Qualcomm Dragonfly Aims to Redefine AI Data Centers with Accelerators, CPUs, Connectivity, and Custom Silicon
Qualcomm is making a major push into the AI data center market with Dragonfly, a new full-stack platform designed to bring AI compute accelerators, high-performance CPUs, advanced memory technology, next-generation networking, and custom silicon together under one ecosystem.
First teased around Computex 2026, Dragonfly is now being positioned as a complete data center platform for next-generation artificial intelligence and general-purpose computing. Rather than focusing on a single chip or component, Qualcomm is building a broader infrastructure strategy aimed at hyperscalers, cloud providers, enterprises, and AI companies that need more compute performance, higher memory bandwidth, better power efficiency, and faster connectivity.
The Dragonfly portfolio includes AI accelerators, the Dragonfly C1000 CPU, high-bandwidth capacity memory technology, scale-up and scale-out networking solutions, optical and copper interconnects, and custom chip design services. Together, these technologies show Qualcomm’s ambition to compete in one of the fastest-growing sectors in computing: AI data center infrastructure.
Qualcomm Dragonfly AI300 Targets Next-Generation AI Inference
One of the most important parts of the Dragonfly ecosystem is Qualcomm’s upcoming AI accelerator roadmap. The company has already introduced the AI200 and AI250 accelerators, with AI200 currently being sampled and AI250 planned for 2027.
The AI250 is expected to be the first Qualcomm accelerator to use HBC Gen 1 memory technology. This memory approach is designed to deliver extremely large LPDDR capacity while improving effective bandwidth and power efficiency. Qualcomm says AI250 can support up to 43TB of LPDDR memory capacity across air-cooled and direct-liquid-cooled rack configurations.
While the raw capacity remains similar to the AI200, which uses LPDDR5X, the addition of HBC technology is expected to deliver a major leap in memory performance. Qualcomm claims AI250 can provide up to 18 times more effective bandwidth and 5 times better bandwidth per watt.
By 2028, Qualcomm plans to sample its next-generation AI300 accelerator series. These accelerators are expected to support both air cooling and direct liquid cooling at the rack level, but the major upgrade will be HBC Gen 2 memory technology.
HBC Gen 2 is designed to push memory performance even further, with Qualcomm targeting up to 54 times higher effective bandwidth compared to AI200 and up to 8 times better bandwidth per watt compared with HBM-based solutions. This could be especially important for large language models, multimodal AI models, and agentic AI workloads, where memory capacity and bandwidth often become key performance limits.
The AI300 platform is being designed for high-throughput, low-latency inference, especially in disaggregated AI deployments. In modern AI data centers, compute, memory, and networking resources are increasingly separated and scaled independently. Qualcomm’s focus on memory bandwidth, power efficiency, and flexible infrastructure suggests Dragonfly is being built with that future in mind.
Qualcomm is also preparing AI300 for large-scale deployments through support for UALink, also known as Ultra Accelerator Link, and ESUN, or Ethernet for Scale-Up Networking. These technologies are intended to help multiple accelerators work together more efficiently inside a rack or across broader AI infrastructure.
The company also plans to support scale-out networking using both copper and optical connections, allowing AI clusters to expand across racks, rows, and potentially larger campus-scale environments.
Dragonfly Connectivity Platform Tackles AI Data Movement Bottlenecks
AI data centers are not limited by compute alone. Moving data quickly and efficiently between processors, memory pools, accelerators, storage, and network infrastructure has become one of the biggest challenges in modern AI deployments.
Qualcomm’s Dragonfly Connectivity Platform is designed to address that problem with a broad lineup of die-to-die, copper, optical, and campus-reach interconnect technologies.
The company is developing advanced SerDes technologies supporting 112 Gbps, 224 Gbps, and eventually up to 448 Gbps speeds. These interconnects are designed to move data within racks and between racks using active electrical cables and other high-speed links.
Qualcomm is also exploring optical technologies, including co-packaged optics and network packaged optics. These approaches are becoming increasingly important as AI data centers demand higher bandwidth, lower latency, and improved energy efficiency. By moving optical connectivity closer to compute and switching silicon, data centers can reduce power consumption and improve signal performance.
