Frustrated by long lead times for off-the-shelf processors, ByteDance is now designing its own silicon — with two very different architectures in play.
ByteDance has had enough of waiting months for processors, so it's going to make them itself. The company behind TikTok and video editing app CapCut is developing its own CPUs to better support its AI infrastructure. Given the AI industry's hunger for silicon, it's hardly surprising another massive tech company is taking hardware into its own hands. The plan is in the early stages and not yet public, but Reuters spoke with a number of anonymous sources familiar with the matter. Several external partners have already been approached with regard to design work for ByteDance's chip, and securing capacity at manufacturing foundries was apparently also discussed.
Currently, ByteDance is exploring two different chip architectures: Arm and RISC-V. Both architectures are used for chips in data centres throughout the wider industry, though Arm presents a proprietary ISA with a fixed feature set, whereas RISC-V is a modular, royalty-free, open-source architecture. The planned chip is intended to be deployed throughout ByteDance's AI data centres as the AI industry reaches towards CPU-intensive inference.
Arm vs. RISC-V: Two Paths Forward
Arm is a proprietary instruction set architecture widely used in data centers. ByteDance would pay licensing fees to Arm Ltd. and work within a fixed feature set. The ecosystem is strong, with mature software, tools, and operating support well established across the industry.
RISC-V is open-source, royalty-free, and modular. Not having to pay Arm's licensing fees may be especially appealing to a major player looking to make its own chips. However, the trade-off is a thinner software ecosystem, meaning ByteDance would need to handle more validation and tooling work internally.
Sources indicate ByteDance is actively evaluating both architectures. The choice will shape every downstream design decision — from cache hierarchy and memory controllers to the software stack. What is clear is that ByteDance is not waiting for a single vendor to deliver the ideal chip.
Why AI Inference Drives the Decision
ByteDance's planned chip is intended specifically for inference — the CPU-intensive phase where a trained model processes real-time requests. As the AI industry reaches toward heavier inference workloads, data-center operators need chips that handle high throughput, low latency, and energy efficiency within a data-center power envelope.
Off-the-shelf chips from major vendors often overprovision general-purpose features that may not align with a specific company's workloads. A custom chip would allow ByteDance to tailor its silicon to the demands of serving AI models at scale, potentially reducing power consumption and improving throughput per dollar.
What This Means for the Chip Industry
ByteDance is not alone in this shift. Other major tech companies have also taken hardware into their own hands to support their AI infrastructure. The trend signals that companies controlling both the algorithms and the scale are increasingly valuing vertical integration over relying solely on vendor relationships.
For the supply chain, foundries stand to benefit — whoever manufactures ByteDance's chips will gain additional capacity orders. Toolchain and design vendors would also collect fees depending on which architecture is chosen. The open-source RISC-V angle is particularly notable: if ByteDance goes that route, the RISC-V movement gains a powerful backer, which could accelerate the ecosystem for everyone.
FAQ: Quick Answers to Common Questions
Will ByteDance sell these chips to other companies?
No evidence suggests that. The plan as reported is for internal AI infrastructure only.
How long until the chips are ready?
The plan is described as being in the early stages. Custom CPU development typically takes several years from architecture definition to production deployment.
Does this affect TikTok's recommendation algorithm?
Only indirectly. Faster inference chips could reduce latency and cost of running AI models, but the algorithm itself is software, not hardware.
