AI Infra Summit 2026: Building Infrastructure for the Agentic AI Era

What you should know: 
• Qualcomm is bringing its Qualcomm Dragonfly™ portfolio to AI Infra Summit 2026, 
showcasing a rack-scale platform purpose-built for agentic AI inference.  
• Tony Pialis, EVP & GM, Data Center, will share Qualcomm's vision for AI 
infrastructure, centered on a new metric for the AI era: tokens per watt, not FLOPS.  
• Vinesh Sukumar, VP, product Management, will highlight how Qualcomm is 
positioning itself at the forefront of the agentic AI era by championing a distributed 
AI architecture. 
• Demonstrations will highlight breakthroughs in inference efficiency, memory 
architecture, connectivity, sovereign AI, and enterprise-ready agentic AI 
deployments. 
• A Hackathon is scheduled to bring together developers and innovators building on 
top of Qualcomm platforms. 
As AI moves from generating responses to reasoning, planning, and acting autonomously, 
infrastructure requirements are changing fundamentally. Agentic AI introduces a new class 
of workloads that demand dramatically higher inference capacity, greater memory 
efficiency, lower latency, and more sustainable economics. 
At AI Infra Summit 2026, Qualcomm Technologies is showcasing how it is helping the 
industry make that transition. The shift toward agentic AI plays directly to Qualcomm 
Technologies’ strengths. For decades, we have focused on solving complex compute 
challenges under strict power constraints. Today, we're bringing that expertise to the data 
center through Qualcomm Dragonfly™, our rack-scale AI infrastructure portfolio designed 
for the next generation of inference workloads. Our vision is straightforward: the future of AI 
infrastructure will be defined by how efficiently systems generate useful AI work, making 
tokens per watt a more meaningful measure of performance than traditional compute 
metrics alone.  
Tony Pialis: Reshaping Infrastructure for the Agentic Era 
During his AI Infra Summit keynote on Wednesday, Sept. 16 at 2:40 p.m. on the Main Stage, 
Tony Pialis, EVP & GM, Data Center, Qualcomm Technologies, Inc., will expand on the 
strategy we introduced at Investor Day earlier this year: agentic AI is driving the industry's 
transition from infrastructure optimized for training toward infrastructure optimized for 
large-scale inference.  
As agentic workloads continue to scale, data centers face increasingly difficult challenges 
around power consumption, memory bottlenecks, and total cost of ownership (TCO). We 
believes the industry must move beyond traditional approaches that prioritize peak 
performance at any cost and instead embrace architectures built around efficiency, 
scalability, and system-level optimization.  
At the center of this strategy is Qualcomm Dragonfly, a rack-scale platform spanning IP, 
silicon, cards, software, and complete rack solutions across four product areas: CPUs, AI 
accelerators, custom silicon, and connectivity. Together, these technologies are designed 
to help hyperscalers, cloud providers, and enterprises deploy AI infrastructure that delivers 
more intelligence within the same power envelope.  
Qualcomm Dragonfly solutions are set to drive industry-leading TCO through superb 
efficiency, with metrics including: 
• Up to 8x better tokens per second per watt versus GPU-based systems1 
• Up to 200x greater memory capacity per watt versus SRAM-based approaches2 
• Up to 6x higher memory bandwidth per watt versus conventional HBM-based 
designs3 
• More than 2x better performance per watt than the latest server CPU systems4 
These innovations are designed to address one of the central challenges of modern AI 
infrastructure: keeping increasingly sophisticated models fed with the memory bandwidth 
and capacity they require while maintaining sustainable power consumption.  
As Tony often notes, the next phase of AI won't be constrained by compute alone. It will be 
defined by how efficiently infrastructure can transform energy into useful AI output.  
