How Many People Work at an AI Data Center? 2026 Surprising Data
An AI data center runs on fewer people per megawatt than a conventional one, not more. Meta’s 5 gigawatt Hyperion campus in Louisiana is staffed by 1,000 permanent employees, about 0.2 per megawatt. AI sites employ more people per building only because each building holds ten times the power.
That single sentence resolves the argument you will find everywhere else on this topic.
One camp says AI data centers barely employ anyone. The other says AI is driving technician demand through the roof. Both are right, and they are measuring different things.
This guide gives the real headcount by facility size, separates AI training campuses from AI inference sites, and explains why announced job numbers almost never match operations headcount.
Last updated: September 3, 2026. Staffing figures come from the Zage Business of Energy Initiative at the University of Southern California (December 2025), the Uptime Institute 16th Annual Global Data Center Survey (July 2026), Kelly’s 2026 Data Center Salary Guide (August 2026), Build’s training and inference design analysis (June 2026), and publicly announced project figures current to August 2026. Every figure below carries its source and date.
The short answer: dozens to a few hundred people
A typical AI data center employs between 15 and 300 permanent on-site workers once it is operational.
Only the multi-gigawatt megacampuses reach four figures, and just one of them has published a number.
The exact count depends on three things: the size of the facility in megawatts, how automated it is, and whether it runs AI training hardware, AI inference hardware, or ordinary cloud servers.
Below roughly 20 MW the number is not set by megawatts at all. It is set by shift coverage, and it bottoms out around 8 to 12 people.
Brookings research published in 2026 put the broader point plainly: large data center projects often promise only dozens to a few hundred permanent workers, and the associated construction jobs are temporary.
AI data center staffing in six numbers
DCGeeks compiled these figures from company announcements, a state tax-abatement filing, and two primary research sources published between December 2025 and August 2026.
- 1,000 permanent roles, staffing for Meta’s 5 gigawatt Hyperion campus in Richland Parish, Louisiana, against more than 7,500 workers at peak construction (Meta, July 13, 2026)
- 375 permanent employees, already hired at Microsoft’s Fairwater AI campus in Mount Pleasant, Wisconsin, rising to about 800 when the second building opens (Microsoft, reported May 12, 2026)
- 800 permanent jobs, against 2,400 construction jobs, across Anthropic’s $50 billion Texas and New York build (Anthropic, November 12, 2025)
- 73 permanent jobs, over ten years at Vantage’s $3 billion AI campus outside Reno, Nevada, against more than 4,000 construction jobs (Vantage application to the Nevada Governor’s Office of Economic Development, March 21, 2024)
- 0.15 to 0.20 FTE per megawatt, permanent operations staffing at automated hyperscale facilities, against 0.30 to 0.35 at labour-intensive enterprise sites (Zage Business of Energy Initiative, University of Southern California, December 3, 2025)
- 100 to 200 kW per rack, AI training rack density, against an industry modal density of 11 kW (Build, June 22, 2026; Uptime Institute 16th Annual Global Data Center Survey, July 2026)
Read the last two together. They are the whole argument.
Why AI data centers employ fewer people per megawatt, and more per building
Data centers are among the least labor-intensive buildings in the economy. The servers do the work, and software watches the servers.
Modern facilities run on data center infrastructure management (DCIM) platforms that monitor power, cooling, and hardware health around the clock, so a small team can oversee tens of thousands of machines. Much of the monitoring happens remotely from a network operations center covering several sites at once.
A single hyperscale campus built by Amazon, Google, or Microsoft can become one of the largest taxpayers in its county while employing fewer people than a mid-size grocery store.
Now add AI to that picture, and here is where almost every article gets confused.
An AI training rack draws 100 to 200 kW, and frontier systems are approaching 1 MW per rack, according to Build’s June 2026 analysis of training and inference design. The Uptime Institute’s July 2026 survey put the industry’s modal rack density at 11 kW.
One AI hall therefore carries the electrical load of roughly ten conventional halls in the same footprint.
That is why both stories are true at once. Headcount per building goes up, because the building is doing ten times as much. Headcount per megawatt goes down, because automation and standardised design scale faster than payroll.
The money buys chips, power, and cooling, not payroll.
AI data center staffing by facility size
The clearest way to estimate headcount is to start with the facility’s power capacity in megawatts (MW), then check it against the shift-coverage floor.
