What does it really cost to own a data center, and how much money can one actually make?
Data centers are the hidden infrastructure behind the modern digital economy. Every cloud application, AI model, streaming platform, online service, and business system depends on massive facilities filled with servers, power systems, cooling equipment, networking infrastructure, and backup systems. But behind all that technology is a fascinating business model.
In this video, we break down the economics of owning a data center from the ground up. We explore the enormous cost of building a data center, electricity and cooling expenses, server infrastructure, power capacity, utilization rates, long-term customer contracts, financing, maintenance, reliability, expansion, and the major risks that can turn a seemingly profitable facility into a financial burden.
We also examine how the rapid growth of artificial intelligence is changing the data center industry. AI workloads require enormous amounts of computing power, electricity, and increasingly advanced cooling systems, creating a race for data center capacity and access to reliable power.
Most importantly, you'll learn where data centers actually make money, what determines their profitability, why utilization matters so much, how owners generate recurring revenue, and what separates a valuable data center from an expensive building full of unused capacity.
Whether you're interested in data centers, cloud computing, AI infrastructure, technology businesses, real estate, investing, or the economics of modern industries, this video gives you a complete look at one of the most important businesses powering the digital world.
Owning a data center is less about real estate and more about operating an industrial power conversion machine. The business centers on securing massive-scale power and cooling, where grid interconnection agreements are more valuable than the land itself. The rise of AI has further intensified the demand for power density and specialized liquid cooling, shifting the industry's core value to energy control.
Okay, so you want to own a data center.
Find some land, build a big warehouse,
fill it with servers, and let companies pay you to use them.
Money flows in. Simple.
Except it is not simple.
Not even a little bit.
And the people who walked into this business thinking it was simple
have spent the last decade either quietly exiting at a loss,
or watching their carefully constructed financial models
collapse under the weight of an electricity bill they did not see coming.
Here is what a data center actually is.
Forget the image of a room full of blinking computers.
A modern data center is, at its core, an industrial power conversion machine.
It takes electricity from the grid, converts it into computing power,
generates enormous amounts of heat as a byproduct,
removes that heat with cooling systems the size of aircraft hangars,
and then does it all over again, 24 hours a day, 365 days a year, without stopping.
The computers inside are almost secondary.
The real product is reliable, uninterrupted power delivery at a massive scale.
That distinction is not a technicality.
It is the entire business.
So before you spend a single dollar, before you hire an architect or call a real estate broker,
you need to answer one question that has nothing to do with buildings or servers.
You need to ask, can I get enough electricity?
Because without the answer being yes, nothing else matters.
Here is why that question is harder than it sounds.
When a data center needs power, it does not just plug into the wall.
It connects to the electrical grid through a substation,
which is a large piece of infrastructure that steps high-voltage electricity down to a level the building can use.
In major data center markets in the United States, like Northern Virginia,
which is the largest data center market on Earth,
and handles something like 70% of the world's internet traffic,
the wait to get a new substation connection approved and built can take 5 to 7 years.
Not 5 to 7 months. 5 to 7 years.
Because the transmission lines that carry power across the region
are already under enormous strain from all the data centers that came before you.
So the most valuable thing in the data center business right now is not a building.
It is not a server.
It is a piece of paper that says a utility company has agreed to give you megawatts of power,
and has given you a date when that power will be available.
That document is called a utility interconnection agreement.
And in some markets, it is worth more than the land it sits on.
This is why smart developers have started doing something called land banking.
Buying parcels in regions where power is available,
locking in grid connections early,
and sitting on the land for years.
A parcel might cost $10 to $15 million,
but the grid connection that comes with it,
already approved and in queue,
can make that land worth 3 to 4 times as much to a competitor who waited too long.
Now let us talk about where to build,
because location in the data center world operates nothing like location in any other real estate business.
In a normal real estate business, you want to be where your customers are.
In the data center business, you want to be where the electricity is cheap,
the fiber connections are dense, the taxes are low, and the weather is cold.
Those factors together determine whether you will ever make money.
The cheapest land in the country is not in Virginia or Texas or Arizona.
It is in rural areas with no infrastructure whatsoever.
And that cheap land will destroy your economics before you ever open the doors,
because you will spend tens of millions of dollars building your own fiber connections,
your own substations, your own utility infrastructure,
and your own roads just to make the site functional.
