Qualcomm Investor Day: one CPU, one memory technology, and a $40 billion target

Text | Bo Yang

Editor | Xu Qingyang

On June 25, local time in the U.S., Qualcomm held its 2026 Investor Day in New York.

Qualcomm announced a complete roadmap for data center AI infrastructure, unveiling the Dragonfly C1000 CPU, the AI300 inference accelerator, and high-bandwidth computing (HBC) technology. It also announced multi-generation collaboration with Meta, expanded cooperation with Hugging Face, and the acquisition of AI software company Modular.

Qualcomm’s fiscal 2029 non-handset revenue target

On the financial side, Qualcomm raised its fiscal 2029 non-handset business revenue target to $40 billion, nearly double its prior long-term target. Of this, data center business revenue in that fiscal year will exceed $15 billion.

In after-hours trading, Qualcomm’s stock rose by as much as 16%.

01 Data center revenue to exceed $15 billion

Qualcomm CFO Akash Palkhiwala predicted during the event that in fiscal 2027, Qualcomm’s data center business would generate “billions” of dollars in revenue. By fiscal 2029, annual revenue for this business will exceed $15 billion.

Judging from the company’s overall revenue mix, by fiscal 2029, QCT (semiconductor) segment non-handset business revenue will reach $40 billion, while the long-term target set in 2024 for that figure was $22 billion.

By fiscal 2029, Qualcomm’s handset business will account for only about one-third of QCT revenue.

The remainder is shared by several growth engines: automotive revenue of $10 billion and IoT revenue exceeding $14 billion. Among this, IoT includes industrial, networking, and robotics ($8 billion), as well as personal AI and computing ($6 billion).

Profit guidance was also raised.

Analysts’ average expectation for Qualcomm’s adjusted earnings per share in fiscal 2029 is $15.26, while Qualcomm’s own target is above $18. This gap is the direct reason for the after-hours jump in the stock price.

When explaining the growth logic, CEO Cristiano Amon put the focus on changes in how AI is used. He believes AI is moving from simple question-and-answer to agentic applications—models that can autonomously execute multi-step tasks. Workloads like this demand more low-power computing, and Qualcomm’s capabilities in mobile chips are precisely in this area.

Amon also said that AI computing is entering automobiles, everyday electronic devices, and robotics, and chip demand in these areas will continue to “open up.”

02 Dragonfly C1000 makes its debut, and Meta becomes the first customer

The main highlight of the hardware announcements was Dragonfly C1000, a CPU Qualcomm designed specifically for data centers.

Dragonfly C1000 is based on a custom-designed Oryon core, using a multi-chiplet (Chiplet) architecture, integrates more than 250 cores, and runs at frequencies above 5GHz. Qualcomm’s performance tests show that its per-watt performance is more than double that of existing server CPU competitors’ benchmarks.

Dragonfly C1000 supports PCIe Gen 7 and CXL connectivity. Its memory system uses low-power memory technology and includes RAS features such as built-in ECC, fault isolation, and error recovery. The cooling solution is compatible with both air and liquid cooling, and the rack meets the OCP ORv3 standard.

Rack configurations equipped with Dragonfly C1000 were also released: 43TB of DRAM, with sample shipments expected in fiscal 2026.

Qualcomm has outlined three sub-segments for this CPU:

The first category is the agent CPU, targeting high-throughput agent orchestration and low-latency interactive AI tasks.

The second category is a general-purpose CPU, balancing two types of needs: when running first-party workloads, it targets optimal TCO (total cost of ownership) performance; for elastic usage by third parties, it targets optimal vCPU (virtual central processing unit) performance.

The third category is an AI head node CPU. Its role is to complete host processing with low overhead, so the XPU can run as fully as possible in generative AI computing.

What truly gave Dragonfly C1000 weight was Meta’s endorsement.

Qualcomm announced that the two sides have signed a “multi-year, multi-generational” agreement. Meta will use Dragonfly C1000 for its next-generation server cluster, with the chip planned for mass production in the second half of 2028. Subsequent CPU iterations are also within the scope of the cooperation.

Qualcomm CFO Palkhiwala said that through mobile chips and other existing products, Qualcomm already has business relationships with nearly all hyperscalers. “This is not a newly established relationship.” This implies that Meta is very likely not the only negotiation target, and more customers may still be in discussions.

In response to the outside question of whether Qualcomm’s entry into the data center is too late, CEO Amon said: “When people ask whether it’s too late to enter the data center now, you should think about scale and execution capability, engineering capability, or operations and the supply chain.”

His point is that the large-scale systems engineering capabilities Qualcomm accumulated during the mobile era remain effective in this market.

03 AI accelerator plus HBC to break the “memory wall”

Beyond the CPU, Qualcomm also updated its roadmap for AI accelerators.

After the previously released AI200 and AI250, the AI300 inference accelerator was unveiled at this Investor Day, and the three products iterate on an annual cadence.