Facility | Typical MW | Permanent operations FTE | FTE per MW | What sets the number |
|---|---|---|---|---|
Edge / micro | 1 to 5 MW | 5 to 15 | Not meaningful, the floor governs | Minimum staffing to keep a site attended at all |
Enterprise or colocation, single building | 10 to 20 MW | 10 to 20 | Not meaningful, the floor governs | Shift coverage, not megawatts |
Conventional hyperscale hall | 50 to 100 MW | 20 to 50 | 0.25 to 0.50 | Megawatts, with automation pulling the ratio down |
AI inference facility | 20 to 100 MW | 8 to 28 | 0.24 to 0.28 | Customer-facing operations and network work on top of facilities |
AI training campus | 500 to 1,000 MW | 75 to 200 | 0.15 to 0.20 | Hardware replacement volume and liquid cooling maintenance |
Multi-gigawatt AI megacampus | 5,000 MW (Meta Hyperion) | 1,000 | ~0.20 | The only US megacampus with a published headcount |
How these numbers were built. Per-megawatt intensities come from the Zage Business of Energy Initiative at the University of Southern California, whose December 2025 workforce forecast publishes 0.15 to 0.20 operations FTE per MW for automated hyperscale sites, 0.24 to 0.28 for standard regional facilities, and 0.30 to 0.35 for labour-intensive enterprise sites. Facility-size bands for AI training and inference come from Build’s June 22, 2026 analysis. The megacampus row is anchored to a single real site.
Counts are direct operations staff. Adding on-site contract security, cleaning, and trades roughly doubles total site headcount, because both JLARC in Virginia and Sage Policy Group in Maryland found about half of a data center’s on-site workforce is contract rather than employed.
Two honest gaps. No published survey breaks operations headcount out by AI training versus AI inference. The rows above apply per-megawatt intensities to the facility sizes each workload actually uses, and we will replace them the moment somebody publishes a real count.
Below roughly 20 MW, headcount is set by shift coverage rather than megawatts. Four shifts each need two or three critical facility technicians plus a supervisor, and that floor of 8 to 12 people holds whether the building draws 5 MW or 20 MW. Per-megawatt math only takes over above that line, which is why the top two rows show no ratio.
Conventional vs AI training vs AI inference data centers
Most coverage treats “AI data center” as one kind of building. It is two, and they staff differently.
Training campuses are where models get built. Inference facilities are where finished models answer requests. They are sited for different reasons, cooled differently, and hire for different roles.
Conventional data center | AI training facility | AI inference facility | |
|---|---|---|---|
Typical facility size | 5 to 50 MW per building | 500 to 1,000 MW campuses | 20 to 100 MW |
Rack density | 11 kW modal, 7.8 kW average (Uptime, July 2026) | 100 to 200 kW, frontier racks approaching 1 MW | 30 to 150 kW mainstream, up to 370 kW at the leading edge |
Permanent ops FTE per MW | 0.24 to 0.35 | 0.15 to 0.20 | 0.24 to 0.28 |
Permanent headcount, real site | About 50 FTE per 250,000 sq ft, Virginia and Maryland averages | 1,000 at Meta Hyperion, 5 GW | 375 hired at Microsoft Fairwater, about 800 planned |
Cooling | Air: CRAC, CRAH, chilled water | Direct-to-chip liquid, mandatory at this density | Hybrid: liquid on the accelerators, air everywhere else |
What drives headcount up | Shift coverage | Hardware replacement volume, coolant loop maintenance, power quality management | Network and customer-facing operations, multi-tenant change control |
Where it gets built | Near the enterprise it serves | Power first: rural Texas, Louisiana, the Midwest, the Nordics | Latency first: inside metros, near end users |
Hardest role to fill | Electrician | Liquid cooling commissioning specialist, 92% of employers report high demand, and GPU cluster operations technician at 84% (Kelly, August 2026) | Network engineer, plus the same medium-voltage electrical gap at 88% (Kelly, August 2026) |
Career read | Stable, local, long tenure | Highest pay, most remote locations, newest skills | Metro-based, closest to a conventional IT career path |
Training campuses pay the most and get built where the power is, which usually means moving.
Inference facilities are going up inside metros, they staff more densely per megawatt than training campuses do, and they are the version of this job you can take without relocating.
If you are choosing where to aim, the inference build-out is the one that will put a site within commuting distance of a city you already live in. The training build-out is the one that will pay you a premium to go somewhere you do not. Our guide to AI data center careers breaks down the roles and pay bands on both sides.
Construction vs operations: two very different numbers
The single biggest source of confusion is the difference between building an AI data center and running one.
Construction is a temporary surge of thousands of skilled tradespeople. Operations is a permanent crew of a few dozen to a few hundred.
Meta’s Richland Parish campus in Louisiana shows the gap in stark terms. In a July 13, 2026 post, Meta put the site at more than 7,500 jobs at peak construction and 1,000 roles once operational, across 5 GW of compute capacity and more than $50 billion of investment.
Anthropic’s November 2025 announcement of its $50 billion Texas and New York build follows the same shape at smaller scale: 2,400 construction jobs against approximately 800 permanent ones.
For scale on the construction side, Virginia’s Joint Legislative Audit and Review Commission (JLARC) found in December 2024 that a single data center building puts about 1,500 workers on site at peak.