The states that have attracted the most data center development have done so not because of cheap land,
but because of cheap power and favorable tax policy.
Virginia eliminated the sales tax on data center equipment purchases,
which in a $100 million server purchase saves you $7 or $8 million in a single transaction.
States like Nevada and Oregon have offered similar incentives.
Meanwhile, electricity prices in the Pacific Northwest, where hydropower is abundant,
have historically run around $0.04 to $0.05 per kilowatt hour,
compared to $0.12 to $0.15 in the mid-Atlantic region.
For a facility using 50 megawatts of power,
that difference in electricity price alone is worth tens of millions of dollars every single year.
Now let us talk about what it actually costs to build one of these things.
If you are thinking of a data center as a warehouse, stop.
A data center is not a warehouse.
In a standard warehouse, the building itself, the concrete, the steel frame, the roof,
represents about 80% of what you spend.
In a data center, the building shell is roughly 15% to 20% of your total cost.
The other 80% to 85% go into electrical systems, cooling systems, backup power, and network infrastructure.
You are not really building a building.
You are building a power plant that happens to have a roof.
The numbers here are not small.
A mid-sized data center, say one that delivers 10 megawatts of power to customer equipment,
will cost somewhere between $80 million and $120 million to build.
A larger facility delivering 50 megawatts might cost $400 to $600 million.
And an AI-optimized facility, built to handle the dense power requirements of modern artificial intelligence hardware,
can run to $14 million or more per megawatt of capacity.
Meaning a 100-megawatt AI campus could cost over $1.5 billion before a single customer ever moves in.
Let us break down where the money goes.
The electrical systems eat the largest share, typically 40% to 50% of total construction cost.
This includes the substation on your property, the transformers that step voltage down to usable levels,
the switchgear that routes power through the building, and the uninterruptible power supply systems,
which are massive battery banks that keep the lights on for 10 to 30 seconds during a grid outage while the backup generators spin up.
Those generators, diesel-powered, industrial-scale machines, sit outside the building
and can run the entire facility for 24 to 72 hours on stored fuel.
They represent another significant line item.
And all of this redundant electrical infrastructure exists for one reason.
Because a customer whose business goes offline for 30 seconds can sue you.
The cooling systems take another 15 to 20%.
Every server in your building converts electricity into heat. All of it.
100% of the electrical energy your customers use comes out the other side as heat.
And if you do not remove that heat continuously, the servers destroy themselves within minutes.
Traditional cooling uses large air conditioning units that pump chilled air under raised floors and through server rows.
More modern facilities pump chilled water through pipes to heat exchangers inside the data hall.
And the newest high-density facilities, especially those built for artificial intelligence workloads,
are moving toward liquid cooling systems that pipe coolant directly to the chips on the servers.
Because air simply cannot move heat fast enough when a single rack of AI hardware is drawing 60, 80, or 120 kilowatts of power.
The building shell, by the time you have added reinforced concrete floors capable of supporting server cabinets
that weigh as much as a small car, fire suppression systems, security infrastructure including biometric access controls and camera systems,
and a finished interior takes another 20 to 30% of your budget.
And then there are the fiber connections.
Running underground conduit from multiple directions, connecting to multiple carriers,
ensuring that there is no single cable whose accidental severing can take your entire facility offline.
You have spent all of this money, and you have not made a single dollar yet.
This is where most people underestimate the data center business.
They see the construction cost and think that is the risk.
The construction cost is not the risk.
The construction cost is just the price of entry.
The real risk begins the moment you open the doors and start trying to fill the building.
A data center that sits empty is a financial disaster in slow motion.
You have borrowed money at some interest rate, currently for large infrastructure projects,
somewhere in the range of 6 to 8% per year, to build a facility that costs hundreds of millions of dollars.
The interest on that debt does not wait for customers to arrive.
It accrues every single month.
A $100 million project at 7% interest costs $7 million a year in interest expense alone,
even if nobody has moved a single server into your building.
So the urgency of filling that building is enormous.
And this is where the business model, what you are actually selling, becomes critically important.
There are a few different ways to sell data center space, and they produce very different economics.
The simplest model is called retail co-location.
A small business, say, an accounting firm or a regional bank, does not want to run its own server room.
It pays you to put its servers in your building, in a locked cage or a dedicated cabinet,
and you provide the power, cooling, and physical security around it.