Connectivity Matters More Than Ever 
As AI factories grow larger and more distributed, networking becomes a critical 
determinant of overall system performance. During the summit, Tony will join Lisa 
Spelman, CEO of Cornelis Networks, during her session, “The Network Is the Computer 
— and It's Time It Acted Like One,” on Tuesday, Sept. 15 at 3:00 p.m. on the Data 
Movement Track, to discuss the evolving role of connectivity in large-scale AI systems.  
The discussion will explore how efficient data movement, advanced fabrics, optical 
networking, and tightly integrated connectivity architectures are becoming increasingly 
important as agentic AI workloads scale across racks, clusters, and entire data centers. 
Powering Scale Across the AI Compute Continuum 
Qualcomm Technologies' presence at AI Infra Summit extends beyond the data center. On 
September 17 at 9:55 a.m. on the Main Stage, Vinesh Sukumar, Vice President of Product 
Management, will deliver a keynote, "Powering Scale: The Evolution of AI Systems," 
exploring how AI is evolving from conversational experiences into autonomous agents that 
reason, take action, and operate across devices, networks, enterprises, and cloud 
environments. As agentic AI drives exponential growth in token demand, Vinesh will 
discuss why the industry's next challenge is not simply building more compute, but 
building AI systems that can deliver lower cost, lower latency, and greater efficiency at 
scale. His presentation will highlight Qualcomm Technologies' vision for a distributed AI 
future where devices, the intelligent edge, on-premises infrastructure, networks, and cloud 
data centers work together as a unified compute fabric. Through real-world demonstrations 
and ecosystem collaborations, Vinesh will showcase how Qualcomm® technologies are 
helping create the distributed foundation needed to support agentic AI across the compute 
continuum, unlocking better economics and more scalable AI experiences for businesses 
and consumers alike. 
Real-World AI in Action 
Visitors to the Qualcomm booth will see firsthand how Qualcomm Dragonfly™ technologies 
translate into practical AI deployments. 
Running Extremely Large Models on Qualcomm Dragonfly 
One featured demonstration highlights the scalability of the Qualcomm Dragonfly AI 
accelerator architecture by running Kimi-K2.5, a one-trillion-parameter model, on a single 
Qualcomm Dragonfly AI200 accelerator card. This proof-of-concept showcases memory 
capacity and system-level design advantages that make it possible to support extremely 
large models within a compact footprint. Rather than focusing on benchmark 
comparisons, the demonstration illustrates how Qualcomm Technologies' architecture 
approaches memory scalability differently to unlock new deployment possibilities for 
advanced inference workloads. 
Bringing Security and Governance to Agentic AI to Drive Enterprise Innovation 
As autonomous AI systems become more capable, enterprises face new governance and 
security challenges. That's why Qualcomm Technologies is collaborating with Confidential 
Core AI (CCAI) to develop infrastructure designed specifically for agentic AI applications on 
Qualcomm Dragonfly solutions. Together, the companies are demonstrating a layered 
approach, combining technologies that help keep AI agents operating within defined policy 
boundaries and an encryption environment that keeps enterprise know-how and 
proprietary information under the enterprise's own ownership and control.  
The platform is designed to help organizations establish alignment and confinement 
controls, manage risk, and maintain ownership of sensitive enterprise information. By 
enabling sovereign deployment models, encryption technologies, and infrastructure 
optimized for AI inference, Qualcomm Technologies and CCAI are helping enterprises 
adopt agentic AI without sacrificing security, compliance, or control over proprietary data. 
Token Economics: More Models per Rack 
We are also working with Multiverse Computing to show how their CompactifAI model 
compression and Qualcomm Dragonfly AI infrastructure can work together to improve 
large-scale inference economics by reducing model memory footprint and bandwidth 
requirements.  
The story is not simply faster inference; it is about making memory the advantage. By 
reducing how much memory each model requires, operators can fit larger models, support 
longer contexts, and create more workload capacity per accelerator per rack, serving more 
sessions within the same rack power envelope, and running more models concurrently 
without expanding infrastructure.  
