Qualcomm Stock Hits Record on ByteDance AI Chip Order Report

Qualcomm closed at a record $248.82 on Tuesday after Bloomberg reported the chipmaker had landed an agreement to ship millions of custom inference parts to ByteDance, the Beijing-based parent of TikTok. Shares touched an intraday all-time high of $258 and finished up about 4%, capping a roughly 24% run over the past month.

The reported deal puts a Chinese internet major on the customer list of a company that, twelve months ago, sold almost exclusively to phone makers and automotive customers. It also drops one more weight onto the side of the scale tilting against Nvidia’s near-monopoly in AI accelerators.

The Tuesday Move and the Shape of the Order

The Bloomberg sourcing, first published on May 26, describes a multi-million-unit order for Qualcomm application-specific integrated circuits (ASICs, chips fabricated for one workload rather than general computing) to power ByteDance’s AI agent software. Volume, dollar value, and delivery schedule were not disclosed. Neither Qualcomm nor the buyer has confirmed the agreement publicly, and Reuters said it could not independently verify the report.

What is on the record is the trading reaction. Qualcomm opened higher on Tuesday and ran past its prior 52-week ceiling, with the intraday peak at $258 marking the strongest single-session breakout since the company’s late-October AI accelerator unveil. The stock has now closed above $200 for nine of the last ten sessions, a level it had not touched at any point in 2024.

Two pieces of context shape how the order is being read. Qualcomm’s chief executive Cristiano Amon told analysts on the April 29 fiscal second-quarter earnings presentation that initial shipments of its custom data center chips would begin in December, to “a large hyperscaler” he declined to name. And custom ASICs already cost as much as 65% less than merchant GPUs for inference workloads at scale, per TrendForce, which is why every major AI buyer now keeps a second sourcing track open.

Amon’s Three-Chip Roadmap, Built Through 2025

Qualcomm’s data center push spans three product lines, layered into a roadmap Amon has been quietly assembling since the back half of 2025.

The first is the AI200, a rack-scale inference accelerator unveiled on October 27, 2025 and slated for commercial availability this year. Built around Qualcomm’s Hexagon neural processing units (NPUs, the AI cores it scaled up from smartphone silicon), the AI200 supports up to 768 gigabytes of LPDDR memory per card, runs on direct liquid cooling, and tops out at 160 kilowatts of rack power. Saudi Arabia’s Humain joint partnership for 200 megawatts of AI capacity was the first announced customer.

The second is the AI250, due in 2027, which Qualcomm says will deliver more than 10x the effective memory bandwidth of the AI200 through a near-memory compute architecture.

The third, and the one that matters for the ByteDance arrangement, is the custom ASIC track. Here Qualcomm is positioning itself less as a chip vendor and more as a design partner, willing to take a customer’s in-house silicon blueprint and shepherd it through to volume manufacturing. Amon hinted at the breadth of inbound interest during a May appearance with CNBC.

You should expect that we’re working not only with them, but most of the AI companies today. So the design engagement is very robust.

That engagement spans multiple unnamed clients beyond the lone hyperscaler Qualcomm has publicly acknowledged. The TikTok parent, if Bloomberg’s sourcing holds, would be the first such partner to come into public focus.

The 44.6% Growth Gap Reshaping AI Silicon

Beyond the headline order, an industry projection from TrendForce published earlier in May captures the deeper structural pull on Qualcomm’s stock.

• 44.6% projected growth in custom ASIC shipments for 2026, per TrendForce.
• 16.1% projected growth in merchant GPU (graphics processing unit) shipments for the same year.
• 27.8% ASIC share of total AI server shipments expected in 2026, up from a single-digit base three years earlier.
• ~80% Nvidia’s current share of AI accelerator revenue, per Silicon Analysts.

ASIC Shipments Pull Ahead for the First Time

The 2026 projections mark the first year of the modern AI build-out in which custom silicon shipment growth outpaces general-purpose GPU growth, and not by a slim margin. A near three-to-one ratio does not flip the installed base, where Nvidia still ships the dominant share of training systems. But it does mark the moment the inflection becomes measurable rather than theoretical.

Inference Is the Pivot Point

The split between training and inference is where the case for ASICs lives. Training a frontier model needs general-purpose flexibility and tight networking, which is what Nvidia’s Blackwell-class systems deliver. Inference, the work of running a trained model at production scale, runs on more predictable matrix math and benefits from chips fabricated for that exact workload. Inference now accounts for roughly two-thirds of all AI compute, according to analysts tracking hyperscaler capital spending.

