Qualcomm is reportedly in talks with ByteDance to design customized chips for the TikTok parent company. The negotiations show yet another front in the race by tech companies to design AI chips outside the dominance of Nvidia.

Success in these negotiations will add to the success of the relationship that has yielded results before. In May, Qualcomm had signed a deal to provide ByteDance with millions of application-specific integrated circuits (ASICs) for its AI data center applications, as revealed by Reuters. These latest negotiations seem to extend to the chip design by Qualcomm, using the technology of AlphaWave Semi, a high-speed connectivity firm bought by Qualcomm last year.

Qualcomm-ByteDance deal?

According to reports, the new negotiations are apparently focused on creating video processing chips that fit ByteDance’s architecture, with an initial target of mass production by late 2026. The chips will be using Qualcomm’s technology, specifically, IP cores from AlphaWave Semi, a company Qualcomm purchased in 2023 for $2.4 billion, for ultra-fast connectivity solutions.

Qualcomm CEO Cristiano Amon said in May that the company is developing three types of data center chips: CPUs, inference accelerators, and custom ASICs. That puts Qualcomm in direct competition with Broadcom and Marvell, each company earning billions of dollars per year for custom chip manufacturing for hyperscale providers.

Neither Qualcomm nor ByteDance provided any comments to the reports. The sources remained anonymous, citing the confidential nature of the negotiations and cautioning that the outcome is not guaranteed.

Why ByteDance need customized chips?

ByteDance’s internal chip development has been on the rise steadily alongside its growing efforts to source chips externally. According to Reuters, on May 28, ByteDance began making its CPU-based chips using Arm and RISC-V architectural designs as a result of bottlenecks in the supply chain, as well as rising component costs.

The overall market for server CPUs has been experiencing longer-than-average lead times. Intel has been reporting to its Chinese customers that it could take as long as six months to have their orders filled, while AMD CEO Lisa Su has said, according to Reuters, that CPU availability in the global market is very “tight.”

Cost pressures are also a significant factor that ByteDance is facing. According to two sources to Reuters, the cost of CPU sourcing for ByteDance has increased by approximately 10% to 35% compared to the previous quarter.

It appears that ByteDance has been required to obtain financing to support its continued purchases of hardware. Reuters also reported that ByteDance is currently in negotiations with banks for a $20 billion 3-year term loan, which may be extended for 2 additional years for offshore financing. If this loan is approved, it will represent the largest financing package ever for ByteDance, indicating continued infrastructure growth.

AI chip race is on

The global semiconductor supply chains have been transformed by the AI chip race.

As Qualcomm and ByteDance continue negotiations, they are now a reflection of the larger drive toward proprietary silicon stacks in the industry; Google has its TPUs for its internal AI workloads, Amazon has created its own Trainium and Inferentia chips for training and inference tasks; Meta has developed its MTIA family of accelerators; and Microsoft is working to introduce Maia AI accelerators. Ultimately, these systems will help to reduce the cost of inference on a per-token basis and also reduce the reliance upon limited supply chains for GPUs, particularly on Nvidia’s much-demanded hardware.

Computing the density threshold has also become important in the context of export controls. As of the 2023 rules of the U.S. Commerce Department, restrictions on advanced AI chips depend on some metrics, such as 4,800 tera-operations per second (TOPS) of integrated circuit compute density limits in particular categories of controlled hardware.

Qualcomm is simultaneously navigating a softer smartphone cycle. Global handset shipments are expected by industry analysts to post one of their steepest annual declines on record, according to Reuters reporting, while memory chip inflation has added cost pressure across the supply chain.

Export controls loom

China remains subject to U.S. export restriction laws for any AI hardware with customers located in China, whether it is the compute performance of the hardware itself or the amount of data that can cross from one customer’s machine to another using the device interconnects.

The earlier ASIC agreement between Qualcomm and ByteDance was structured for the chips to remain below the U.S. export rules for allowed compute limits. According to Reuters, the VPU types discussed last month will be treated similarly (it is unclear if the VPU types will be subject to the same restrictions).

A critical variable regarding deployment to various regions is whether the installation will be reviewed by U.S. authorities based on location or type of placement. For example, if the data center is installed in mainland China, it will likely be subject to significantly greater regulatory scrutiny than if it were installed outside of China in a country with less restrictive regulations on U.S. exports, such as Singapore, which is the most popular regional center for cloud computing infrastructure.

Washington continues to broaden the scope of regulatory oversight of high-performance AI systems, including the addition of other custom-designed accelerators and custom-designed network interface components to the existing list of regulatory restrictions that are applicable to GPUs.

What to expect?

Investors are interested in whether the conversations lead to a tape-out and production-ready designs, which would signal a shift from exploratory talks to committed production.

Qualcomm’s other M&A activity and overall positioning are part of the background here; they are reportedly close to a purchase of AI chip startup Alphawave-related assets (Modular) for $4 billion and are in separate negotiations to purchase Tenstorrent in a deal estimated between $8 billion and 10 billion, according to Bloomberg via Reuters. If those deals all converge within Qualcomm’s AI stack, it could rapidly shift Qualcomm from a smartphone-centric semiconductor supplier to a vertically integrated data centre AI player.

If you’re reading this, you’re already ahead. Stay there with our newsletter.