Rippling is trying to shift the narrative with these sensationalized claims,” a Deel spokesperson told CNBC in an email. “We deny all legal wrongdoing and look forward to asserting our counterclaims.”
Rippling confirmed its findings earlier this month. The company’s general counsel sent a letter to three Deel executives that referred to a new Slack channel, and the Deel spy quickly looked for it. Rippling subsequently served a court order to the spy at its office in Dublin, Ireland requiring him to preserve information on his mobile phone.
“Deel’s spy lied to the court-appointed solicitor about the location of his phone, and then locked himself in a bathroom — seemingly in order to delete evidence from his phone — all while the independent solicitor repeatedly warned him not to delete materials from his device and that his non-compliance was breaching a court order with penal endorsement,” Rippling said in Monday’s filing. “The spy responded: ‘I’m willing to take that risk.’ He then fled the premises.”
Rippling hired the person whom it calls the Deel spy for a management role in 2023, as the two companies were becoming more competitive, the filing says. Deel had used Rippling’s software, but Rippling opted to not renew Deel’s contract, according to the legal filing.
The spy repeatedly accessed information about Rippling customers, quotes, sales calls, demos and support requests in internal Slack repositories, according to the filing. He found and downloaded Rippling’s guidance on how to go up against Deel for prospective business, too, the filing says.
Then, in February, a reporter at The Information sent an inquiry to Rippling that included Slack messages from inside Rippling, which the startup concluded were collected by the Deel spy, the filing says. Additionally, email records suggest that the spy met with Deel executives in December, Rippling said in the complaint.
“We always prefer to win by building the best products and we don’t turn to the legal system lightly,” Parker Conrad, Rippling’s co-founder and CEO, said in a Monday X post. “But we are taking this extraordinary step to send a clear message that this type of misconduct has no place in our industry.”
This isn’t Conrad’s first legal entanglement over data access. In 2015, ADP dropped a defamation lawsuit that claimed his previous HR startup, Zenefits, had obtained information from clients in order to provide them with payment processing services.
