McKinsey's 6.7 trillion dollar number is 219 gigawatts of data-center capacity by 2030. Most of it does not exist yet

$6.7 trillion is the number that travels. The number that does the work is 219.

McKinsey's 2026 report projects global data-center capacity reaching 219 gigawatts by 2030, split as 156 GW serving AI workloads and 64 GW serving everything else. Getting there requires a 19 to 22 percent compound annual growth rate sustained from 2023 through 2030, and it carries a bill: $5.2 trillion of capital for AI-capable data centers plus $1.5 trillion for traditional IT infrastructure, a cumulative $6.7 trillion of capex by the end of the decade. The trillions get quoted at conferences. The gigawatts are where a developer should point the flashlight, because a dollar figure is an abstraction and a gigawatt is a place.

Most of those 219 gigawatts do not exist yet. That is the entire post, worked backward.

Divide the target by the calendar

McKinsey's incremental AI-driven capacity between 2025 and 2030 is 125 GW on its own. Spread across the five years in that window, the AI slice alone averages 25 GW of new capacity every year through 2030.

Now put the observed pace next to the required one. SemiAnalysis counts 20 GW of total new data-center capacity coming online in 2026, with roughly 30 GW expected in 2027. Those are the two biggest delivery years this industry has ever recorded, and they bracket the 25 GW annual average McKinsey's AI number needs. Which means the projection assumes the buildout runs at approximately the best pace ever achieved, every single year, for the rest of the decade, with no off years, and that is before touching the non-AI capacity in the 219.

Targets do not pour concrete.

Say it less politely: there is no slack in this schedule. A 19 to 22 percent CAGR sustained for seven years is not a trend line, it is a promise that nothing systemic goes wrong between now and 2030. No financing winter and no moratorium wave spreading county to county through the exact jurisdictions where land and power happen to coexist. It also quietly assumes interconnection queues in every major market keep clearing at record volume. Each of those failure modes has already appeared somewhere at small scale. The projection requires that none of them appear at large scale for four more years.

Two serious firms are 450 terawatt-hours apart, and that gap is information

McKinsey's companion power figure is data-center demand reaching 1,400 TWh by 2030. The IEA's 2026 Electricity report puts the 2030 number at 950 TWh. Different scopes, different methodologies, two institutions that do not publish casually, and a 450 TWh spread between them.

I am not going to pretend to reconcile those numbers, and you should distrust anyone who does. The honest reading is that the spread itself is the finding. When two careful forecasters differ by nearly the size of the IEA's entire 2025 baseline, the error bars around this buildout are wide enough to drive a substation through, and any pro forma that plugs in either figure as settled fact is reporting false precision to its investment committee.

The spread has a practical use, though. Underwrite against the band, not the point. A deal that only works in McKinsey's 1,400 TWh world is a bet that the most aggressive published number survives four more years of county votes intact. A deal that still works in the IEA's 950 TWh world has priced its forecast risk instead of wishing it away. If a deck does not say which world it lives in, assume it quietly picked the bigger one.

What an entitlement-aware developer does with a number like 219

Stress-test the calendar, not the capital.

The capital is the input this projection has never lacked; $6.7 trillion is being assembled in public, quarter by quarter, in hyperscaler guidance. The calendar is the input nobody controls. The capacity built so far was, mostly, the easy capacity: entitled land in counties that still said yes, next to substations with headroom. Every incremental gigawatt from here is a site, a substation, and a county hearing, and the marginal one is harder than the last because the friendly jurisdictions fill up first. A model that assumes 25 GW a year of AI capacity through 2030 is implicitly assuming entitlement timelines hold flat while volume runs at record pace through progressively more contested ground. My working assumption is the opposite: timelines stretch as volume grows, because hearing rooms are a fixed resource and opposition organizes faster at scale.

So when the 219 GW figure shows up in a deck, the underwriting questions are concrete. Which years of the ramp does this deal depend on. What happens to the exit if the industry delivers 20 GW in a year the model needed 25. And if an energization date slips a year, the model should already say who eats the carry. A projection this large is not wrong because it is large; it is untested because almost none of it has happened yet.

My bet, dated so it can be scored: 2030 arrives with global capacity under 219 GW, and the post-mortems trace the shortfall to power and permission rather than to capital or demand. If the industry actually hits the number, it will have sustained a construction pace it has never once managed for even three consecutive years, and I will happily reread this post in public. Either way, the developers who underwrote the delay instead of the target will own the decade, because they priced the one thing McKinsey's model holds constant and a county commission does not.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.