Analysis
US construction spending on AI data center "shells" -- the warehouse-scale structures later filled with server racks, chips, memory and fiber connections -- jumped to an annualized pace above $75 billion in July, rising nearly 60% from July 2025 levels, according to Census Bureau data released this week, Axios reported. The figure covers only the physical building costs, not the far larger cost of the computing hardware that eventually goes inside.
The acceleration is happening even as political backlash against data centers grows in communities where they're sited -- concerns over electricity costs, water use and local grid strain have generated organized opposition in multiple states this year, which Pulse has tracked alongside a parallel push by AI investors and the Trump administration to counter that backlash with public messaging campaigns defending the buildout.
Why this number matters for the IPO pipeline
Shell construction spending is a leading indicator for exactly the kind of infrastructure companies now filing for public listings. Nscale is pitching IPO investors on tens of billions in contracted revenue tied to data centers still under construction; SB Energy's own IPO filing discloses roughly 8 gigawatts of Ohio capacity leased primarily to OpenAI, with construction spending like the figure in this report representing the physical work that has to happen before any of that leased capacity can actually generate revenue. A $75 billion annualized construction pace is the upstream number that determines whether infrastructure IPO pipelines like Nscale's and SB Energy's contracted-revenue figures can plausibly convert into delivered, revenue-generating capacity on the timelines both companies have told investors to expect.
The gap between construction pace and revenue recognition is also where the risk concentrates: a shell under construction today doesn't generate the tenant revenue underlying an infrastructure IPO's growth story until it's actually built, powered, equipped and occupied -- a multi-year process during which financing costs accrue regardless of whether the eventual tenant relationship performs as modeled. The nearly 60% year-over-year jump in construction spending is real evidence of demand, but it's also evidence of how much capital is now locked into projects that won't generate a dollar of revenue for one to three years, a timing mismatch every infrastructure IPO investor this fall is implicitly underwriting.
Reading the number against Anthropic's own compute-buying pace
The $75 billion annualized construction pace is a useful sanity check against the demand side of the equation, too -- Pulse has tracked Anthropic alone signing roughly $90 billion in disclosed multi-year compute commitments over the past month, spread across deals with Volta, Nscale and Lambda. A single lab's compute-buying pace running close to the entire industry's quarterly shell-construction spend is either evidence that current construction capacity is genuinely undersupplied relative to demand, or evidence that compute-purchase commitments are being signed well ahead of the physical capacity that will eventually fulfill them -- and distinguishing between those two readings is exactly the kind of diligence infrastructure IPO investors need to do before pricing contracted-revenue figures at face value. The next Census Bureau release, covering August, will show whether July's jump was a one-month spike or the start of a sustained new baseline.
