Qualcomm announced the Dragonfly C1000 data center CPU at its Investor Day on June 24, targeting $15 billion in annual chip sales by 2029, with Meta as its first named customer in a multi-generational supply agreement.
Three days ago, Qualcomm stood up at its Investor Day in New York and told the market something it hadn't quite said before: the smartphone business is no longer the whole story. The company unveiled the Dragonfly C1000, a data center CPU built on its custom Oryon cores, featuring more than 250 cores clocked at 5GHz, PCIe Gen7 and CXL support, and a chiplet architecture aimed squarely at Intel and AMD's server dominance. Mark Zuckerberg confirmed on the same day that Meta has signed a multi-generational supply agreement to deploy the C1000 across its facilities, with production scheduled for the second half of 2028. As CNBC reported, this is Qualcomm's most direct bid yet at owning a slice of the infrastructure powering the AI era.
The timing matters. Qualcomm also has its AI200 and AI250 inference accelerators in the pipeline, both based on its Hexagon NPU architecture and targeted at rack-scale AI inference. The AI200 carries 768GB of LPDDR memory per card and is expected to ship commercially in 2026. The AI250, due in 2027, introduces what Qualcomm calls a near-memory computing architecture, claiming over ten times the effective memory bandwidth of conventional designs. Qualcomm positions both against Nvidia not on raw peak performance, where GPUs still hold the architectural edge, but on power efficiency and total cost of ownership. For enterprises running large-scale inference workloads around the clock, that argument has real weight.
But the more interesting story for founders and investors isn't the data center chip. It's what Qualcomm is doing at the other end of the compute stack, and how the two ends are starting to converge.
On Snapdragon 8 Elite hardware, Qualcomm's Hexagon NPU can now run models up to 20 billion parameters entirely on-device, reaching up to 220 tokens per second. On a Samsung Galaxy S25 Ultra, models like Granite 4.0-h-350M hit 92 tokens per second on the NPU with up to nine times better energy efficiency than running the same workload on the CPU. Developers have already built fully functional edge AI inference servers on these devices requiring zero cloud connectivity. These aren't benchmarks in a lab. They're deployable today.
That changes the unit economics for any founder currently routing inference through an API. A mobile app making 100,000 daily API calls to run a medium-sized language model is paying for cloud compute, round-trip latency, and the data exposure that enterprise customers increasingly flag during procurement. Move that inference to the device and you cut all three. You also stop paying per token. The model cost shifts from operational to capital, and for consumer apps with heavy engagement, that math flips quickly.
Don't underestimate the enterprise angle either. Corporate buyers have spent the last two years demanding that sensitive data stay off third-party servers. On-device inference gives them that guarantee structurally, not just contractually. Healthcare, legal, finance: the verticals where data residency concerns kill deals fastest are exactly the ones where Qualcomm's edge silicon becomes a selling point a cloud-dependent competitor can't match.
Qualcomm's Investor Day target tells you how seriously the company reads this moment. The company raised its non-handset revenue target for fiscal 2029 to $40 billion, nearly double its prior figure. Data centers and edge AI are doing that work. The handset business, still substantial, is no longer the growth engine.
What this means if you're building
For founders building AI-native mobile applications, there's a real architecture decision forming that didn't exist eighteen months ago. Cloud inference is flexible and easy to iterate on. Edge inference on Snapdragon silicon is faster, cheaper at scale, and structurally private. The right answer depends on your model size, your update cadence, and what your users are willing to tolerate in terms of latency and data handling. But the question is now live in a way it wasn't before, and ignoring it means leaving a cost and differentiation advantage on the table.
For investors, Qualcomm's move is a signal worth tracking carefully. The company isn't just trying to sell more Snapdragon chips. It's positioning itself as an inference layer provider across every form factor, from the phone in your pocket to the rack in Meta's data center. That puts it in a different competitive conversation than it was in twelve months ago, one where Nvidia's inference dominance is being challenged not by a GPU rival but by a company with Oryon CPU cores, Hexagon NPUs, and an installed base of over a billion mobile devices. Qualcomm declined to comment on specific deployment timelines beyond what was presented at Investor Day. What it did say, clearly, is that the bet is on inference everywhere, and it's placed the chips to prove it.
Also read: AI's power hunger is turning energy infrastructure into the hottest new IPO category • Japan's Sakana AI and China's Qwen are rewriting the AI pecking order in ways US labs did not expect • Washington's AI export controls are handing Asian labs the frontier race they were supposed to lose