Together, Multiverse and Qualcomm Technologies are reinforcing the conversation toward 
the metrics that matter most in real deployments: served capacity, rack-scale value 
proposition: more models per watt, lower cost per token, and ultimately better economics 
for large-scale AI inference deployments. 
Visit Qualcomm Technologies at AI Infra Summit 
The demos above represent only a portion of what Qualcomm Technologies will be 
showcasing throughout the event. 
Whether you're interested in hyperscale AI infrastructure, sovereign AI deployments, 
networking innovation, custom silicon, memory architecture, or agentic AI applications, 
our experts will be available throughout the conference to discuss the technologies 
shaping the next era of AI infrastructure. 
Stop by the Qualcomm booth 206 to explore the full Qualcomm Dragonfly portfolio, see live 
demonstrations, and engage directly with the engineers and leaders building the future of 
AI inference. 
We also encourage developers and innovators to check out the Qualcomm Technologies 
sponsored AI Infra Summit Hackathon where you can learn more about our AI software 
offerings including tools, framework integrations, and model optimizations. You will also 
have an opportunity to build production-grade AI applications on a variety of Qualcomm® 
platforms, including the latest AI PCs with Snapdragon® processors and Arduino® UNO™ Q 
boards. 
Join the Team Building What's Next 
As Qualcomm Technologies expands its data center business, we are actively looking for 
engineers, architects, researchers, and innovators who want to help solve some of the 
industry's hardest challenges in AI, compute, memory, networking, and system design. If 
you're passionate about rethinking infrastructure for the agentic AI era and helping build 
technologies that can power the next generation of intelligent systems, we'd love to meet 
you. 
The infrastructure revolution is underway. At AI Infra Summit 2026, we're excited to show 
what's possible when efficiency, intelligence, and scale come together. 
Go Deeper 
What is Qualcomm Dragonfly, and how does it fit into the shift toward agentic AI? 
As AI moves from generating one-off responses to continuously reasoning and acting, the 
infrastructure underneath has to change with it. Qualcomm Dragonfly is our answer: a 
rack-scale portfolio spanning IP, silicon, cards, and complete rack solutions across CPUs, 
AI accelerators, custom silicon, and connectivity, designed so hyperscalers, cloud 
providers, and enterprises can coordinate compute, memory, and connectivity as one 
system. Explore Qualcomm Dragonfly and learn how it anchors our broader data center 
vision for the agentic AI era. 
What makes Qualcomm Technologies’ approach to data center AI inference different 
from the status quo? 
The industry has long optimized infrastructure for peak compute throughput, but agentic AI 
changes the math: what matters now is how much useful output a system can sustain 
within a fixed power budget. What sets our approach apart is that we bring together 
leading-edge AI, high-performance low-power computing, and unrivalled connectivity as 
one coordinated stack rather than discrete parts, reorienting the whole system around 
efficiency and useful output instead of raw FLOPS. Learn more about how our inference
first architecture and efficiency-first design translate into real infrastructure. 
Qualcomm branded products are products of Qualcomm Technologies, Inc. and/or its 
subsidiaries. Qualcomm, Snapdragon and Qualcomm Dragonfly are trademarks or 
registered trademarks of Qualcomm Incorporated. Arduino and UNO are trademarks or 
registered trademarks of Arduino S.r.l. 
¹ With Qualcomm® High Bandwidth Compute. Qualcomm estimates compared to 
contemporary GPU-based architectures on decode performance for select models.  
² With Qualcomm® High Bandwidth Compute . Qualcomm estimates compared to 
competing published product specifications normalized at rack-level.  
³ With Qualcomm® High Bandwidth Compute . Qualcomm estimates compared to 
competing published product specifications normalized at card-level. 
⁴ With Qualcomm Dragonfly™ C1000 CPU. Qualcomm estimates compared to existing 
product benchmarks for server CPU competitive offerings based on specs.