ByteDance's $4B AI revenue leads China as it restructures Lark into Doubao chatbot division

ByteDance's plan to dominate AI

During the summer of 2023, inside a converted brick warehouse in downtown Shanghai, dozens of ByteDance employees embarked on a secret mission. The project had no name at the time and its engineers only knew they were part of the "AI innovation" unit. Their first task was to build the Chinese company's next super-app, an AI chatbot for the masses akin to OpenAI's ChatGPT. ByteDance had once before broken the dominance of global tech giants with its addictive short-video app Douyin, which later became TikTok in overseas markets. The Shanghai unit spent two months building a prototype chatbot, drawing on vast computing resources and experience building consumer applications. Eventually, the prototype became Doubao, which translates to "bean bun". Two years on, Doubao is China's most popular AI app, with 324mn monthly users. It is now so ubiquitous that for some Chinese consumers the word "doubao" has become synonymous with AI itself. Yet ByteDance founder Zhang Yiming has bigger ambitions. As well as designing consumer-facing apps, Zhang wants ByteDance to become China's leading company across all aspects of AI, from frontier models and video generation to cloud infrastructure and custom chips -- the components of what Nvidia founder Jensen Huang has since called the "five-layered AI cake". Over the past three years, ByteDance has accordingly invested in AI more aggressively than any of the other Chinese tech giants, building out its network of data centres and hiring researchers. It has doubled down on its cloud unit, Volcano Engine, which sells AI solutions to enterprises, and buttressed its video generation capabilities with its model Seedance. It also has ambitions to develop a custom AI chip. There is a broader, deepening investor optimism about China's advances in AI, particularly with the latest models from Moonshot AI and the blockbuster initial public offering of memory-chip maker CXMT. The value of ByteDance's shares on the secondary market has risen to up to $600bn, a 50 per cent jump from last year. But investors and analysts say ByteDance's metamorphosis represents a big gamble at a moment when some are questioning the vast sums being spent on AI. Semiconductor stocks have plunged in recent weeks over concerns about oversupply and the sustainability of the AI boom. Much like Meta in the US, ByteDance has long considered itself a social media company with a cash-spinning platform that sells consumer attention to advertisers. The majority of its $186bn in annual revenue last year was from advertising, with a small but growing portion from ecommerce. Past deviations from this core business have not always worked out, including previous forays into gaming and education technology. Nonetheless, ByteDance is now ploughing proceeds into AI infrastructure, hardware and software as it raises money to expand into new areas such as chips. "Over time, ByteDance will look less like a social media company and more like an AI platform company with consumer distribution at massive scale," says Charlie Dai, a tech analyst at consultancy Forrester. There are considerations beyond capital expenditure. Like other large tech companies operating in a highly competitive market, including Alibaba, Tencent, Meta and Google, ByteDance has struggled to retain top talent. It must also balance the twin priorities of commercialisation and frontier research. "ByteDance is now spending pretty much every dollar they make on AI, because it's the next big thing and they can't afford to be left behind," says an early investor in the company who asked not to be named. "It's a big bet." Modest beginnings Fourteen years ago, ByteDance consisted of a dozen or so staff in a Beijing apartment so small that if they stretched, they touched each other. Sharing a dining table with his team in the living room, Zhang wrote the first lines of code that later evolved into ByteDance's secret sauce of success: algorithms that micro-analysed user behaviour and precisely predicted their real interest, keeping them hooked on apps. Zhang opted to keep headquarters in an apartment because Beijing's office buildings typically shut down air conditioning systems at night, the most productive time for coders. Not long before, Zhang had convinced an angel investor to put $80,000 into the start-up. The pitch meeting took less than 10 minutes, according to a person with direct knowledge, as the then 29-year-old entrepreneur drew his idea on a napkin. One of its first successful products was a news aggregator called Toutiao, or "top headline" in English. Two years later, ByteDance's valuation exceeded $500mn. In 2016 it launched a short video-sharing platform, Douyin, based on the same recommendation algorithm. A year later, TikTok, the international version, was born. The surging popularity of Douyin and TikTok allowed ByteDance to enter ecommerce among other internet services. It became one of the most profitable Chinese tech companies, surpassing Alibaba and Tencent in 2023. Zhang, who became one of China's richest people with wealth exceeding $50bn, moved his team to fancier and larger office buildings along the way. But with each move he cautioned his team against "large corporate disease", meaning bureaucracy and politics. "ByteDance had and still has very much a do-er mentality -- your work should speak for you," says one veteran employee. Following the release of OpenAI's ChatGPT in late 2022, Zhang decided that ByteDance needed to build its own version of a consumer AI chatbot, assigning the task to the Shanghai unit. This was the first plank in a longer-term AI strategy. In an effort to win over users, the unit added voice functions so that