ByteDance Designs Custom AI Chips to Reduce US Dependence
ByteDance is developing its own AI processors designed for inference tasks, according to Reuters. The move reflects the company's effort to reduce reliance on US chipmakers amid tightening export controls and geopolitical pressure.
The chips take inspiration from Groq's language processing units, which optimize performance for running AI models rather than training them. ByteDance is evaluating both Arm and RISC-V architectures for the design, which remains in the concept and planning phase.
Memory and Manufacturing Strategy
ByteDance has partnered with Chinese startup InnoStar Semiconductor for memory technology, potentially avoiding the need to purchase expensive HBM chips from Samsung and other US suppliers. Both ByteDance and Alibaba have invested in InnoStar.
The company lacks internal chip design teams and will rely on external partners for both design and manufacturing. This mirrors the approach taken with its SeedChip accelerator, developed with TSMC starting in 2024 and expected to reach mass production this year.
Context: Export Controls and Market Dynamics
China's government banned purchases of Nvidia's H200 Blackwell chips after the Trump administration tightened export restrictions. ByteDance will likely continue using hybrid server architectures with Nvidia processors in the near term, but custom silicon offers a longer-term alternative.
Nvidia holds substantial pricing power in the market. Intel and AMD increase prices quarterly, while Nvidia's newer Vera chips remain US products. Custom chips reduce exposure to these supply and pricing constraints.
ByteDance's Broader AI Infrastructure
Beyond TikTok, ByteDance operates Doubao, a Chinese AI chatbot app, and maintains multiple AI models in production. The company's chip development reflects its shift toward controlling its own infrastructure stack.
For development teams working on AI for IT & Development, this trend signals broader changes in how companies approach generative AI and LLM deployment. Custom hardware optimized for inference workloads may become standard as inference-heavy agentic AI expands.
Your membership also unlocks:
