According to the latest report from The Information, ByteDance is working with TSMC to mass-produce two chips designed by it in 2026, thereby reducing its dependence on Nvidia.

So far this year, ByteDance has ordered more than 200,000 Nvidia H20 chips worth more than $2 billion, but is still waiting for all to be delivered.
The performance of H20 is 1/4 of that of H100. Due to US export restrictions, domestic manufacturers need to buy more H20 to achieve a single cluster with the same computing power, which undoubtedly increases costs.

According to the latest plan, ByteDance will order hundreds of thousands of training and inference chips designed by it from TSMC, hoping that its own chip-driven cluster can achieve four times the computing performance of a single H100 at the same cost, saving billions of dollars in expenses.

In June this year, Reuters reported that ByteDance was working with Broadcom to develop a 5 nm custom ASIC chip and outsourced it to TSMC, which was only one generation behind TSMC's most advanced 3 nm process and the same as the new generation Blackwell architecture chip announced by Nvidia in March.

Currently, ASIC chip development has these advantages:
Optimize performance: Customized specifically for set tasks to provide higher efficiency (i.e. LLM reasoning; news push); Reduce costs: Avoid using third-party hardware, thereby reducing operating costs; Energy efficiency: Customized ASIC consumes less power, which is very important for large data centers; Enhance security: Direct control of hardware can enhance data security and privacy;

Of course, ASIC chips also have many disadvantages, such as the need to invest a lot of money in design and manufacturing in the early stage, and it takes a long time from design to deployment; In addition, rapid technology iterations can easily cause ASIC chips to become obsolete.
As early as 2022, Broadcom became a business partner of ByteDance. It publicly stated that ByteDance purchased its 5nm high-performance switch chips and Bailly switches for AI computer clusters. Customized ASIC chips will help the platform better handle intensive video tasks and optimize user experience and operational efficiency.

Source: Youxin