ByteDance’s in-house self-developed CPU will be finalized no later than early 2027, with mass production slated for the second half of the year—pulling in Qualcomm and fighting to secure wafers.

According to three sources familiar with the matter cited by the South China Morning Post, ByteDance plans to complete the design finalization of its next-generation self-developed CPU by early 2027 at the latest, with the goal of starting mass production and large-scale deployment in data centers in the second half of 2027 to support the inference needs of its agentic AI services such as Coze.
(Previous context: Qualcomm is in talks with ByteDance about custom chip design! VPU chips to enter mass production by end of year, stock price boosted in after-hours trading)
(Background supplement: UBS and TD Cowen raised Arm's target price to $475 on the same day, citing future revenue from self-developed CPUs)

Key Takeaways

• Self-developed CPU design finalized by early 2027 at the latest; early version already running internally since late 2025
• In talks with Qualcomm to collaborate on securing TSMC fab capacity; custom VPU to begin mass production as early as end of 2026
• Evaluating Arm and RISC-V architectures, aiming to bypass Intel and AMD and fill the supply gap caused by Nvidia export controls

ByteDance is building its own CPU, aiming for mass production in the second half of 2027 and large-scale deployment in its data centers. The South China Morning Post, citing three sources with knowledge of the matter, reported that the next-generation self-developed CPU design will be finalized by early 2027 at the latest, and an early version has been running internally since late 2025. To the surprise of outsiders, the assisting party is U.S. chipmaker Qualcomm.

AI enters the inference era: CPU is the next main character

The AI industry is shifting gears. It is rapidly moving from the early "training phase" of burning GPUs to train large models to "inference" (the process where models actually perform tasks externally and generate answers). The next core application is agentic AI, i.e., systems that can autonomously break down goals, plan steps, and execute repeatedly until complex tasks are completed. ByteDance's Coze AI development platform is actively promoting such applications. This type of workload demands far more CPU resources than GPU-intensive training. By developing its own CPU, ByteDance is essentially betting directly on the next major AI battlefield, bypassing the queue behind Nvidia's GPU supply chain.

In terms of architecture choice, ByteDance is evaluating two paths: the instruction set architecture of SoftBank-owned Arm and the fully open-source RISC-V, with the goal of removing Intel and AMD from the core of its supply chain.

Because Nvidia's most advanced AI accelerators have long been subject to U.S. export controls, with most models no longer able to enter the Chinese market, relying on Nvidia allocation quotas is no longer a viable option for ByteDance.

The source added that due to urgent expansion needs, the tape-out (the critical step where chip design is formally sent to the fab for trial production) of the next-generation CPU may start earlier than originally planned. ByteDance's actual expansion pace for AI infrastructure has already outpaced the originally scheduled rhythm.

Partnering with Qualcomm to gain time

ByteDance's rationale for bringing in Qualcomm lies in the supply chain, not just technology. Qualcomm operates a fabless model, building no factories of its own and long producing chips through foundries like TSMC. In a time of extreme tightness in global advanced process capacity, Qualcomm's order quota at TSMC is itself a scarce resource. Cooperating with Qualcomm is equivalent to leveraging its manufacturing relationships to buy time.

Qualcomm itself is also happy to see such cooperation, as it is actively diversifying away from the smartphone market and, through the acquisition of AlphaWave Semi completed in December 2025, has entered the high-speed connectivity and custom chip business.

Sources said the scope of cooperation under discussion includes not only assisting CPU development but also building a VPU (video processing unit) for ByteDance, with mass production as early as the end of 2026, earlier than the timeline for the self-developed CPU. These are two related but independent product lines; do not mistake the custom chip assisted by Qualcomm as ByteDance's own self-developed result.

However, the two sides have not yet signed any formal agreement, and ByteDance may also turn to other suppliers. Meanwhile, ByteDance is reportedly seeking about $20 billion in loans to support the overall expansion of its AI infrastructure. Google, Amazon, and Microsoft have been on this path for years. The wave of self-built CPUs now has one more major player confirmed.

Frequently Asked Questions

Why does ByteDance need to develop its own CPU?

Nvidia's most advanced AI accelerators are subject to U.S. export controls, with most models unable to enter China, forcing ByteDance to build its own chip supply chain. At the same time, AI is shifting from training to inference, and agentic workloads demand far more CPU than GPU. Developing its own CPU allows direct optimization for the next wave of AI scenarios, without relying on Nvidia quotas.

What role does Qualcomm play in ByteDance's CPU plan?

Qualcomm operates a fabless model, producing chips through foundries like TSMC, and its manufacturing quota is a scarce resource. ByteDance leverages cooperation with Qualcomm to secure advanced fab capacity, and the two sides are also negotiating a custom VPU (video processing unit). However, no agreement has been signed yet, and the outcome remains uncertain.