The core logic of this platform is “decoupled rack-scale AI inference.” Tony Pialis, Executive Vice President and General Manager of Qualcomm’s Data Center Business, explained that agent workloads require coordination among the CPU, AI accelerator, and interconnect technology, rather than relying on a single chip. What Qualcomm is doing now is integrating compute, AI, memory, and connectivity into a unified rack-scale platform.

Within this platform, the memory problem is a hurdle you can’t bypass, and Qualcomm’s solution is high-bandwidth computing (HBC).

This is a technology designed to break the “memory wall.” The so-called memory wall refers to the bandwidth bottleneck in AI computing when data is moved between the processor and memory. HBC’s approach is to tightly integrate compute units and memory through 3D stacked silicon technology, following a near-memory computing route.

Qualcomm provided several sets of data to illustrate HBC’s potential.

With AI250 featuring HBC Gen 1, the effective memory bandwidth per card reaches 133 TB/s, which is an 18x improvement over AI200 using LPDDR5X. With AI300 featuring HBC Gen 2, the bandwidth increase compared with AI200 will reach 54x.

Compared with today’s mainstream HBM (high-bandwidth memory), HBC delivers 6 times the bandwidth at the same power consumption. Compared with SRAM (static random-access memory), at the same power consumption, HBC offers 200 times the capacity.

In other words, with HBC, the amount of data that can be processed per unit of power is significantly increased, which has a direct impact on data center total cost of ownership (TCO). Commercial samples of AI250 are expected to be provided in mid-2027, while commercial samples of AI300 will arrive in 2028.

Connectivity products are Qualcomm’s long-established strength, and this time it was also not absent. The company provides interconnect solutions from Die-to-Die, copper cables, and optical fiber to campus-level, supporting 800G and 1.6T rates, covering scenarios from within the data center to the longest distance of 20 kilometers.

More than 35 technology ecosystem companies have publicly expressed support for this roadmap. The list includes Supermicro, Lenovo, SK Hynix, Micron, Samsung SDS, and Arista.

04 Acquires Modular and partners with Hugging Face

Beyond hardware, Qualcomm also took dense action in its software ecosystem.

First is the acquisition of AI software company Modular. The acquisition consideration is approximately $3.9 billion worth of Qualcomm stock. The deal is expected to close in the second half of 2026, subject to regulatory approval.

Modular’s core product is an open, AI-native software stack that allows models to run on different chip architectures such as CPU, GPU, NPU, and custom ASIC, so developers do not need to rewrite code for each type of hardware. Modular was co-founded by Chris Lattner and others, and its platform is regarded in the industry as an open alternative to Nvidia’s CUDA.

When commenting on the acquisition, Amon said that after agents expand from the data center to the edge, the industry needs a more open and modern software foundation. Qualcomm hopes that through this acquisition, it will give customers truly viable deployment choices across diverse computing environments.

Second is expanding cooperation with Hugging Face. The cooperation consists of three parts:

* Bring Hugging Face’s internal and developer workloads into data centers driven by Qualcomm Dragonfly;

* On the Hugging Face platform, more than 3 million open-source models can be directly loaded onto devices and data center racks equipped with the Qualcomm platform, simplifying the process from experimentation to deployment for developers;

* Develop “Hugging Face Agents” to orchestrate AI workloads in hybrid environments spanning both the device side and the cloud, dynamically assigning tasks based on performance, cost, and latency requirements.

Hugging Face co-founder and CEO Clément Delangue explained: “We are enabling our 16 million developers to run open models easily anywhere—from the devices in your hands to the complete racks in the data center.”

There is also a specific arrangement within the partnership. Hugging Face will provide Hugging Face PRO access to customers using devices or cloud systems powered by Qualcomm platforms, including advanced storage, compute, and collaboration features.

This step lowers the barrier for developers to build applications with open models.

05 Automobiles, robotics, China

Beyond the main data center storyline, Qualcomm also updated progress across other businesses.

For the automotive business, the “automotive design project win pipeline” has expanded to $65 billion, and Qualcomm raised its fiscal 2029 revenue target to $10 billion. Underlying demand for automotive chips is the continued penetration of ADAS and autonomous driving.

The IoT business has been broken down into finer categories. Industrial, networking, and robotics are listed separately, with a target revenue of $8 billion; personal AI and computing targets $6 billion. Qualcomm believes that agents will trigger a new upgrade cycle for smart connected devices. The company estimates that by 2030, the combined scale of these businesses will reach $1.7 trillion.

Regarding the China market, Amon gave a brief response at the event. The U.S. government currently has regulations on exporting AI-related hardware to China, but he said Qualcomm will have data center chip versions that do not trigger export restrictions. He did not go into specifics about the plan, but this statement indicates that the opportunities in the China market have not been put on hold.

Overall, the signals released by Qualcomm at this Investor Day were fairly complete. From C1000 to HBC, from acquiring Modular to collaborating with Hugging Face, from the $15 billion data center target to $18 in earnings per share, all are verifiable milestones. Customers are secured, there are product sample timelines, and the financial model has been laid out.

In the next few quarters’ earnings reports, there will be a first round of testing of Qualcomm’s roadmaps.