AI in practice

ByteDance and Broadcom partner to develop AI chips

Matthias Bastian
hand-drawn illustration of a semiconductor chip. It should convey the essence of a microprocessor or integrated circuit using bold, simplified geometric shapes and lines. The colors should be more abstract and artistic, using broad strokes of purples, oranges, and greens to suggest a reflective surface without the intricate details of the actual circuitry. The overall design should be modern and minimalistic, focusing on the square shape of the chip with suggestive outlines of internal components

Midjourney prompted by THE DECODER

TikTok parent company ByteDance is collaborating with US chipmaker Broadcom to create an AI processor, according to Reuters sources.

The custom 5-nanometer chip aims to comply with US export regulations and will be manufactured by TSMC in Taiwan. The project seeks to secure a supply of high-performance AI chips while reducing costs. Production is slated to begin next year. The two companies already collaborate in other areas.

ByteDance is researching generative AI and implementing the technology in TikTok, most recently with digital avatars. However, political restrictions limit the Chinese company's access to AI chips compared to its international rivals.

Reuters reports that ByteDance spent $2 billion on Nvidia chips last year. This is a small fraction of the investment made by social media competitor Meta, one of Nvidia's biggest known customers alongside Microsoft.

Mark Zuckerberg recently announced that Meta plans to deploy 340,000 Nvidia H100 GPUs and around 600,000 total GPUs by year-end. If generative AI becomes critical to social media services, ByteDance risks falling behind if it doesn't have enough compute.

Moreover, ByteDance's competitors not only have better access to Nvidia cards but are also developing their own AI chips. Meta's "Artemis" chip is set to enter production this year and will be used in the company's data centers for AI model inference.

Microsoft has also developed its own AI chips, Azure Maia, and an Arm-based CPU, Azure Cobalt, for its cloud infrastructure.

OpenAI is considering developing its own AI chips and is working with RainAI, a company creating a neuromorphic processing unit (NPU). OpenAI CEO Sam Altman is reportedly planning to raise significant funding to build a global chip infrastructure as an alternative to Nvidia.

Sources: