For two decades, software ate the world on the assumption that hardware was somebody else’s problem. Cloud providers absorbed the physical complexity, developers built on top of an abstraction layer that felt infinite, and “infrastructure” became a line item rather than a discipline. In 2026, that assumption has broken. I spend my days underwriting cloud, ML, data infrastructure, and security startups, and nearly every serious technical conversation I have now eventually turns into a conversation about copper, optics, silicon, and voltage.
The parallel to the mid-1990s is not just rhetorical. When commercial internet traffic exploded between 1994 and 1997, the primary constraint was a physical network designed for phone calls, not data. The buildout of switches, routing foundations, and fiber was ultimately an engineering response to a physics problem: too much data, not enough pipe.
AI infrastructure is running the same playbook today, just with a different bottleneck- not the wire between cities, but the wire between a chip’s memory and compute cores, and the substation feeding the building that houses both.
The Physics Budget
“Software is eating the world” has run into the local utility. Gartner now projects that roughly 40% of AI data centers will be power-constrained by 2027, not capital-constrained or chip-constrained, but physically unable to draw the electricity their racks are rated for. In the PJM Interconnection territory, the average span from interconnection application to commercial operation has stretched from under two years in 2008 to more than eight years in 2025. Globally, over 2,500 gigawatts of generation and storage capacity is sitting in interconnection queues, more capacity than is currently installed and operating in the entire U.S. grid.
That queue is why Microsoft structured a 20-year power agreement tied to restarting a nuclear unit, alongside a multi-gigawatt gas supply deal with Chevron, why Amazon has stitched together a nuclear-adjacent campus, and why the IEA flagged a wave of on-site natural gas generation being built specifically to route around slow grid connections. None of this is optional set-dressing; it’s the actual rate-limiting step on how fast a hyperscaler or a neocloud can bring capacity online.
The daily work of a serious AI infrastructure engineer has shifted accordingly- less academic model tuning, more of what the industry now calls “tokenomics”: maximizing useful output per dollar of power and per rack of heat budget. A model that can’t stay inside a building’s electrical and thermal envelope doesn’t matter how good its benchmark scores are.
The Cost of Moving Data
Compute is, for the moment, less scarce than the ability to feed it. The main constraint in AI infrastructure is memory bandwidth, making High Bandwidth Memory (HBM) the tightest link in the entire supply chain. While next-generation HBM4 helps close that gap, every new accelerator arrives hungrier than the memory supply chain can comfortably support. Long context windows and the explosion of agentic applications only further stress the KV Cache.
Every centimeter of physical distance between memory and compute costs bandwidth and burns power as heat, which is why data center thermal load has become a first-order design constraint rather than a facilities afterthought. This is precisely the class of problem systems engineers solved for telecom networks in the 1990s- minimize hops, minimize distance, keep the expensive resource (then bandwidth, now compute) from sitting idle waiting on data. It is now the single most important variable in AI infrastructure economics, and it is why some of the more interesting seed-stage companies I’m seeing aren’t building models at all- they’re building the interconnect and memory-orchestration layer underneath them.
Upgrading What We Have
The popular narrative that AI’s future belongs to whoever builds the newest greenfield campus ignores how long greenfield actually takes now. With interconnection queues running years and gas turbine lead times stretching alongside them, the more immediate opportunity is retrofitting the data center capacity that already exists. Direct-to-chip liquid cooling has gone from a novelty to something close to mandatory above roughly 100kW per rack, and the retrofit market around it has gotten real capital fast.
That’s the pattern I look for as an investor: a large, urgent, physically constrained problem that incumbents can’t solve by writing better software. Squeezing another 15% out of an operating data center this year is worth more, in practice, than a theoretical chip architecture that won’t ship for two years. Capital is rotating toward the companies solving that problem now- cooling, power routing, high-speed interconnect- over ones promising a cleverer model wrapper.
Back to the Wires
There’s a useful, if humbling, coda to the 1990s comparison: the fiber buildout that solved the bandwidth bottleneck also overshot it. By 2002, only a sliver of the fiber laid in the boom was actually lit and in use. Of course, in the fiber era, when you were digging up the streets, there was effectively 0 incremental cost to install one fiber or 100. The scenario in the GPU is quite different, where revenue and Cost of Goods Sold are directly tied to GPU spend and efficient utilization.
But the underlying shift is real regardless of how the buildout ultimately paces itself. The engineers and the investors who matter most over the next several years won’t be the ones with the cleverest model architecture; they’ll be the ones who understand power interconnection queues, memory bandwidth budgets, and thermal design as fluently as they understand transformer architectures. That’s a different hiring profile, a different diligence checklist, and, I’d argue, a different definition of what counts as “AI infrastructure” altogether. The plumbing is the story now, and it’s worth looking past the model layer and engineering through the wire.
