Unlock the Editor’s Digest for free

The writer is global head of digital Infrastructure for KKR

When Thomas Edison lit up a block of lower Manhattan in 1882, the story wasn’t only about the lightbulb. It was also about the grid that needed to be built to light up the world. The early years of electrification were messy, with speculative booms, bankruptcies and rival standards. What endured was the infrastructure that propelled the modern economy.

Artificial intelligence is at a similar turning point. Breakthroughs in models and applications are breathtaking, and each leap increases pressure on the physical systems that power them.

It is tempting to compare the ascent of AI to the internet boom in the late 1990s. According to Wall Street estimates, so-called AI hyperscalers in the US and sector companies are expected to spend more than double what they spent in 2022 on data centre capital expenditure in 2025. AI-related capex accounts for about 5 per cent of US GDP and is growing by roughly 10 per cent a year.

Such growth makes many uneasy as they contemplate the parallels with the dotcom rush to build out infrastructure. After the dotcom bubble in market valuations burst, there was a brutal drawdown when the Nasdaq index plunged 78 per cent from peak to trough.

And so it is fair to ask whether AI is a bubble that will burst. It is true that hypergrowth can create periods of over-investment at the beginning of an industrial technology build-out, as enthusiasm outruns economics. But as a big investor in AI infrastructure, my firm sees strong potential both for returns and for societal benefits in general.

History offers perspective that infrastructure built is rarely wasted and forms a foundation that outlasts the cycle — even if not every investor is spared in the disruption. Fibre optic cable in the late 1990s is a case in point: many builders like WorldCom, Exodus and Global Crossing went bankrupt, yet the more than $500bn that sector companies invested in the boom on capital expenditure developed the backbone of the modern internet. Similarly, railway investors in the 19th century lost fortunes, yet the tracks they financed stitched together national markets.

In hindsight, far from being overhyped, the internet itself turned out to be underhyped. Entire industries like travel, retail, advertising and financial services were remade, even as many of the leading companies disappeared, and some of the biggest winners (eg, Google) didn’t even exist during most of the “bubble”.

AI’s physical footprint — data centres, power generation and transmission, cooling, fibre networks, etc — fits the same pattern as rail, electricity and telecom. Undoubtedly, there will be overshoots and shake‑outs. And some losers. Just as in an economic boom, not every lender will make prudent credit decisions.

The question though isn’t whether AI infrastructure is “needed”, but which infrastructure, including where, and under what constraints, will matter most. The next decade’s infrastructure winners will control inputs unlikely to be commoditised including power procurement and delivery; permits and land; grid interconnections; and adjacency to networks.

Recommended

Small edges are material at scale: Even a tiny difference in power pricing creates a multimillion-dollar annual impact. A move of just 1 cent per kWh on a hyperscaler using 50MW of annual power for a facility is roughly $4.4mn a year. Across the 200GW of AI-driven extra power capacity that Bain has forecast will be needed globally by 2030, that same penny swing equates to nearly $18bn a year in costs. That’s why power cost optimisation and long-term energy contracts are critical for data centre economics.

AI models and applications will keep advancing. We are now in a phase for AI of embedding machine learning into everyday workflows, scientific discovery, logistics, finance, and more — in fact, into every facet of our professional and personal lives. The systems able to power such developments at scale will have real value.

In the meantime, investors and policymakers should look past the hype and focus on the questions that matter — can power supply scale to demand? Where will permits, grid capacity and land actually unlock gigawatts? Who bears the capital cost of ensuring resiliency of systems as AI workloads surge? Data centre headlines or “bragawatts” aren’t the point; delivery is. Not all picks-and-shovels strategies will be equally effective.

Adam Selipsky, the former CEO of Amazon Web Services and current KKR senior adviser, also contributed to this article.