This talk delivers a data-driven summary on the physical and economic bottlenecks in the AI infrastructure market today. We will cover the diverging strategies between traditional hyperscalers and specialized "Neoclouds," supported by deep supply chain data. The audience will gain understanding of trends across the industry, from DRAM price hikes to CoWoS constraints, inference performance to behind the meter power generation for GW-scale datacenters.
What the audience will learn
- The Reality of the Power Wall: Why the commercial grid is failing to meet AI demand and how the industry is pivoting to "behind-the-meter" power solutions (like onsite gas turbines) to bring gigawatt clusters online.
- Neocloud vs. Hyperscaler Economics: A comparative framework for understanding the cost-performance trade-offs between renting from AWS/Azure versus specialized GPU clouds, backed by real-world benchmarking data.
- Supply Chain "Ground Truth": A detailed look at the actual constraints in the hardware supply chain (specifically HBM yield and advanced packaging) and a realistic timeline for when these bottlenecks will alleviate.
- The Next-Gen Network: Insights into the battle between InfiniBand and Ultra Ethernet, and why optical interconnects are becoming critical for scaling beyond the individual rack.
Interview:
What is your session about, and why is it important for senior software developers?
This session analyzes the physical and economic bottlenecks in the AI infrastructure market. It provides data to inform people on constraints in the supply chain, and describe how hardware and power density dictate software scaling limits. Understanding these physical constraints is necessary for building large scale AI systems. We want to be able to anticipate the future.
Why is it critical for software leaders to focus on this topic right now?
AI demand has exceeded the capacity of our existing cloud infrastructure. Many AI leaders must evaluate the cost structures of specialized providers to make their business work. These infrastructure choices define the performance limits and operating costs of AI deployments in 2026.
What are the common challenges developers and architects face in this area?
Architects face technical trade offs when scaling systems up. Balancing performance with cost and hardware availability is the primary challenge for building and deploying AI-powered products.
What's one thing you hope attendees will implement immediately after your talk?
Attendees should audit their infrastructure roadmap and prepare for supply constraints. They should do this while keeping reliability, security, and performance in mind.
What makes QCon stand out as a conference for senior software professionals?
QCon prioritizes practitioner led technical sessions over marketing. It facilitates technical exchange between developers who are in pursuit of truth.
Speaker
Jordan Nanos
Member of Technical Staff @SemiAnalysis, Previously Distinguished Technologist @HPE
Jordan Nanos is a Member of Technical Staff at SemiAnalysis and a former Distinguished Technologist at HPE. He lives in Squamish, British Columbia, Canada.
