admin The agreement covers application-specific integrated circuits, or ASICs, designed to support AI workloads in ByteDance’s expanding data-centre operations. The chips are expected to be used for AI agent software, a fast-growing area in which companies are racing to lower inference costs while serving millions of users at scale. ByteDance, the owner of TikTok and one of China’s most aggressive AI developers, is set to buy millions of the chips, making it one of the first large customers for Qualcomm’s push into custom AI silicon.
Shares of Qualcomm rose sharply after details of the agreement emerged, reflecting investor optimism that the company can turn its mobile-chip expertise into a broader data-centre business. The move also places Qualcomm in closer competition with suppliers such as Broadcom and Marvell, which have benefited from demand for custom chips designed for large cloud and internet companies.
The reported deal comes as global technology groups seek alternatives to general-purpose graphics processors, particularly for AI inference, where trained models are deployed to answer queries, generate text, process images and run autonomous agent workflows. Nvidia remains the dominant player in AI accelerators, but the enormous cost of deploying models has encouraged internet platforms and cloud providers to use custom silicon tuned for their own software stacks.
Qualcomm has been preparing for this shift. Its AI200 and AI250 accelerator products, unveiled as part of its data-centre strategy, are scheduled for commercial availability in 2026 and 2027. The company has promoted the chips as rack-scale inference systems built around high memory capacity, energy efficiency and lower total cost of ownership. The architecture draws on Qualcomm’s Hexagon neural processing technology, long used in mobile devices, but scaled for data-centre deployment.
For ByteDance, the deal supports a wider drive to secure computing capacity despite tight global supply conditions and export controls on advanced AI chips. The company has invested heavily in AI models and applications, including Doubao, its flagship AI assistant in China. Doubao passed 100 million daily active users during the Lunar New Year holiday in February, helped by its integration with a national television gala and a surge in consumer interest in AI tools. That scale places heavy pressure on inference infrastructure, especially as users move from simple chatbot queries to multi-step agent tasks.
ByteDance has also explored in-house chip development, including an inference chip project known as SeedChip and manufacturing discussions with Samsung Electronics. Plans under discussion earlier this year involved prototype chips and production targets running into hundreds of thousands of units. A partnership with Qualcomm could give ByteDance access to a more production-ready path while allowing it to tailor hardware for its own AI workloads.
The arrangement will still have to operate within Washington’s restrictions on advanced chip exports to China. Qualcomm, like other US semiconductor companies, must ensure that products supplied to Chinese customers comply with export-control rules. Those restrictions have reshaped the AI hardware market by limiting access to the most advanced processors and encouraging Chinese technology groups to diversify suppliers, develop domestic alternatives and optimise software for less powerful hardware.
Qualcomm’s motivation is clear. The company reported $44.3 billion in revenue for fiscal 2025, up 14 per cent from the previous year, with its chip division benefiting from handset, automotive and internet-of-things demand. Yet smartphones remain a mature market, and the company has been looking for growth in automotive systems, PCs, connected devices, robotics and data-centre AI. A major ByteDance order would strengthen the case that Qualcomm can compete for large-scale infrastructure contracts rather than remain primarily associated with mobile devices.
Cristiano Amon, Qualcomm’s chief executive, has framed the company’s strategy around “intelligent computing everywhere”, with AI running across phones, PCs, vehicles, edge devices and data centres. That strategy depends on convincing customers that Qualcomm can deliver performance and efficiency at scale, not just low-power processing for consumer hardware.
The ByteDance deal also underlines a broader industry trend: AI companies are moving from model-building experimentation to industrial deployment. As usage grows, the cost of serving each query becomes a strategic issue. Custom ASICs can reduce power consumption and improve efficiency when matched closely to a company’s software and workloads, but they require large volumes, long-term planning and deep integration between chip designers and platform operators.
Competition is likely to intensify. Nvidia is defending its lead with faster chips, networking systems and software tools. AMD is expanding its accelerator line-up. Broadcom and Marvell are targeting custom silicon for hyperscale customers. Cloud providers and internet companies are designing their own processors to gain more control over cost and supply. Qualcomm’s advantage lies in energy-efficient AI processing and decades of experience integrating compute, memory, connectivity and software into compact systems.
