Qualcomm has reportedly reached a deal to supply ByteDance with millions of custom application-specific integrated circuits (ASICs) for AI data centers, according to a Bloomberg report on May 26. The news sent Qualcomm shares up more than 6% to a record high near $247.91 — and signals that the TikTok parent is becoming one of the first major customers for Qualcomm's nascent AI-silicon business.
Custom ASICs, not off-the-shelf accelerators
The nuance matters for anyone tracking the inference-silicon market. Per Bloomberg's reporting, the deal will help ByteDance turn an already-completed in-house chip design into a production-ready semiconductor. In other words, Qualcomm is acting as a custom-silicon partner — closer to the model Broadcom and Marvell run for hyperscalers — rather than simply selling its branded AI200/AI250 data-center inference parts off the shelf.
ByteDance is set to procure the chips to support its AI agent software, including the workloads behind Doubao, which was China's most-downloaded AI chatbot for much of last year. For Qualcomm, landing a customer at this volume validates a custom-ASIC strategy that until now has been mostly forward-looking commentary.
Qualcomm's pivot beyond the smartphone
The deal is a concrete step in Qualcomm's effort to expand past mobile processors into AI infrastructure. On the company's late-April earnings call, CEO Cristiano Amon flagged a "leading hyperscaler custom silicon engagement" and described discussions on track for initial shipments later this year, framing it as a multi-generation partnership. Reporting now points to ByteDance as the anchor customer behind that language, though Qualcomm has not publicly named it.
That positions Qualcomm as a credible third option for custom inference silicon alongside Broadcom and Marvell — a market that has so far been dominated by US hyperscalers designing their own accelerators (Google's TPU, Amazon's Trainium, Microsoft's Maia).
ByteDance's compute diversification
For ByteDance, the move fits a broader push to secure compute outside constrained Nvidia GPU supply. The company has raised its 2026 AI infrastructure budget by roughly 25% to about 200 billion yuan (~$29.4 billion) and has pursued domestic and custom silicon to reduce dependence on export-controlled hardware. Owning the chip design while outsourcing productization lets ByteDance tune silicon to its own agent and recommendation workloads without building a full chip-manufacturing apparatus.
What changes for builders and buyers
The signal here is that custom inference ASICs are broadening beyond the handful of US cloud giants. If Qualcomm can convert a single "leading hyperscaler" engagement into a repeatable custom-silicon line, enterprise buyers and infra teams gain another vendor for high-volume inference — and the economics of agentic workloads at scale get a new reference point. Watch for shipment timing later this year and any follow-on customers as proof the business is more than a one-deal headline.