children and older people who do not type can still interact with the chatbot and may feel like they are talking to a friend. To find a voice for its avatar, Doubao's engineers screened more than 100 voice samples from men and women and ran tests over a month. They found that a young woman's voice tested the best, giving rise to the cartoon-style girl synonymous with the Doubao app. One person involved in training Doubao says they chose a voice in keeping with a character who is "gentle and sweet". "We found that men do not want to interact with a voice that is too authoritative," they say. The chatbot took off with users, becoming the market leader. "Doubao has become pretty much just another word for AI in our family," says Jimmy Zhu, a 37-year-old electric engineer from Shanghai. "My seven-year-old son calls Doubao instead of us whenever he has a question, and my parents in their seventies use Doubao for just about everything -- planning trips, asking what to do when they are not feeling well." Keen to build on this initial success, Zhang, who closely follows the latest academic papers and trends, decided in early 2024 that ByteDance needed to strengthen its research operations and improve its model-training capabilities, according to people familiar with his thinking. This in turn would help improve the Doubao app, with models trained to deliver better intelligence while using fewer computing resources. At the time, ByteDance did have an AI research team, but it was dwarfed by larger rivals Alibaba and Baidu and was seen as peripheral to the company. ByteDance formulated a plan to buy more Nvidia cards for AI training, while Zhang personally oversaw a hiring spree for top research talent. Zhang's resolution hardened with DeepSeek's launch of its open-sourced reasoning model R1 in January 2025 -- proving that a singular focus on AI research brought breakthroughs that could be parlayed into products. "DeepSeek overtook Doubao almost overnight in daily active users," says a former Doubao employee. "We built a great product with lots of functions and nice user interfaces. DeepSeek had just one boring web page. It was disheartening and a wake-up call for us that no matter how good your product looks, the underlying model decides how smart it really is." ByteDance acted quickly, hiring Wu Yonghui, a senior scientist from Google DeepMind, in February 2025 to lead its foundation model team Seed. Wu expanded his Seed team to its current size of 2,000 -- one of the largest Chinese AI labs -- with some members in Singapore and California but the bulk of them in China. The team includes core researchers, infrastructure engineers, a data labelling team and translators. "ByteDance is hiring aggressively not only in China but also in Silicon Valley," says a person close to its San Francisco-based team. When ByteDance's HR caught wind that a senior engineer was being interviewed by Alibaba, they hastily arranged a call between him and Zhang and put an offer on the table the next day, according to a person with knowledge of the matter. While ByteDance has what insiders call a "high talent density" in the Seed team, it has still seen senior researchers depart to found start-ups, join fast-growing AI labs or jump over to Tencent -- which raided its ranks last year. One former Seed member said that Tencent was offering more responsibilities in a smaller team, and that ByteDance had grown so large with overlapping research functions that it was more competitive to gain ownership of projects. US tech companies including Meta and Google have similarly seen talent drain away to Anthropic and OpenAI, as researchers and engineers bet they can do more impactful work in a fast-moving start-up as well as picking up stock. Zhang stepped down as ByteDance CEO in 2021, but remains deeply involved in the company's strategy, especially in AI. He personally spends a lot of time with the Seed team, getting into technical discussions with its engineers, according to multiple former and current employees. "He is modest and asks good questions," says one of them. Head in the cloud The most ambitious plank of ByteDance's AI strategy is its pivot to selling AI solutions to companies and organisations through Volcano Engine. ByteDance has previously tried to launch enterprise-focused businesses, notably with Lark, its internal communication and collaboration tool. But it has not reached the same scale and success as its consumer apps. Founded in 2020, ByteDance's cloud unit Volcano Engine was a new entrant in an already saturated market where Tencent, Baidu, Huawei and Alibaba all offer business services, in addition to state-owned cloud companies such as China Mobile. ByteDance has gained market share by offering cut-throat pricing on its fleet of products and models. Its most competitive offering is Seedance 2.0, which it sells to advertising agencies, media groups and AI content creators. The product has been praised by users for its high-quality cinematic shots with complex framing and easy-to-use editing capabilities. ByteDance faces competition in video generation from Alibaba and Kling, owned by Kuaishou, another short-video company in China. But analysts say the company has an edge that is difficult for other companies to replicate. Training video AI models requires vast amounts of high-quality footage, data that ByteDance has in abundance through its ownership of short-video apps such as TikTok. Since video content, unlike text, cannot easily be scraped at scale, it is difficult for competitors to catch up. Sales of the Seedance model to media companies represent a significant portion of Volcano Engine's overall revenues, according to people familiar with the matter. Despite its relative inexperience in cloud, ByteDance has become China's second-largest