(Yicai) April 20 -- ByteDance’s net profit plummeted more than 70 percent in 2025 from the year before as the Chinese operator of TikTok hiked investment across the entire artificial intelligence chain, underscoring the growing importance that leading Chinese tech firms are placing on AI as a critical driver of long-term competitiveness, Yicai learned from a source familiar with the matter.

ByteDance significantly increased expenditure on computing power, infrastructure construction and research and development in the second half of last year, which weighed on profitability, the source said.

Although ByteDance has not officially confirmed the figures, based on the USD33 billion in net profit that it reported in 2024, the Beijing-based firm is likely to have made about USD9 billion in net profit last year.

Fluctuating financial data among leading technology companies demonstrate that investments in computing power are an important ticket to the endgame of AI, the source said. The sharp drop in ByteDance's net profit last year signals the company’s need to offset the ‘computing power deficit’ arising from talent retention pressures, increased computing expenses and challenges associated with evolving technological pathways with new high-growth initiatives.

AI represents a larger core opportunity for the industry than that of personal computers and the worldwide web, Chief Executive Officer Liang Rubo said at a company-wide meeting at the beginning of the year. In the near term, ByteDance aims to advance its AI chatbot Doubao and its international version Dola, as well as its enterprise-focused Model-as-a-Service offerings.

ByteDance’s China revenue soared 20 percent in 2025 from the year before, while overseas revenue surged almost 50 percent, the source said. Revenue from its international segment now accounts for a record 30 percent of total revenue, up from 25 percent in 2024, with TikTok’s e-commerce business serving as a key driver.

ByteDance’s investments reportedly include the self-developed Dayu AI liquid-cooled server cabinets that can each support between 64 and 128 graphic processing units, a MegaScale training cluster capable of supporting tens of thousands of GPU units and designed for stable training of trillion-parameter Mixture-of-Experts models, diversified procurement of computing resources and self-developed initiatives in optical chip technology.

Editor: Kim Taylor