Beautiful insanity: AI buildouts lead to long lead times

The scramble to build AI factories has resulted in long lead times and power access concerns, affecting hyperscalers, operators and enterprises alike.

Hyperscalers are investing heavily in capacity and infrastructure to scale AI, but they don't necessarily know what to expect in the next few years, said Jeff Wabik, CTO at DC Blox, a digital infrastructure provider, during a fireside chat at Fiber Connect 2026 in Orlando.

"I see them planning and executing feverishly -- what if we need more fiber, more conduit, more power, more splices, more cross-connects," Wabik said. "It's beautiful insanity."

Companies like AWS, Google and Meta are spending hundreds of billions of dollars on these massive, centralized data centers that are currently focused on GPU-intensive model training. But as AI use cases mature -- along with the real-time inference that comes with them -- they'll need to shift to a distributed edge architecture that can process, analyze and respond to data within 10 milliseconds, closer to the data sources, said Peter Cresse, president of Entropy, a consulting firm, during a panel about data center innovations.

The existing internet ecosystem simply can't support that scale right now, said Brent Legg, executive vice president for government affairs at Connected Nation and IXP.US, which builds physical carrier-neutral data centers across the U.S. That ecosystem, which routes data traffic through centralized hubs in big cities, will break in 24 to 36 months, he predicted.

"As we move from large language model training to latency-sensitive applications that are inference-based, we need more places where networks interconnect and exchange traffic regionally, not just in big cities like New York and Atlanta," Legg said. "That evolution hasn't happened."

This has all led to a building and purchasing frenzy among hyperscalers and operators, causing equipment-buying lead times the industry hasn't seen since the Covid-19 pandemic, according to Robin Olds, senior sales business development manager at Cisco.

"At Cisco, we've been zapped by hyperscalers buying more than what they need -- they see what's coming," Olds said during the fireside chat with Wabik. "They're actively purchasing as much GPU, compute, electronics and optics as they can."

Another complicating factor is the fact that these AI factories are power-hungry. In various sessions, panelists repeatedly mentioned the challenge of generating adequate energy to support these AI facilities. As Sachin Gupta, vice president of business and technology strategies at Oklahoma-based ISP Centranet, said during a panel on scaling AI, "Whoever wins the power race wins the AI race."

With the level of compute that occurs in these data centers, hyperscalers require hundreds of gigawatts of energy -- but they can't access it, said Jason Eichenholz, founder and CEO of Relativity Networks, a fiber optics provider specializing in hollow-core fiber. For example, the lead time for transformers is between five and eight years, he added.

The industry is considering alternative energy options, such as solar and small modular reactors for nuclear energy. SMRs are more compact, provide stable power and have low carbon emissions. But commercial SMR-powered data centers aren't likely for another few years. Additionally, most hyperscalers have enough challenges contending with jurisdiction and permitting delays, let alone community concerns about nuclear power, DC Blox's Wabik said.

As a result, AI buildouts look a bit like throwing sprinkles in the air, watching them fall and plotting the course based on the outcome, he added.

"That's where the inference nodes are going, wherever you can find dirt, wherever you can find a political ecosystem that's not going to clobber you, wherever you can find a natural gas main so you can actually give them the power to make it work," Wabik said.