Where the Mix Is Heading

The longer-dated forecasts get bolder. Silicon Analysts and Introl have both modeled Nvidia’s inference market share falling from above 90% today to somewhere between 20% and 30% by 2028 if the ASIC ramp holds. That is the assumption priced into a 4% Tuesday in Qualcomm.

ByteDance’s Three-Track Compute Plan

The Bloomberg report is best read against the TikTok parent’s already-public sourcing strategy, which has split into three parallel tracks over the past 18 months.

• Track one: Nvidia, routed offshore. The Chinese internet major is on track to spend roughly $14 billion on Nvidia AI GPUs in 2026, with delivery routed through compliant cloud operators in Malaysia. A cluster of 500 NVL72 GB200 rack systems, holding roughly 36,000 Blackwell chips, will be operated by Aolani Cloud in Malaysia rather than landed inside China.
• Track two: domestic silicon. The company has reportedly engaged Huawei, Cambricon, and an internal design team for chips that can sit inside Chinese data centers without triggering US export controls.
• Track three: merchant ASICs from US suppliers. The reported Qualcomm engagement is the public face of this leg, with Qualcomm both selling its own AI200-class parts and helping push an in-house ByteDance design toward volume production.

Sourcing resilience is the point of the three tracks. No single vendor, no single jurisdiction, and no single architecture carries the full AI buildout. That posture is now standard for any AI infrastructure buyer touching China, and it is the structural reason Qualcomm has a seat at the table at all.

Broadcom and Marvell Already Own the Tier Qualcomm Just Entered

The merchant ASIC business Qualcomm is pushing into already has incumbents. Broadcom and Marvell have spent the better part of a decade building it.

The Merchant ASIC Field

Broadcom is the design partner behind Google’s TPU (tensor processing unit) v7 Ironwood and Meta’s MTIA (Meta Training and Inference Accelerator) program. Marvell holds the Amazon Trainium 3 design win and a slice of Microsoft’s Maia 200 line. Together those two companies represent the bulk of non-Nvidia AI silicon shipping inside US hyperscalers today.

Where Qualcomm Plugs In

What Qualcomm carries that the incumbents do not is a credible inference-first NPU lineage. Hexagon was built to run trained models on phones efficiently, which is the same fundamental problem hyperscalers now face at rack scale. The pitch to the TikTok parent, on the public reporting, was both AI200 parts and engineering support to take an internal design through tape-out and production. That dual-track offer is rare in the merchant ASIC tier and is why Bloomberg’s sourcing pegged the win as a category breakthrough rather than a one-off purchase order.

For context on how aggressively AI infrastructure spending is rotating across the broader supply chain, optical transceiver volumes have caught the same wave; see our prior reporting on AI optical spending and the AAOI revenue surge.

What Has to Hold Before the June Investor Day

Tuesday’s move has priced a deal that neither party has confirmed. Three things have to hold for the trade to keep paying.

The order has to formalize without running into the US Bureau of Industry and Security (BIS, the agency that administers chip export controls) compute thresholds that limit advanced silicon sales to Chinese end users. Industry coverage suggests Qualcomm has been designing the relevant ASIC SKUs at or just below those caps, but the BIS has revised the thresholds three times since 2022.

December shipments to the unnamed “large hyperscaler” Amon referenced must also arrive on schedule. Qualcomm’s June 19 investor day will set both the revenue framing and the customer-name disclosures that translate design wins into modeling lines. A slip there reopens the gap between narrative and recognized revenue.

The inference-mix thesis itself has to hold through Nvidia’s response. Blackwell Ultra and the Rubin generation arriving in 2027 are explicitly aimed at clawing back inference share with denser scale-up networking and software lock-in. If Nvidia’s CUDA moat absorbs the inference workload that custom silicon is supposed to take, the 44.6% growth gap compresses fast.

If the deal hardens and Humain plus the Chinese order plus the unnamed hyperscaler all hit shipment by the first half of 2027, Qualcomm finishes next year as the only smartphone-pedigree silicon vendor with a multi-customer AI infrastructure book. If any one of those three legs slips, the breakout above $250 will look, in hindsight, like a single Bloomberg headline doing too much work.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. Securities mentioned carry market risk, including the possibility of loss of principal, and reported deals may not formalize on the terms described. Consult a licensed financial professional before acting on any of the information here. All figures are accurate as of publication.