Qualcomm's Dragonfly Bet: Meta, Microsoft, and $15 Billion

For eight months, the bear case on Qualcomm’s data center ambitions was one word: customers. The AI200 accelerator reveal last October moved the stock 11% in a day, but skeptics could fairly ask who, besides a Saudi state-backed AI venture, was actually buying. At its Investor Day in Manhattan on June 24, Qualcomm answered with the two names that matter most in a capex cycle: Meta and Microsoft. The stock jumped another 12% after hours. Time to take the numbers apart.

What was announced

The event formally unveiled Dragonfly, the umbrella for Qualcomm’s data center silicon, and its first CPU — the Dragonfly C1000: more than 250 cores per socket, high memory bandwidth, PCIe Gen 7, and CXL support. Alongside it:

• Meta signed a multigenerational agreement to put C1000 CPUs into its next-generation server fleet — announced with Mark Zuckerberg lending his face to the deal.
• Microsoft confirmed Qualcomm silicon is going into some Azure data centers.
• Qualcomm announced a $3.92 billion acquisition of Modular, the AI-software company whose inference stack is one of the more credible attempts at a portability layer above CUDA.
• Management guided the data center business to roughly $5 billion in revenue by fiscal 2027 and $15 billion by fiscal 2029.

For scale: Qualcomm’s whole company did $10.37 billion in revenue in the quarter it announced the AI200. Management is promising to conjure, in four years, a business more than half that size per year out of a segment that barely existed on its income statement in 2025.

The strategy is a decade of acquisitions converging

The Dragonfly story reads like a receipts pile. The 2021 Nuvia acquisition ($1.4 billion) bought the Oryon CPU team that now spans Snapdragon laptops to the C1000. The Alphawave deal ($2.4 billion, closed December 2025) bought the high-speed connectivity IP — SerDes, chiplets — that data center parts live or die by, and Alphawave’s founder now runs Qualcomm’s data center business. The AI200 and AI250 accelerators scaled up the Hexagon NPU from phones. Modular is the software layer meant to make any of it usable without a CUDA rewrite. And a 2025 agreement to support Nvidia’s NVLink Fusion means Qualcomm CPUs can sit inside Nvidia-accelerated racks rather than only competing with them.

Each piece is defensible. The question was always whether they’d assemble into revenue. Meta and Microsoft are the first evidence that they might.

Why those two names move the needle

A CPU deal with Meta is not a science project; Meta’s server fleet is one of the largest on earth, and “multigenerational” implies design-in, not a trial rack. It’s also a telling choice given that Meta designs its own MTIA accelerator in-house — apparently the build-versus-buy math still favors buying CPUs.

Microsoft cuts the other way, and that’s what makes it interesting: Azure already deploys Microsoft’s own Arm CPU (Cobalt). Adding Qualcomm anyway suggests hyperscalers want a second-source lever against both Intel/AMD pricing and their own internal silicon roadmaps. Second-source is not a glamorous position — but in a market this size it’s a lucrative one.

The earlier anchor still matters too: Humain’s 200 megawatts of AI200 capacity begins deploying this year, giving the accelerator line a flagship reference just as the AI capex boom makes every hyperscaler hunt for cheaper inference. At today’s data center economics, 200 MW is a billion-dollar-class commitment before a single chip margin is counted.

The bear case hasn’t disappeared

• Qualcomm has been here before. Centriq, its 2017 Arm server CPU, was technically respectable and commercially dead within about a year, abandoned amid cost cuts. Data center customers remember who leaves.
• The accelerator fight is brutal. Inference is the growth segment, but AMD is attacking it with HBM-heavy parts and Nvidia owns the software gravity. Qualcomm’s memory-first design must prove itself in production, not keynotes.
• Targets are not bookings. $15 billion by fiscal 2029 assumes flawless execution across CPU, accelerator, networking, and a just-acquired software stack — while integrating three acquisitions.
• The customers hedge by design. Meta and Microsoft anchor several vendors at once; being one of many second sources caps pricing power.

The takeaway

June 24 was the day Qualcomm’s data center story stopped being hypothetical: a named CPU with 250-plus cores, two hyperscaler anchors willing to say so on stage, a $3.92 billion software acquisition to blunt the CUDA problem, and hard revenue targets — $5 billion by fiscal 2027, $15 billion by 2029 — that management can now be graded against. The diversification logic is sound for a company still leveraged to handsets, and the after-hours 12% says the market wants to believe. Belief has a maturity date, though: watch fiscal 2027, when the first $5 billion is due.