Those construction jobs are real, well paid, and last one to three years per project, but they pack up and move to the next site when the building is done. Between 75 and 98 percent of site employment ends at commissioning, depending on the campus.
Real AI data center examples
Announced headcount for named AI campuses makes the scale concrete. Every figure below is traced to the document it came from.
AI data center | Capacity | Peak construction | Permanent jobs | Source |
|---|---|---|---|---|
Meta Hyperion (Richland Parish, LA) | 5 GW | 7,500+ | 1,000 | Meta, July 13, 2026 |
Microsoft Fairwater (Mount Pleasant, WI) | Two buildings, $7.3bn+ | Not disclosed | 375 hired, ~800 planned | Microsoft, May 12, 2026 |
Anthropic and Fluidstack (TX and NY) | $50bn programme | 2,400 | ~800 | Anthropic, November 12, 2025 |
OpenAI Stargate (Abilene, TX) | 1.2 GW, eight buildings | 6,400+ daily | “Nearly 1,700 jobs onsite”, see note below | Oracle and OpenAI fact sheet, September 23, 2025 |
Vantage (near Reno, NV) | 1.1M sq ft, $3bn | 4,000+ | 73 over ten years | Vantage filing to Nevada GOED, March 21, 2024 |
Meta (Lebanon, IN) | $10bn+ project | 4,000+ | ~300 | Quartz, May 15, 2026 |
Why announced job numbers do not match operations headcount
Look closely at the Abilene row, because it demonstrates a trap that runs through every economic impact announcement in this industry.
Oracle’s fact sheet says the Abilene campus will carry “nearly 1,700 jobs onsite.” Meta says Hyperion will carry 1,000 roles once operational. Abilene is 1.2 gigawatts and Hyperion is 5 gigawatts.
Those two numbers cannot both be permanent operations headcount. If they were, Abilene would be staffing about seven times as densely as Meta’s largest site.
The difference is definitional. A phrase like “jobs onsite” in a developer fact sheet or a tax-abatement filing usually blends operations staff, on-site contractors, and in some cases the construction crew. A phrase like “roles once operational” usually means the operator’s own permanent payroll.
When you read a data center jobs announcement, the question to ask is not how many. It is how many, of what kind, and for how long.
Vantage’s Reno campus is the cleanest example of why this matters, because the figure comes from a filing rather than a press release. Business records and Vantage’s own March 2024 application to the Nevada Governor’s Office of Economic Development, reported by The Nevada Independent in January 2026, put the site at 73 permanent jobs over the next decade against more than 4,000 temporary construction jobs near Reno.
A Sage Policy Group study for the Maryland Tech Council in 2025 found that a typical 250,000-square-foot data center has about 50 full-time workers, roughly half of them contractors. JLARC reached the same split independently in Virginia, which is what makes it a reliable rule of thumb rather than one consultant’s estimate.
Why AI facilities need more hands-on staff per rack
Per megawatt, AI sites staff leaner. Per rack, they staff heavier. Three mechanisms explain the second half of that sentence.
Hardware fails constantly at AI density
Meta’s own published account of training Llama 3 recorded 419 unexpected interruptions across 54 days on a cluster of 16,384 NVIDIA H100 GPUs, roughly one every three hours.
Hardware caused 251 of them. 148 traced to GPU faults and 72 to HBM3 high-bandwidth memory, per The Llama 3 Herd of Models, published July 29, 2024. Every faulted GPU eventually becomes a physical swap that a person performs.
The caveat matters as much as the number. Meta also reported that significant manual intervention was needed only three times in those 54 days, because automation handled the rest, and the run still achieved more than 90 percent effective training time.
That is the honest shape of it. The failure volume is enormous, the response is mostly automated, and the residual human work is hardware replacement rather than firefighting. Anyone quoting the 419 figure as proof that AI data centers are chaos is quoting half a finding.
Liquid cooling adds a maintenance system air-cooled halls never had
At 100 kW per rack and above, direct-to-chip liquid cooling stops being optional.
That introduces coolant distribution units (CDUs), manifolds, quick-disconnect fittings, leak detection, and fluid chemistry monitoring, all of which need a mechanical technician who understands them.
It is why liquid cooling commissioning is the single scarcest skill set in the sector, with 92 percent of employers reporting high demand and short supply in Kelly’s 2026 Data Center Salary Guide, published August 25, 2026. If you want the roles this creates, start with liquid cooling data center careers.
AI loads swing, and the electrical room absorbs it
Training clusters produce load fluctuations of 30 to 60 percent within milliseconds, which calls for harmonic filtering and fast-response UPS infrastructure that inference and conventional facilities do not carry, according to Build’s June 2026 analysis.
