ByteDance is preparing to spend about 100 billion yuan ($14 billion) on artificial intelligence chips from Nvidia in 2026, a sharp increase from roughly 85 billion yuan this year, according to people familiar with the matter who spoke to SCMP.
The proposed spending, which underscores both the scale of its AI ambitions and the strategic importance of access to U.S. chip technology, hinges on whether Nvidia is ultimately allowed to sell its H200 graphics processing units to Chinese customers, a decision that remains subject to regulatory approval in Beijing despite Washington recently easing restrictions on the chip.
The budget, sources cautioned, could still change depending on policy outcomes and market conditions.
Register for Tekedia Mini-MBA edition 19 (Feb 9 – May 2, 2026): big discounts for early bird.
Tekedia AI in Business Masterclass opens registrations.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab: From Technical Design to Deployment (next edition begins Jan 24 2026).
The planned outlay forms part of ByteDance’s broader push to dramatically expand capital expenditure on artificial intelligence next year. The Beijing-based company, which is privately valued at about $500 billion, is preparing to lift total AI investment to as much as 160 billion yuan in 2026, according to a Financial Times report earlier this month.
At the heart of the strategy is ByteDance’s surging demand for computing power across its sprawling ecosystem, which includes TikTok and Douyin, its rapidly growing cloud business Volcano Engine, and its expanding suite of large language models.
Volcano Engine, ByteDance’s enterprise cloud arm, is set to become the exclusive AI cloud partner for China Central Television’s Spring Festival Gala, the most-watched television broadcast in the country. The partnership highlights how deeply embedded ByteDance’s infrastructure has become in high-profile national platforms, further intensifying its need for reliable, large-scale AI compute.
The company’s chatbot, Doubao, illustrates the pace of that growth. Doubao is now processing more than 50 trillion tokens per day this month, up from just 4 trillion tokens in December 2024. Tan Dai, president of Volcano Engine, recently said the platform has served more than 100 corporate clients, collectively processing over 1 trillion tokens to date.
Alongside its reliance on Nvidia, ByteDance is working to reduce long-term exposure to supply constraints and geopolitical risk by building internal semiconductor capabilities. Its in-house chip design unit, which employs around 1,000 staff, has made progress on a processor that matches the performance of Nvidia’s H20 chip — a China-tailored product — but at a lower cost, according to one of the sources.
The H20 was designed by Nvidia to comply with earlier U.S. export controls before being blocked, and its absence has accelerated efforts by Chinese firms to develop domestic alternatives. ByteDance is not alone in this push. Other Chinese technology groups are pursuing similar strategies to gain greater control over chip supply, pricing, and long-term planning.
Beyond processors, ByteDance has also been investing in memory technologies, including high-bandwidth memory, through a combination of internal research and equity stakes in specialized start-ups. These components are critical for training and running large AI models efficiently, and shortages can be as constraining as limited access to GPUs.
Internally, ByteDance’s chip arm works closely with Seed, the company’s frontier AI research team, highlighting how semiconductor design is becoming increasingly central to core AI development rather than a peripheral function.
Reflecting growing geopolitical sensitivity, ByteDance transferred its chip unit to a Singapore-incorporated subsidiary, Picoheart, in September. The move came amid intensifying China–U.S. tensions over technology. Some employees based in mainland China have been asked about relocating to Singapore, particularly those working on projects considered more sensitive, according to staff familiar with the discussions.
On the supply side, Nvidia is racing to capitalize on pent-up Chinese demand. The company hopes to ship its newly approved H200 chips to Chinese customers before the Lunar New Year holiday in mid-February, Reuters has reported. Interest from Chinese technology groups has been strong, though Beijing has yet to approve any purchases, leaving timelines uncertain.
The H200 is Nvidia’s second-most powerful AI chip and a significant step up from products currently available to many Chinese firms. That performance gap explains why companies like ByteDance are willing to allocate tens of billions of yuan to secure access, even as domestic chipmakers gain ground.
Beijing, for its part, is attempting to strike a careful balance. While authorities want to accelerate AI development to support economic growth and technological leadership, they are also keen to promote the adoption of domestic chips from companies such as Cambricon Technologies, Huawei’s Ascend unit, Moore Threads Technology, and MetaX Integrated Circuits.
Whether ByteDance ultimately gets the green light to deploy Nvidia’s H200 at scale will be a key test of how that balance is managed — and a signal of how far China is willing to lean on foreign technology as it races to build its AI future.