ByteDance Planned to Spend $7 Billion on Nvidia Chips Next Year
The Takeaway
• ByteDance is talking to data center operators about accessing Nvidia Blackwell chips
• ByteDance operates the top consumer AI chatbot in China
• Its co-founder said capital expenditures for AI would surpass those of other Chinese tech giants
ByteDance continues to use Nvidia’s most advanced artificial intelligence chips despite U.S. efforts to block Chinese companies from using them.
The Chinese owner of TikTok recently told suppliers it planned to spend up to $7 billion to access Nvidia chips outside China in 2025, according to a person who has been involved in the plan. The effort, led by co-founder Zhang Yiming, would make ByteDance one of the world’s biggest consumers of such hardware.
Taking advantage of loopholes in the U.S. restrictions, ByteDance has been using Nvidia’s Hopper AI chips, which cannot be exported to China under the rules, at data centers outside China. And Zhang recently has been in talks with data center operators in Southeast Asia and elsewhere to access Nvidia's next-generation Blackwell chips when they become available next year, according to two people who have been involved in the company’s plans.
ByteDance’s moves show how difficult it’s been for the Biden administration to cut off China’s access to critical hardware for AI, as Chinese companies including ByteDance have accessed such chips outside China—including by renting them in the U.S.—without violating the U.S. restrictions. They also reflect the seriousness of Zhang’s push to make ByteDance a global AI powerhouse on par with rivals such as Meta Platforms, despite a looming TikTok ban in the U.S.
The U.S. is now working on new rules restricting the sale of advanced AI chips to countries in Southeast Asia and the Middle East, which could impact ByteDance’s efforts.
Hopper chips became available in late 2022 and quickly became a hot commodity, leading to shortages as AI developers including OpenAI, Anthropic and ByteDance began using them. Anticipation has been building for Blackwell chips, set to launch en masse next year, as AI firms develop bigger and bigger clusters of AI chips to train and run large language models and video models.
Capex Commitments
Zhang has been involved in some of the negotiations with suppliers for AI-focused data centers, which the company is leasing or building. Zhang told some suppliers ByteDance intends to spend more than any other Chinese tech giants on capital expenditures for AI—hardware, land and building construction—in the next two years.
He also said he will personally oversee these development projects while making sure the company has enough computing power to train and run its AI models, according to one person who is involved in the conversations. (American AI developers such as OpenAI from time to time have struggled to get enough servers to develop and power their products.)
A ByteDance spokesperson said some information in this article was “not correct,” without elaborating. An Nvidia spokesperson declined to comment.
Like Meta, ByteDance was taken aback by the leap OpenAI made two years ago with the conversational AI models that power ChatGPT. ByteDance has been developing its own models and applications, including video-generating ones, and now operates China’s biggest AI chatbot app by active users, Doubao, which is similar to ChatGPT.
The Doubao mobile app drew 60 million monthly active users last month, according to Aicpb.com, a Chinese website that tracks AI products. (OpenAI says ChatGPT has 300 million weekly users across web and mobile devices.) ByteDance has also been developing versions of its AI for markets outside China.
ByteDance has told some data center suppliers and banks it expects to spend more than $20 billion next year on AI chips, data centers and other kinds of hardware such as undersea cables, according to one of the people working with ByteDance and another person who’s involved in financing these projects. The company hasn’t yet finalized its 2025 budget for such spending, the two people said.
Nearly $7 billion of that budget is allocated to overseas cloud services giving the company access to Nvidia chips for AI, according to one of the people working with ByteDance. In comparison, Google has ordered over 400,000 GB200 chips, which are part of the upcoming Blackwell series and could cost more than $10 billion including server hardware. Similarly, Meta has made a substantial order for GB200s that would also cost at least $10 billion, including servers.
TikTok’s Billion Users
Chinese spending on AI hardware pales in comparison to spending by U.S. firms. Meta, which generates roughly the same revenue as ByteDance, projected $38 billion to $40 billion of capital expenditures this year, including for data centers and servers, or nearly a quarter of its projected revenue.
ByteDance leaders won’t hesitate to boost spending if the company’s rivals do, said one of the people working with ByteDance on data centers.
ByteDance is a major customer for a number of global data center operators, and those operators are planning to build new data centers in Southeast Asia and Europe in anticipation of strong AI computing demand from the Chinese tech giant over the next decade. TikTok has more than 1 billion monthly active users outside China.
ByteDance, for example, is the biggest customer of Chindata Group, a China-based data center operator backed by Bain Capital that is working on a $3 billion project to build new data centers in Malaysia—mainly to serve TikTok, according to a person involved in financing the project. Chindata didn’t respond to a request for comment.
Multiple European data center operators are also discussing plans to build new facilities whose main customers would include ByteDance, this person said. The talks with European data center operators are part of ByteDance’s plan to store TikTok’s European user data in local data centers.