provider of AI infrastructure and software, according to market intelligence firm IDC. ByteDance accounted for 16 per cent of China's AI cloud services revenue, which totalled $1bn in 2025, behind only Alibaba. However, Volcano Engine's share of the overall cloud market in China is less than 5 per cent. A key advantage Volcano Engine has over other Chinese cloud providers is its ability to expand overseas, leveraging ByteDance's operational knowhow in more than 100 countries gathered from the past decade. ByteDance has, for example, a dedicated team that is working with data centre operators to build new facilities in Malaysia, Singapore, Brazil, Ireland and Norway. This infrastructure will not only support ByteDance's businesses on the ground but also provide cloud and model services to local clients, according to a person with knowledge of the plans. Yet some question whether ByteDance can broaden its cloud sales beyond the sectors it is most familiar with, such as media. In finance and tech, Alibaba still has the upper hand, analysts say. ByteDance's models have also lagged behind in AI coding -- an area that it has prioritised to improve given the booming enterprise demand for coding tools. "ByteDance doesn't have as much industry knowhow compared to AliCloud, which for over 15 years has served a wide range of enterprise customers," says Edison Lee, head of China tech research at investment bank Jefferies. "ByteDance has been lowering prices to gain market share, but this strategy doesn't work for large enterprises who aren't just motivated by prices." Another bottleneck for all Chinese AI companies is the lack of advanced computing power in the country. Doubao's popularity demonstrates the potential demand. But the more users the AI app has, the more computing power is needed, which is expensive and increasingly scarce in China due to US export controls. Doubao's daily token consumption has reached about 120tn, one of the highest globally among consumer AI apps. ByteDance is the best-positioned tech giant in China in terms of its raw computing capacity, having been an early and aggressive purchaser of Nvidia chips as well as domestic inference chips from Cambricon, say industry insiders. The company is pursuing a "portfolio strategy", says analyst Poe Zhao, as it seeks diverse AI chips from multiple sources, both inside China and overseas. ByteDance is a leading tenant in the rapidly growing network of data centres being built in Malaysia and Singapore. It is also designing its own AI chip for internal inference workloads. Whether ByteDance's audacious transformation succeeds may ultimately depend on whether it can secure enough computing resources. US export controls, which could be tightened, complicate the picture. "That is the constraint that binds everything else," Zhao says. Data visualisation by Haohsiang Ko

ByteDance tops China's AI revenue at $4bn, folds Lark into Doubao

ByteDance now makes about $4bn a year from AI, the largest figure any Chinese company has publicly claimed, according to the South China Morning Post. To press the lead, it is folding Lark, its Slack-style workplace app, into the teams behind its Doubao chatbot and its cloud business. It is one piece of a costly bet to lead China across every layer of AI. ByteDance is quietly winning a race in China that most people are watching in America. The company behind TikTok now makes about $4bn a year from artificial intelligence. That is the largest figure any Chinese firm has publicly claimed, the South China Morning Post reported. To press that lead, ByteDance is tearing up part of its own structure. It is folding Lark, its Slack-style workplace app known as Feishu at home, into the teams behind its Doubao chatbot and its cloud business. The plan comes from an internal memo seen by the paper. Absorbing the office app The reshuffle is telling. Lark's product team will merge into the group that runs Doubao, ByteDance's answer to ChatGPT, under Zhao Qi, who until now oversaw Doubao alone. Lark's sales and support staff move to Volcano Engine, ByteDance's cloud arm, to sell AI models and software to businesses. Its engineers join Flow, a division focused on AI apps. In plain terms, ByteDance is breaking up a standalone product and redistributing it around AI. The company said the goal was "stronger AI product competitiveness and customer service capabilities." The message to staff is clear. The chatbot and the cloud, not the office suite, are the future. Big in China, small next to the US The $4bn puts ByteDance ahead of China's better-known AI names. It is more than DeepSeek, Moonshot or Alibaba have disclosed. But it is still tiny beside the American leaders. OpenAI and Anthropic each run at many times that. The figure says more about the gap with US AI than about any Chinese lead. Money is the thing China's labs keep circling. Most give their best models away to win users. They are only now working out how to charge for them. A firm that can already book $4bn has a head start. A bet across the whole stack The restructure is one piece of a larger push. ByteDance is spending across every layer of AI at once, from frontier models and its Seedance video generator to infrastructure and chips. The Financial Times reported the effort in detail, and noted that some see it as a big gamble. The video bet is already visible. ByteDance's Seedance model is powering a wave of AI-made short dramas in China, compressing months of production into weeks. The research firm DataEye expects China's AI short-drama market to reach 40 billion yuan ($5.9bn) this year, with viewers already past 600 million. Set alongside its other AI experiments, it points to a company trying to be everywhere in AI at once. The Lark reshuffle is simply where that ambition met the org chart.