ByteDance develops custom CPUs to power AI infrastructure buildout

ByteDance is building its own central processing units to feed a fast-growing artificial intelligence infrastructure push, a move that shows how China’s biggest tech companies are trying to regain control over cost, supply and performance as the hardware market tightens.

The custom chip project is aimed at ByteDance’s own servers and data centers, where it would support internal operations and a broader rollout of agent-based products, including the company’s Coze platform. That matters because AI spending is moving beyond model training and into inference and agentic workloads, where systems execute tasks, coordinate tools and keep services running at scale. In that environment, CPUs matter more than they once did, especially for software stacks that need predictable performance and tighter integration across infrastructure.

The effort is still early, but its logic is clear. ByteDance has been pushed by surging chip prices and prolonged shortages that make it harder to expand AI systems at the pace it wants. A proprietary CPU could give the company more control over long-term deployment costs, improve tuning for its own software, and reduce exposure to market bottlenecks that can delay new capacity.

The timing also lines up with ByteDance’s spending plans. Reports in May said the company had raised its 2026 AI infrastructure capital expenditure plan to more than 200 billion yuan, or about $30 billion, up from a preliminary 160 billion yuan plan discussed late last year. That scale of investment points to a company racing to secure enough compute to keep pace with the AI arms race, while also trying to avoid being boxed in by outside suppliers.

Supply risk has become a central part of that calculation. In February, Intel and AMD warned Chinese customers of long delivery delays for server CPUs, with Intel lead times reaching up to six months. For a company like ByteDance, which runs huge data-center fleets and is expanding AI services quickly, that kind of lag can translate directly into slower product launches and higher costs.

ByteDance is not alone in taking this path. Google has its Axion Arm-based processor for general-purpose workloads, including CPU-based AI. Amazon Web Services says its Graviton chips are designed to improve price-performance in cloud workloads. Microsoft announced Cobalt 200 in November 2025 as its next-generation Arm-based CPU for cloud-native workloads. The pattern shows a broader industry shift toward in-house silicon, with companies seeking insulation from both pricing pressure and supplier dependence.

For ByteDance, the move also carries a geopolitical edge. A custom CPU strategy offers a measure of self-sufficiency in a market shaped by export controls, supply-chain strain and escalating competition over AI infrastructure. It signals that China’s leading internet firms are no longer just adapting their software for the AI era. They are rebuilding the hardware stack beneath it.