For scale-out networking, Qualcomm is working on a QAM16 coherent-lite optical solution designed for links of up to 20 kilometers. This could be useful for large campus data centers where AI workloads need to run across multiple buildings or facilities. The company is also developing PAM4 optical SerDes technology for optical reach of up to 2 kilometers.
In the 2026 to 2027 timeframe, Qualcomm’s Dragonfly Connectivity lineup is expected to include the O200, which supports 1.6T optical modules and active copper optical solutions, along with the CU200, which supports 1.6T active electrical cables.
By 2028, Qualcomm plans to expand the portfolio with the CO1600 for 1.6T long-reach and 3.2T FR2 optical connectivity, the O400 for 3.2T optical modules, and the CU400 for 3.2T active electrical cables.
This roadmap highlights Qualcomm’s goal of becoming a major provider of AI data center connectivity, not just AI compute chips. As AI models grow larger and more distributed, high-speed networking will be essential for maintaining performance and efficiency.
Custom Silicon Becomes a Key Part of Qualcomm’s AI Strategy
Beyond AI accelerators and connectivity, Qualcomm is also expanding into custom silicon for AI and cloud data centers. This could become one of the most important parts of the Dragonfly strategy.
Many hyperscalers and cloud providers are increasingly interested in custom chips designed for their own workloads. Instead of relying only on general-purpose processors or off-the-shelf accelerators, companies want specialized silicon that can deliver better performance per watt, lower latency, and tighter integration with their software and infrastructure.
Qualcomm plans to offer performance-optimized silicon at scale for next-generation AI and cloud platforms. The company is also preparing bespoke custom chips for agentic AI, specialized AI workloads, and other demanding applications.
The custom silicon offering includes end-to-end co-design across silicon, systems, and software. That means Qualcomm can work with customers to tune chips for specific performance, power, and integration requirements.
Advanced packaging is another major part of the plan. Modern AI processors increasingly depend on chiplet designs, stacked memory, high-speed interconnects, and modular architectures. Qualcomm says its approach will focus on improving performance, energy efficiency, and scalability through advanced packaging and modular design.
The company will also provide a proven IP stack and streamlined execution process to help customers reduce design risk and bring products to market faster. Qualcomm says it can support the entire process from design through high-volume manufacturing, backed by ecosystem and supply chain relationships.
This move puts Qualcomm in a stronger position to serve companies that want customized AI infrastructure rather than standard hardware platforms.
Why Dragonfly Matters for the AI Data Center Market
Qualcomm Dragonfly represents a major shift in the company’s data center ambitions. Instead of entering the market with only a CPU, accelerator, or networking product, Qualcomm is building a complete AI infrastructure platform.
The Dragonfly ecosystem combines several key components needed for future AI data centers: efficient AI accelerators, high-capacity memory technology, powerful CPUs, high-speed networking, optical and copper connectivity, and custom silicon design.
The focus on inference is especially important. While AI training often gets the most attention, inference is where many companies face the largest long-term cost and efficiency challenges. Running large language models, multimodal AI systems, recommendation engines, AI agents, and enterprise AI services at scale requires massive compute resources and significant power.
If Qualcomm can deliver strong performance per watt, large memory capacity, and efficient connectivity, Dragonfly could appeal to cloud providers and enterprises seeking alternatives to traditional GPU-heavy infrastructure.
The roadmap also shows that Qualcomm is thinking beyond individual chips. AI data centers are becoming complex systems where compute, memory, networking, power, cooling, and software all need to work together. Dragonfly is designed to address that full-stack challenge.
A Bold AI Infrastructure Play for 2027 and 2028
As the Dragonfly platform matures through 2027 and 2028, Qualcomm could become a more serious player in the AI data center market. The AI250 accelerator with HBC Gen 1 memory, the upcoming AI300 with HBC Gen 2, and the expanding Dragonfly Connectivity Platform all point toward a long-term strategy focused on scalable, power-efficient AI infrastructure.
The addition of custom silicon services also gives Qualcomm another way to compete, especially with customers that want workload-specific chips for AI, cloud computing, and specialized data center deployments.
Qualcomm Dragonfly is more than a product launch. It is a signal that the company wants to play a much larger role in the future of AI computing. With accelerators, CPUs, memory technology, connectivity, and custom silicon under one roof, Dragonfly could become a compelling platform for organizations looking to build the next generation of AI data centers.