More equipment in the electrical room means more preventive maintenance and more people qualified to do it. That pressure is already visible in the market: the Uptime Institute’s 16th Annual Global Data Center Survey, fielded across more than 1,600 participants in April and May 2026, found 53 percent of operators reporting difficulty finding qualified candidates, up from 46 percent a year earlier.
For the full picture on that gap, and which electrical and mechanical roles are hardest to fill, see our breakdown of the data center workforce shortage.
One framing note worth keeping straight. Uptime Institute forecast back in 2021 that global data center staffing demand would rise from about 2.0 million full-time roles in 2019 to 2.3 million by 2025. That forecast window has now closed, and it was a projection rather than a measurement, so treat it as history rather than a current figure.
The jobs actually on-site
The permanent crew at an AI data center is small but highly skilled, and it covers a predictable set of roles for 24/7 coverage. Most facilities run three or four rotating shifts to keep people on-site at all hours.
Typical on-site positions include:
- Critical facility engineers who manage power and cooling systems
- Critical facility technicians (CFTs) who monitor HVAC, power distribution, and alarms
- Electrical and mechanical maintenance technicians, including coolant loop specialists at liquid-cooled sites
- DCIM and controls specialists who run the monitoring software
- Server, network, and GPU hardware technicians who swap failed components
- Physical security officers and facility management staff
Meta’s own description of its Richland Parish roles lists electricians, HVAC workers, engineers, server and network technicians, safety workers, security personnel, and even culinary staff.
Pay is strong for the people who hold these jobs. There are simply not many of them per building.
Frequently asked questions
How many people do AI data centers employ?
AI data centers employ anywhere from a few dozen to about 1,000 permanent staff. Microsoft’s Fairwater AI campus in Wisconsin has hired 375 and expects about 800. Meta’s 5 gigawatt Hyperion campus in Louisiana will carry 1,000 roles, the largest figure any US operator has published.
How many people does an AI data center employ?
A single AI facility employs roughly 0.15 to 0.28 permanent staff per megawatt, so a 100 MW site runs 25 to 28 people and a 500 MW training campus runs 75 to 100. Construction employs 3 to 50 times that number, temporarily, depending on the site.
How many people work in an AI data center?
On any given shift, three to eight people work inside an AI data center: critical facility technicians, a network operations centre operator, hardware technicians swapping failed GPUs, and a security officer. Total site headcount across four shifts plus engineering runs 8 to 200 for most facilities.
How many people work at AI data centers?
Across named US AI sites the range runs from 73 permanent jobs at Vantage’s $3 billion Reno campus to 1,000 at Meta’s Hyperion in Louisiana. The middle of the range is a few hundred. Anthropic’s $50 billion Texas and New York build projects 800 permanent jobs.
How many employees does an AI data center have?
Employee count depends on who you count. A 250,000 square foot facility averages about 50 on-site workers, but roughly half are contract electricians, security officers, and cleaners rather than the operator’s employees. Announced “jobs onsite” figures in developer fact sheets usually blend both.
Do AI data centers employ a lot of people?
No, not relative to their cost. Meta’s Hyperion campus carries more than $50 billion of committed investment against 1,000 permanent jobs, about $50 million per job. Vantage’s Reno campus is $3 billion against 73 permanent roles. Construction is where the volume of employment sits.
How many jobs do AI data centers create?
An AI data center creates far more construction jobs than permanent ones. Meta’s Hyperion campus peaks at over 7,500 construction workers against 1,000 permanent roles. Anthropic’s build projects 2,400 construction jobs against 800 permanent. Between 75 and 98 percent of site employment ends at commissioning.
How many people does it take to run an AI data center?
Running an AI facility around the clock takes a minimum of 8 to 12 people, because four shifts each need two or three critical facility technicians plus supervision. AI sites add hardware technicians for GPU replacement and a mechanical technician for the coolant loops.
Conclusion: what the numbers actually mean
The straight answer to how many people work at an AI data center is dozens to a few hundred per facility, with only the rare multi-gigawatt campus reaching four figures.
That number stays small no matter how many billions the project costs, because AI data centers buy compute, not headcount.
What AI changes is not the per-megawatt ratio. It is the density behind it. Ten times the power per rack means ten times the hardware to swap, a coolant system that did not exist on the old sites, and an electrical room that has to absorb load swings of 30 to 60 percent in milliseconds.
The real employment story sits in two places: the temporary construction boom that can put thousands of tradespeople to work for a year or three, and the rising industry-wide demand for skilled electrical, mechanical, and cooling technicians who keep dense AI hardware alive.
If you want to work in this field, target those durable skilled-trade and critical-facilities roles, because that is where AI’s power density is creating steady, hard-to-automate demand. Aim at inference sites if you want to stay in a metro, and at training campuses if you want the premium.
For the full picture on total job creation across construction and operations, see our breakdown of how many jobs does a data center create.
