The Dragonfly has its first test flight at the Qualcomm Investor Day
Qualcomm
At the Qualcomm 2026 Investor Day, the company disclosed it’s data center roadmap and, perhaps more importantly, announced that Meta has entered into a multi-year, multi-generation agreement with Qualcomm to deploy its data center technologies. This is a strong affirmation of Qualcomm’s plans for the age of agentic AI, and extends the collaboration the two companies have embarked upon over the years in augmented reality. Leading up to the event, there were a lot of rumors of who that the data center customer was; Meta is an important win. (Note that Qualcomm, like many in the semiconductor industry, is a client of Cambrian-AI Research.)
The strategy is simple: leverage the company’s renowned strength in connectivity, create custom silicon for large clients, build an architecture that breaks the memory wall, and create rack-scale Oryon CPUs that are industry leaders in power efficiency. But this strategy will take time to materialize, with the Dragonfly Accelerators showing up in FY2027. Meanwhile The new CPU racks are needed now to feed the agentic AI frenzy, and new clients like Microsoft and Meta will have plenty to consume in the meantime.
QUalcomm has a four-pronged strategy to break into the data center: Leverage connectivity, deliver Custom Silicon, AI Accelerators, and fast CPUs.
Qualcomm
There are several technologies that are all coming together that CEO Cristiano Amon believes make up for being “late to the party”. First, the company has developed the Dragonfly C1000 derivative of the Oryon CPU specifically for data centers, doubling the performance per watt while running 250 cores at over five GigaHertz. It also supports the new PCIe Gen 7 with CXL connectivity for shared memory. More critically, it is packaged with the next innovation that Qualcomm believes with finally break through the memory wall, called High Bandwidth Computing.
The new Dragonfly C1000 CPU delivers twice the performance per watt in agentic workflows.
Qualcomm
This new memory architecture, called HBC for short, is an attempt to deliver the performance akin to SRAM, as found in Groq and Cerebras, with the capacity normally associated with DRAM. I suspect more details will be forthcoming as we get closer to first customer shipments.
HBC is intended to provide latency closer to SRAM with the memory capacity closer to HBM.
Qualcomm
First out the door will be racks of C1000 CPUs with 43TB of DRAM, in both air- and liquid-cooled systems. This is probably what Meta is deploying, and it will ship samples in FY26. FY27 will see the first version to support the HBC architecture, also at 43TB of capacity. Then in FY28 we will see a 2nd generation of HBC along with the scale-up UAL networking and optical scale-out. Again, Qualcomm plans to support both air- and liquid-cooled racks throughout the roadmap.
The Qualcomm roadmap for data centers starts this year with the A200, then adds the High Bandwidth Computing memory in 2027, and reaches 54X more bandwidth by 2028.
Qualcomm
All told, Qualcomm is predicting the AI250 will deliver an 18-fold improvement in bandwidth over the initial AI200 thanks to HBC, and 54-fold improvement with HBC Gen 2 the following year with the AI300. That’s some pretty startling numbers in the area of AI that everyone in the industry from Nvidia to AMD to Cerebras to Intel is trying to fix; the memory wall is tumbling down.
There will not be one winner that will challenge Nvidia. AI developers have many fine architectures to choose from, including Google, Amazon, AMD, and scores of others. Software will enable them to adopt these new innovations without a lot of overhead across the various architectures. That is why Qualcomm also announced just prior to the Investor event that they had acquired Modular, which provides an open AI stack akin to Nvidia’s stack to enable high performance and portable AI.
Qualcomm announced the acquisition of Modular, a company with an open, AI-native software stack that enables AI to run efficiently across different hardware architectures.
Qualcomm
I’m sure investors will find something to argue about in the Investor Day presentations, be it timing or veracity or affordability of the roadmap material presented. But I’d be hard pressed to argue with the aggressive and comprehensive roadmap Mr. Amon and his team shared today. Nice job, guys!
Disclosures: This article expresses the opinions of the author and is not to be taken as advice to purchase from or invest in the companies mentioned. My firm, Cambrian-AI Research, is fortunate to have enjoyed many semiconductor firms as our clients, including Baya Systems, Cadence, Cerebras Systems, D-Matrix, Flex, Groq, IBM, Infleqtion, Intel, Micron, NVIDIA, Qualcomm, SImA.ai, Synopsys, Taalas, Tenstorrent, Ventana Microsystems, and scores of investors. I have no investment positions in any of the companies mentioned in this article. For more information, please visit our website at https://cambrian-AI.com.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。