Qualcomm has landed ByteDance as a customer for custom AI chips. The deal covers millions of application-specific integrated circuits for inference workloads, specifically the Doubao chatbot, China’s most-downloaded AI assistant, and the broader agent infrastructure around it. Qualcomm is also helping ByteDance bring ByteDance’s own proprietary chip design from blueprint to volume production, a foundry-plus-supply arrangement that runs through TSMC, Qualcomm’s primary foundry partner.

The chips are designed to comply with U.S. Commerce Department export control thresholds. Both the Qualcomm ASICs and ByteDance’s proprietary designs are engineered to remain below the computing performance limits the current framework sets for chip exports to China. ByteDance, for its part, raised its 2026 AI infrastructure budget by 25 percent to 200 billion yuan, roughly $29.4 billion, and a material share of that is going toward domestic chips. Export controls created the demand. A U.S. chipmaker found a way to fill it legally. Analysts are calling the arrangement a geopolitical tightrope, which is accurate, and also a proof-of-concept, which is the more interesting part: if the thresholds move, the deal breaks.

On the same day, Micron crossed $1 trillion in market cap after its stock surged roughly 19 percent in a single session. The catalyst was a UBS note from analyst Timothy Arcuri, who raised his price target from $535 to $1,625 and argued that AI has permanently changed the memory business from a cyclical commodity trade to something closer to a utility. UBS projects Micron will generate over $400 billion in cumulative free cash flow between 2027 and 2029. Micron CEO Sanjay Mehrotra said the company is currently fulfilling only 50 to 65 percent of key customers’ medium-term high-bandwidth memory demand, and that entire HBM production capacity for 2026 is already sold out.

SK Hynix joined the trillion-dollar club the following day, its stock rising roughly 10 percent in Seoul. SK Hynix holds roughly 70 percent of NVIDIA’s HBM4 orders for the Vera Rubin platform and posted a 72 percent operating margin in its most recent quarter. Samsung, the third major memory producer, crossed $1 trillion earlier in May. The three companies are collectively worth around $3 trillion, which is a lot of money for an industry that lost $90 billion of it in two days in March when a Google research paper suggested AI might need less memory. It did not, or at least the market has concluded it did not, and concluded this emphatically.

The connection between a Qualcomm ASIC deal and a memory stock repricing is the supply chain underneath both of them. Custom inference silicon needs high-bandwidth memory. Inference at ByteDance’s scale, tens of millions of Doubao users generating agent workloads, is exactly the kind of deployment that fills HBM order books. The same TSMC advanced nodes that fabricate Qualcomm’s ASICs also produce the chips that SK Hynix and Micron sell to NVIDIA. Demand is not competing for the same chip. It is competing for the same fab time, the same packaging capacity, and the same memory.

ByteDance is spending $29.4 billion on AI infrastructure. Huawei is proposing new laws of physics to explain why being three years behind TSMC is progress. Qualcomm is designing ASICs that fit precisely under the line the Commerce Department drew. And the memory companies that supply all of them are worth a trillion dollars each, up from considerably less than that two months ago. The AI chip market used to be a story about who could buy the most NVIDIA GPUs. It is now a story about everything else, too.

–
By the Control Plane Editorial Team