Invest1 publisher3 min readPublished
Morgan Stanley puts the last groundbreaking for 2027 data-center capacity in October
Build times of two to three years make an October start the last one that can open before 2028, and JPMorgan already counts more than 60 percent of planned 2027 capacity as unbuilt. The scarce input is a permitted site with power.
The Investor · Invest desk
What happened
- Morgan Stanley is warning that any data center project meant to be operational before the end of 2027 has to begin construction by October.
- Grid interconnection queues in many U.S. regions run five to seven years, so a project with land and permits today may wait most of a decade for utility-guaranteed power at scale.
- More than 300 local moratoriums on new data center construction have been enacted across the United States since 2023.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- contradiction Bernstein's 35 to 40 percent at risk and JPMorgan's 67 percent unstarted or late cannot both set the discount on announced 2027 capacity, and whichever an investor picks changes the write-down by roughly half.
- constraint A five-to-seven-year queue forecloses any 2027 opening that depends on new grid capacity, leaving on-site generation or an already-permitted site as the only routes to that date.
- decision Capital committed to a project that cannot break ground by October is buying a 2028 or 2029 asset, so the choice now is paying up for permitted land with power or accepting the later start.
- capability The shortfall hands pricing power to developers who can serve sites that have already cleared permitting, because the buyer's alternative is a place in the queue.
Two years from an October groundbreaking finishes about two months inside the end-2027 target, and three years finishes about ten months outside it [2]. Morgan Stanley's own range is two to three years at minimum [2], so the October date [1] works only for projects at the fast end of it. A three-year build starting in October is a 2028 asset.
JPMorgan's figure is the concrete one. More than 60 percent of planned 2027 capacity has not started construction, and another 7 percent is already delayed [5]: 67 percent unstarted or late, about a third started and on time [1]. Bernstein's 35 to 40 percent at meaningful risk of delay or cancellation through 2027 is a smaller share of a bigger pool, global rather than domestic [4]. Bernstein is estimating risk; JPMorgan is counting starts. All of these numbers reach the reader through one account of bank research that did not name the analysts behind them [12].
Power sets the date. Interconnection queues run five to seven years in many U.S. regions [7], which is at least two years longer than the longest build in Morgan Stanley's range [4]. The bank puts the shortfall at 38 GW through 2028 [6]. A 2027 project that still needs a new interconnection waits the better part of a decade for firm power at the scale it needs [7].
Spending is not being paced to that. Forecast capital expenditure across Microsoft, Amazon and Alphabet is roughly $785bn in 2026 and approaching $1 trillion in 2027, an increase of about $215bn, or 27 percent [8][3]. If a third of announced 2027 capacity lands late, the money still gets committed: it moves into land, permits and power procured earlier, and into whatever premium a permitted site with grid commitments now commands. Crypto Briefing described the 38 GW gap as an addressable market for anyone able to bring reliable large-scale power to sites that already have permits [10].
The delay thesis can be wrong in three places. Leases of existing permitted capacity and behind-the-meter generation do not show up in a count of construction starts, and Morgan Stanley says hyperscalers are already moving earlier in the development cycle to secure sites and grid commitments before communities and regulators organize opposition, with the 2028 election cycle adding urgency [9]. Capex guidance could be trimmed instead; then guidance falls and the schedule holds. And the 300-plus moratoriums enacted since 2023 may sit in counties that were never carrying 2027 load [3].
Crypto Briefing framed the whole problem this way: "Money is no longer the limiting factor in the AI infrastructure race. Time is." [11] For 2027 delivery, the published figures support that reading. In my view, announced 2027 capacity is worth about a third of its headline until the construction count moves [1], and a permitted site with power is worth a premium over one with neither. Two developments would show that wrong: JPMorgan's unstarted share dropping sharply before October [5], or hyperscalers reporting 2027 capacity fed by on-site generation that never entered an interconnection queue [7].
What to watch
- Whether the moratorium tally climbs past 300 as hyperscalers buy and permit sites earlier in the cycle.
- Contracted capacity announced by utilities and power developers against the 38 GW shortfall, and the price at which it is signed.
- Any revision by Morgan Stanley to the two-to-three-year build range or to the October cutoff itself.