ByteDance is developing custom AI processors designed specifically for running AI models rather than training them, marking a significant shift in how the Chinese tech giant approaches its infrastructure independence. The company is evaluating both Arm and RISC-V architectures for these chips, which remain in the early concept and planning phase.

Why Is ByteDance Building Its Own AI Chips?

The move reflects mounting pressure from US export controls and geopolitical tensions. China's government banned purchases of Nvidia's H200 Blackwell chips after the Trump administration tightened export restrictions, forcing Chinese companies to seek alternatives. ByteDance, best known globally for TikTok but also operating Doubao, a Chinese AI chatbot app, faces the same constraints as other domestic tech firms.

Beyond regulatory pressure, custom chips offer ByteDance protection from pricing volatility. Nvidia holds substantial pricing power in the market, while Intel and AMD increase prices quarterly. By developing its own silicon, ByteDance can reduce exposure to these supply and pricing constraints over the long term.

What Architecture Is ByteDance Using for These Chips?

ByteDance's custom processors draw inspiration from Groq's "language processing units," which are chips optimized specifically for inference tasks, meaning running already-trained AI models rather than training new ones from scratch. This focus on inference aligns with industry trends, as inference-heavy agentic AI, where AI systems take actions autonomously, is becoming increasingly common.

The company is partnering with Chinese startup InnoStar Semiconductor for memory technology, potentially avoiding the need to purchase expensive HBM (high-bandwidth memory) chips from Samsung and other US suppliers. Both ByteDance and Alibaba, China's cloud and e-commerce giant, have invested in InnoStar.

How Is ByteDance Approaching Chip Development and Manufacturing?

ByteDance lacks internal chip design teams and will rely on external partners for both design and manufacturing. This approach mirrors the company's strategy with its SeedChip AI accelerator, which ByteDance developed with TSMC starting in 2024 and expects to reach mass production this year.

• Design Phase: The custom CPU project is currently in the concept and planning stage, with ByteDance evaluating multiple architectural approaches before committing to a final design.
• External Partnerships: ByteDance will outsource both the chip design work and the actual silicon manufacturing to specialized partners rather than building these capabilities in-house.
• Memory Technology: The company is partnering with InnoStar Semiconductor to develop memory solutions, reducing dependence on US-based suppliers like Samsung for expensive HBM chips.
• Hybrid Infrastructure: For the near term, ByteDance will likely continue using hybrid server architectures that combine Nvidia processors with custom silicon as the new chips mature.

What Does This Mean for ByteDance's AI Infrastructure Strategy?

ByteDance's chip development reflects a broader shift toward controlling its own infrastructure stack. The company operates multiple AI models in production beyond Doubao, and custom hardware optimized for inference workloads may become standard as inference-heavy agentic AI expands.

For now, ByteDance will likely continue relying on Nvidia processors in hybrid server setups, as complete independence from US suppliers remains impractical in the near term. However, custom silicon offers a longer-term alternative that reduces both supply chain vulnerability and exposure to pricing pressures from dominant US chipmakers.

This development signals broader changes in how companies approach generative AI and large language model (LLM) deployment. As inference becomes the dominant workload for deployed AI systems, custom hardware optimized for these specific tasks may become increasingly common across the industry.