Start with the definition, because the phrase is used loosely. Workers per megawatt is the number of people a given block of capacity requires — either at the peak of construction, or on an average day, or permanently once operating. It is a conversion rate between two units that finance teams treat separately: megawatts (capacity, and therefore revenue) and headcount (a cost line, and therefore a schedule risk). Treating labour as a cost line is the mistake. In a market where 42% of contractors say worker shortages delayed projects — the most commonly cited cause of delay in the Associated General Contractors of America (AGC) and NCCER 2026 workforce survey of 1,830 firms — labour is not a cost, it is the delivery date.

What are the actual workers-per-megawatt numbers?

The most complete public dataset is Apollo Technical's tracker of large US campuses (100 MW and up), which measures peak construction headcount against capacity for projects that disclosed both. Its median across 19 campuses under 1.5 GW is 4.1 construction workers per megawatt at peak. The tracker's 54 large campuses under construction — 50.5 GW — imply roughly 134,800 workers at peak (range 96,300–157,700), and the figure is lower than 4.1 × 50,500 MW because gigascale campuses are built in phases, where the median falls to 1.9 per MW.

Two developer disclosures check that ratio from the outside, which is the only kind of cross-check that matters here. Meta's 1 GW El Paso campus was announced at 4,000+ workers at peak (Meta, March 2026) — about 4.0 per MW, and independently tracked at 2,300+ on site in July 2026 as the build ramped. Oracle's fact sheet for Stargate Abilene (1.2 GW) puts 6,400+ workers on site daily at the busiest point — about 5.3 per MW. So the planning range for a US AI campus is 4 to 5.3 peak workers per megawatt, with phased megacampuses closer to 2.

SourceScopeWorkers per MWWhat it measures
Apollo Technical tracker19 US campuses under 1.5 GW4.1Peak, median of projects publishing both figures
Meta, El Paso TX1 GW campus~4.0 (4,000+ peak)Peak, developer announcement (Mar 2026)
Oracle fact sheet, Stargate Abilene TX1.2 GW campus5.3 (6,400+ daily)Peak, busiest point (Sep 2025)
Apollo tracker, phased megacampuses8 projects1.9Peak, multi-year phased build
Hamm Institute (Oklahoma State), Nov 2025National forecast0.7–2.0Average day, not peak — different measurement

Why do the published ratios differ by up to 6×?

They are not disagreeing; they are answering different questions. The Hamm Institute for American Energy's November 2025 employment forecast puts data center construction requirements at roughly 0.7 to 2.0 workers per MW, and that is the number a regional economist should use for a multi-year labour-demand forecast, because it averages across the whole build. Peak headcount is a different quantity — the crew on site in month nine of an eighteen-month fit-out — and it is the number a general contractor staffs to, because that is when the electrical contractor needs 40 more journeymen than the market has. Confusing the two is how a county approves a project expecting 400 workers and then meets 4,000.

The electrical share is why the peak is so sharp. Electrical work is 45%–70% of a data center's construction budget (IBEW, cited by multiple contractors), with a competing estimate of 40%–50% from Encor Advisors' cost guide — the ranges differ by scope definition, and we are showing both rather than picking. Either way the building is an electrical project with a shell around it, and the peak crew is not a smooth blend of trades: it is a spike in one trade, in one phase, in one metro, all at once. That is the structural reason labour behaves like a queue and not like a market — you cannot spread a commissioning crew across four years the way you can spread a transformer order.

The 12:1 cliff: what happens when the building opens

Operating ratios are the mirror image, and every independent source lands in the same narrow band. Apollo's median is 0.35 permanent jobs per MW (25 projects). PwC's employment analysis puts long-term operations staffing at 50 to 300 permanent jobs for a hyperscale build. Virginia's JLARC study of a typical 250,000 sq ft facility found about 50 permanent jobs, roughly half direct employees and half contractors. Read as an intensity, those figures cluster at roughly 0.25–0.5 permanent jobs per megawatt — 25 to 50 staff for a 100 MW campus (our arithmetic on the cited numbers). And Meta's El Paso disclosure is 300+ permanent jobs at 1 GW — 0.3 per MW, with the city's incentive agreement requiring only 50.

So the cliff is real and roughly constant: 4.1 at peak, 0.35 forever, about 12:1. A 1 GW campus is a 4,000-person construction project that becomes a 300-person operating site. We are not making a political argument about whether that is a good deal for a county — that debate is well covered elsewhere and depends on tax abatements we are not auditing. We are pointing at the operational consequence: the labour constraint exists almost entirely in the build phase, which means it bites hardest exactly when a developer is trying to convert a signed lease into revenue, and it disappears precisely when the asset starts paying. Every hour of schedule risk sits in the window where nothing is being billed.

What does one month of delay actually cost?

STL Partners modelled a 60 MW single-tenant US facility and priced the slip: about $14.2M per month, decomposing into $10.8M of lost revenue, $2.2M of labour cost overrun (300 contractors plus 60 technicians carried longer than planned) and $1.2M in SLA penalties. The return impact is the part that changes decisions: project IRR falls from 17.1% delivered on time to 15.5% at one month, 12.6% at three and 8.8% at six. Two independent write-ups of that model agree on the figures, and the internal arithmetic checks: 60,000 kW × $180/kW/month = $10.8M, which is the lost-rent component almost to the dollar.

That coincidence is the useful insight, so let us state it as arithmetic rather than metaphor. $14.2M ÷ 60,000 kW = $236.67 per kW-month. The delay penalty and the rent for the same capacity are the same order of magnitude, because the dominant term in the penalty is the rent you did not collect. One month late ≈ one month of rent, plus a 31% surcharge for the labour and contractual tail ($3.4M on $10.8M).

