Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Charles Yeomans, CEO and co-founder of Atombeam
For the past several years, conversations about artificial intelligence (AI) infrastructure have been dominated by compute. Faster accelerators. Larger clusters. More specialized silicon. But by 2026, it will become clear that compute is no longer the limiting factor. Data movement is.
This is not a new problem, but the rapid expansion of AI workloads and hyperscale data centers is finally forcing it into the spotlight. Organizations are investing billions in processing power, yet much of that capacity sits idle, because the data they depend on cannot reach the compute resources quickly, consistently, or economically enough.
The result is a growing mismatch between how fast we can compute and how effectively we can feed that compute with data.
An Old Problem, Exposed at Scale
At a fundamental level, modern AI systems are not compute-bound, they are data-bound. Massive datasets must traverse complex hierarchies: across wide-area networks, between data centers, through racks and boards, into memory, and caches. Each step adds latency, cost, and variability.
This is not unlike challenges the industry has faced before. Decades ago, computing hit similar limits until memory hierarchies were rethought to ensure the right data arrived at the right place at the right time. Today, we are facing a modern version of that same issue; only now the scale is global, and the stakes are significantly higher.
One of the clearest signals of this imbalance is utilization. Even in cutting-edge AI environments, that inefficiency becomes increasingly expensive – and not just in infrastructure costs, but in lost opportunity.
Why Latency Variability Matters
Average performance metrics no longer tell the full story. What increasingly matters is tail latency, especially the 99th percentile. The slowest data transfer often determines when a task completes. One delayed response, one congested link, or one inefficient transfer path can stall an entire pipeline.
By 2026, organizations will recognize that reducing variability in data delivery is just as important as increasing raw throughput. Faster networks alone are not enough. The focus will shift to minimizing the amount of data that must move in the first place – and ensuring it can be accessed efficiently across multiple layers of the infrastructure stack.
This applies not only within data centers, but across cloud environments as well. As data is replicated, cached, and moved between regions, the cost and performance impact of inefficient transfers compounds quickly. Egress fees, bandwidth constraints, and unpredictable latency all become part of the equation.
The Edge Makes the Problem Urgent
The same dynamics are playing out beyond centralized infrastructure. Edge computing is often discussed as a simple extension of the cloud, but in reality it represents a far more distributed and complex system. Data flows from devices through networks that may rely on constrained links such as satellite, cellular, or intermittent connectivity. Events are processed and stored at multiple points before ever reaching centralized applications.
In these environments, efficiency is not optional. Latency budgets are tight. Bandwidth is limited. And repeatedly decoding, transforming, or expanding data at each hop can consume more time than the actual computation.
In 2026, you can expect that edge computing will no longer be framed primarily as a location problem – i.e., where processing happens – but instead as a data movement problem: how information flows, how much of it moves, and how intelligently it is handled along the way.
A Shift in Thinking
The next phase of infrastructure evolution will focus less on raw speed and more on balance. Keeping compute resources fully utilized will require rethinking how data is represented, transmitted, cached, and stored across distributed systems. That means prioritizing efficiency at every layer; not just faster links, but smarter movement.
This shift will also reshape economics. Reducing the volume of data that must traverse networks lowers costs as much as it improves performance. As AI workloads scale, these savings become material.
Most importantly, this is not a problem that will disappear as hardware improves. Faster and larger models only raise the bar. Without corresponding advances in how data moves through the system, inefficiencies will continue to surface, often in the most expensive and least predictable ways.
Next year, the industry will stop treating data movement as a secondary concern – and start recognizing it as a strategic one. Successful organizations will be those that view efficiency not as an optimization, but as a foundational requirement for scalable AI.
##
ABOUT THE AUTHOR
Charles Yeomans is the co-founder and CEO of Atombeam. With 25+ years in executive roles and investment banking, Charles has led multiple firms and founded successful companies, including major insurance brokerages. He’s a former U.S. Navy intelligence officer with an AB from Kenyon and an MBA from Stanford.
