Power, not money, now blocks data centers

By September 2026, the public language from operators and investors had converged on a blunt conclusion: the scarce asset was not financing but deliverable power. On The Data Center Frontier Show, executives said “power is the first part of diligence” and “No power, no customer,” while also arguing that “powered land” had effectively become its own investment class because customers wanted sites with secured interconnection; as one speaker put it, “We remember the days when we were like, oh, wow, you know, 10 megawatt data center… and now the numbers in terms of power are just massive.”

That framing hardened into a measurable constraint by late summer: “One thing decides how fast US data centers grow in 2026,” and it was “the ability to get electric power delivered to the site not money, land, chips, or demand.” The evidence pointed specifically to guaranteed delivery as the bottleneck, with Goldman Sachs Research estimating only 60% of next year's scheduled capacity would arrive on time, dropping to 50% over the following two years; Bloomberg, citing Sightline Climate, said 30–50% of roughly 16 GW planned for 2026 could be delayed or canceled, while large transformers averaged 128-week lead times and generator step-up units 144 weeks.

AI demand has changed data-center siting from a market-sizing exercise into a constraint-matching race for parcels that can actually be delivered. On The Data Center Frontier Show, Colby Cox said the question is no longer where demand can support buildouts but “who can secure the power, the land, the interconnection cues” quickly enough, because projects have moved beyond “incremental expansion” into “significant chunks of megawatts if not get into the gigawatts,” with “multi gigawatt AI campus planning” stretching from roughly 100 upward and forcing developers to align power, land, cooling and timing at once.

That new logic is visible in where projects are going: not necessarily near users, but where energy and land can be assembled on realistic timelines. AI server racks now draw 100-120 kW, up from 10 kW five years ago, while rack densities are surging from 25-30 kW to 600 kW per rack; with AI power demand in the US potentially growing more than thirtyfold by 2035 to 123 GW, grid stress has become a major obstacle, and with grid connections taking 3-7 years against 18-24 month builds, developers are pushed toward off-grid power, alternative fuel access and sites that can support those workarounds. Reuters, citing JLL, found Europe’s average project distance from major hubs rising to 175 kilometres from 46, greenfield share climbing, and inner-city pipeline share falling to 5% from 13%.

The clearest sign that AI data center demand has become system-scale is how fast capital and capacity are being absorbed at once. PR Newswire reported that CleanCore “Lands $800M, 10-Year Deal With Cerebras for Fully Pre-Leased Minnesota AI Data Center,” while JLL said North American absorption hit a record 25 GW in H1 2026, double a year earlier and five times two years prior, with vacancy stuck at 1% for a third straight year despite unprecedented construction and 66 GW underway, 95% of it already pre-committed.

Apollo’s Torsten Slok put the spending surge in macro terms: “data-center capex adds 1.7 percentage points in just two years, from 1.4% of GDP in 2025 to 3.1% in 2027,” after a “massive 2.5 percentage point rise from 0.6% of GDP in 2023,” with investment “building at close to twice the pace” of the mid-2000s housing boom and consensus forecasts holding hyperscaler capex near 3% of GDP through 2029. Slok also cautioned that “a cycle that builds at 0.85 percentage points a year can unwind at a similar pace.” That scale is large enough to make the constraint binding, and speculative fervor around it is visible too, with “S&P 500 call options hit a record 4M+ contracts on Cboe as traders chase upside.”