ComponentPer month (60 MW)Per kW-monthNote
Lost rent$10.8M$180At the model's assumed $180/kW/month, 100% utilised
Labour overrun$2.2M$37300 contractors + 60 technicians carried longer
SLA penalties$1.2M$20Enterprise tenant, $100k/week class
Total$14.2M~$237≈ one month of rent + 31% (our arithmetic)

Then re-price it in the markets that actually set the ceiling. CBRE's Q1 2026 global data puts asking rents for 250+ kW, N+1/Tier III requirements at $235–265 per kW per month in Frankfurt and $330–475 in Singapore. Run the same 60 MW through those rates and a month of slip costs $14.1M–$15.9M in Frankfurt and $19.8M–$28.5M in Singapore before labour overruns and penalties — because the penalty is a rent-shaped number and rent is not the same everywhere. The corollary is uncomfortable for anyone modelling delay as a fixed percentage of capex: it is not a capex percentage, it is a rent percentage, and it scales with the market you built in.

Where does the queue actually break?

In one trade. Contractors surveyed by AGC and NCCER in mid-2026 named electricians as the hardest occupation to hire, followed by HVAC technicians, pipefitters and welders; among the 28% of respondents who had worked on a data center project in the past year, 58% said those projects increased competition for skilled workers, 49% reported wage pressure, and 37% named worker or subcontractor availability as their single biggest challenge. Across all respondents, 87% had openings for hourly craft positions, and of those, 88% said the jobs were as hard or harder to fill than a year earlier. Half said candidates lacked the skills, certificates or licences required — even though more than eight in ten firms had raised base pay.

The supply side is not elastic on any timeline that matters to a project. The US has 757,220 electricians at a median $63,190 a year (BLS, May 2025), with 72,700 openings a year projected through 2035 — a replacement rate of roughly 9.6% of the entire stock, most of it retirements, against 9% projected employment growth. The Associated Builders and Contractors puts the industry's need at 349,000 net new construction workers in 2026 and 456,000 in 2027, and notes about one in five electricians is over 55; NCCER's long-standing estimate is that 41% of the current construction workforce retires by 2031. Electrical apprenticeships run four to five years. You cannot compress that into a project schedule, which is why the shortage shows up as price and delay rather than as an unfilled job posting.

The operating side tells the same story from the other direction. Uptime Institute's 2025 Global Data Center Survey found 46% of operators have difficulty finding qualified candidates, 37% struggle to retain staff, and 35% say staff are being hired away by competitors — four independent write-ups of that survey agree on all three figures. Retention is now as binding as hiring: every campus that opens takes technicians out of the same national pool, so the build boom of 2026 is quietly consuming the operations workforce of 2028.

Is the shortage concentrated in specific markets?

Yes, and the concentration is the useful planning unit. Apollo's metro pressure index compares capacity under construction against the local pool of electricians, HVAC mechanics, construction managers and electrical technicians: Columbus, Ohio's 2,214 MW under construction would need about 6,160 workers at peak — 41% of the metro's entire combined trades pool — and Atlanta sits at the same 41%. Texas carries the most large-campus capacity, 14.2 GW across 17 campuses, much of it in rural counties far from metro labour pools, which is why those projects run on travelling crews, per diems and 75-mile commutes. The index is explicitly a stress test rather than a forecast — it asks what happens if every project peaks at once — but the ranking is what a site-selection decision should actually read, because it prices the labour market the way the interconnection queue prices the grid.

Why can't higher wages fix this?

Because wage competition moves workers between projects; it does not create them. Macrina Wilkins of AGC puts it directly: "You can poach and offer competitive wages, but if you literally do not have the number of skilled individuals needed to get a project done, you are just going to have to wait." The survey data matches — half of firms report candidates who lack the required skills even after pay increases, and 42% report that labour shortages delayed projects, the most-cited delay cause in the survey. McKinsey's oft-recited framing of the AI-infrastructure ask — roughly 130,000 additional electricians between 2023 and 2030 — is a training-pipeline problem, and training pipelines are measured in apprenticeship cohorts, not quarters. Poaching also has a second-order cost that shows up in the delay math: it raises turnover at the site you poached from, which pushes their commissioning date, which pushes the shared substation, which pushes you.

What does this mean if you are buying compute?

Treat announced megawatts as a hiring curve, not a delivery date. A neocloud's 500 MW pipeline is 2,000-plus peak construction workers, 175 permanent operators, and a 12–18 month electrical fit-out in a metro whose trades pool may already be 40% committed. The preleasing number tells you the market already knows this: 80.4% of capacity under construction across US primary markets was pre-leased in H1 2026, up from 74.3% a year earlier, with less than 1,500 MW of available future capacity left across all primary markets — about six months of demand at the current absorption rate (CBRE, H1 2026). When almost everything under construction is already spoken for, the thing you are actually buying is somebody's schedule, and the schedule is a labour plan.

Price the slip, not just the rate. If a month of delay is worth $237/kW, then a contract that promises delivery in Q3 and delivers in Q4 has given you a 3–4% effective discount on the whole deal — worse than most of the rate negotiation you just spent three weeks on — and it has done so in a market where rent is rising 8.3% per half-year for 3–10 MW requirements (CBRE H1 2026), i.e. the delay also moves the benchmark you will re-price against. Contracts that carry liquidated damages for late delivery are not hostile; they are the only clause that prices the queue you are actually standing in.

Ask who is holding the crew. The question that separates a real pipeline from a slide is not "how many MW do you have" but "which electrical contractor, which metro, and how many journeymen are committed in which month." The same discipline this forum applies to utilisation and to reserved-versus-spot spreads applies one layer up: the price of capacity is set by the scarcest input, and at 4.1 workers per megawatt the scarcest input is a person with a licence and a pair of boots.