AI's Ceiling Isn't Compute. It's Glass.
By Brian Newman, Contributing Editor
Insight at the Intersection
Every AI capacity debate of the past two years has fixed on one resource: GPUs. Ciena's fiscal third-quarter earnings call pointed at a different constraint entirely. Revenue reached $1.67 billion, up 37 percent year over year and a company record. Backlog climbed to $8.5 billion, and management guided to more than $10 billion by the close of fiscal 2026.
CEO Gary Smith framed the driver plainly. Ciena expects its addressable market to roughly double, from about $25 billion today to about $50 billion by 2029, as AI-related networking demand expands across three fronts. The first is the traditional wide-area network. The second is a newer category some call the AI WAN, the backbone and data-center interconnect linking distributed training and, increasingly, distributed inference. The third is the optical fabric running inside the data center itself.
Compute has not stopped mattering. It has stopped being the whole story.
The Layer Nobody Priced In
Optical connectivity sits at the center of AI's real bottleneck. Training clusters move enormous data volumes between GPUs inside a single data center, and increasingly between separate data centers as distributed training spreads compute across multiple sites. Inference and agentic workloads add a second demand: constant, latency-sensitive traffic between cloud environments, AI-specific compute, and the applications or users waiting on a response. None of that traffic moves without fiber and the optical systems that light it.
Ciena's own product roadmap tracks the shift. WaveLogic 6 Extreme remains the only 1.6-terabit high-performance modem on the market 18 months after launch. RLS Hyper Rail, the company's newest intelligent line system, is built specifically for data-center interconnect and AI inferencing capacity. Vesta, an open co-packaged optics platform, already has sample orders from anchor customers, with revenue expected to begin in 2027 and ramp through 2028.
KPMG's own reporting adds the layer beneath Ciena's numbers. KPMG US technology principal Phil Wong told RCR Wireless that reliable power is the primary constraint on AI infrastructure scaling over the next three to five years, ahead of supply-chain delays and labor availability. Fiber ranks third on that list, but it is becoming the harder problem to solve in practice. New data-center developments increasingly land outside traditional population and business centers to chase power and land, and the middle-mile and long-haul fiber routes reaching them do not automatically pencil out. Operators now have to underwrite the return on each route individually, in markets fiber was never built to serve.
Inference is changing the traffic pattern, too. Wong noted that agentic AI, which needs constant access to data, context, and memory, is driving more traffic between traditional cloud environments and AI-specific compute than training workloads ever did. That traffic wants low latency, which means it wants proximity to the user. The network map AI infrastructure planners are drawing looks less like a handful of hyperscale campuses and more like a distributed grid of metro fiber, edge nodes, and resilient interconnection points.
Does AI infrastructure planning still treat fiber as a solved problem left over from the broadband build-out, or as a live constraint that determines which sites are even viable? When the industry's compute-scarcity narrative finally fades, optical capacity may turn out to have been the tighter ceiling all along.
Strategic Signal
Watch where control in AI infrastructure sits. Hyperscalers control the compute and the capital, but they do not control the physical path data has to travel, and increasingly they cannot build that path fast enough on their own. Companies that already own long-haul route miles, rights-of-way, and metro fiber plant are the ones now fielding order requests they did not generate. These are the same incumbents telecom spent a decade writing off as big dumb pipes.
Ciena's addressable-market math is one signal of this shift. Zayo has reported more than a billion dollars in AI-related bookings and has announced thousands of new long-haul route miles to reach AI compute sites. US Signal is building more than a thousand miles of new middle-mile fiber across Ohio and Indiana for the same reason. None of these companies build models. All of them are becoming load-bearing infrastructure for the companies that do.
The pattern echoes a shift telecom has already lived through once, when spectrum and network access outlasted several generations of device and application hype cycles. The physical layer takes longer to build and harder to replace, which is exactly why it accrues durable value once demand finally catches up to it. Expect the AI capital-spending conversation to keep splitting in two directions: money chasing a compute layer that gets cheaper and more commoditized every year, and money chasing an optical and fiber layer that gets scarcer and harder to route around every year.
Brian C. Newman is a telecom and AI strategy consultant, course creator, and former Verizon technology leader with more than 30 years of experience across wireless networks, 5G, network operations, infrastructure modernization, and emerging technologies. He helps organizations understand how AI, connectivity, edge computing, and digital infrastructure are reshaping business operations, real estate, public safety, and customer experience.