You might charge this customer somewhere between $100 and $200 per kilowatt of power per month.
On a single cabinet drawing 5 kilowatts, that is $500 to $1,000 a month.
Multiplied across hundreds of customers, this model generates very high revenue per square foot.
But it also requires a large sales team to find and sign all those small customers.
And those customers tend to leave more often than large ones.
The second model is wholesale co-location, or what the industry calls hyperscale leasing.
Here, instead of selling to hundreds of small businesses, you sign a single contract with a massive company.
Amazon Web Services, Microsoft Azure, Google Cloud, or Meta.
And that company leases an entire building, or a very large portion of one, for 10 to 15 years.
The revenue per kilowatt is lower in this model, sometimes as little as $80 to $110 per kilowatt per month.
But the revenue predictability is extraordinary.
A signed 15-year lease with Microsoft is effectively a guaranteed income stream for the next decade and a half.
You can take that contract to a bank and use it to finance your next building at a lower interest rate,
because the bank can see exactly how much money is going to flow in.
This is what makes long-term contracts the most valuable asset in the data center business.
They are not just contracts, they are financing instruments.
A facility with 10 years of signed revenue from creditworthy tenants
can borrow money at rates a half to a full percent lower than a facility with empty space.
And on a billion-dollar project, a half-percent reduction in borrowing costs saves $5 million a year.
If you are enjoying this breakdown of how the data center business actually works,
do subscribe to the channel and give this video a like.
It genuinely helps more people find this content.
But the contract model carries a hidden danger that most people do not see until it is too late.
That danger is called customer concentration.
If you build a 50-megawatt facility and sign a single contract with one cloud provider for the entire building,
your financial statements look spectacular.
100% occupancy, predictable revenue, low operating costs, clean.
But you are also building a business that depends entirely on one customer not leaving.
And when that contract expires in 10 or 15 years,
or if the customer decides not to renew,
you are suddenly left with a purpose-built facility that was designed around that customer's specific requirements,
sitting empty, still accumulating interest charges and maintenance costs,
with no revenue coming in at all.
The history of commercial real estate is full of this scenario.
A large tenant signs a long lease, the developer builds around them,
and 15 years later the tenant leaves and the building becomes a problem.
In the data center world it is more extreme,
because the facility has been optimized for a specific power density,
a specific cooling configuration, a specific network layout.
Repurposing it for a different use is expensive.
The opposite approach, many small customers across many industries,
protects you from any single departure, destroying the business.
But more customers means more salespeople, more billing complexity,
shorter lease terms, and higher churn.
Operating expenses are higher, and you are constantly refilling space that keeps turning over.
Most successful operators end up with a hybrid,
a core of large anchor tenants for revenue stability,
surrounded by smaller tenants for margin and diversity.
Now let us talk about what keeps you awake at night when you own one of these things.
The electricity bill.
Power is not a minor operating cost in a data center.
In most facilities, electricity accounts for somewhere between 60 and 80% of total ongoing operating expenses.
And that expense is driven not just by how much power your customers use,
but by how efficiently your building converts grid power into computing power.
The industry measures this with a metric called power usage effectiveness, or PUEE.
The calculation is simple.
You take the total power flowing into the building,
and divide it by the power that actually reaches the customer equipment.
A perfect facility, one where every watt of incoming electricity goes directly to compute,
with zero losses, would have a PUEE of 1.0.
That does not exist in the real world.
What exists is the range between roughly 1.15,
which represents an extremely efficient modern hyperscale facility,
and 2.0 or higher, which represents an old, inefficient legacy building,
where for every watt of computing power delivered,
another watt is consumed by cooling systems, conversion losses, and lighting.
Why does that matter in dollars?
Consider a facility consuming 50 megawatts of total power at an electricity rate of 7 cents per kilowatt hour.
At a PUEE of 1.5, it is consuming roughly 17 additional megawatts beyond what the customer's equipment actually needs.
Just in overhead.
At 7 cents per kilowatt hour, that overhead costs approximately $10.5 million per year.
Bring that PUEE down to 1.2, through better cooling systems and power management,
and the overhead drops to 10 megawatts, saving about $5 million annually.
That $5 million goes straight to the bottom line.
The electricity bill is not a utility expense.
It is the single most important lever in the entire business.
Now, about the AI revolution that everyone is talking about,
artificial intelligence has fundamentally changed the economics of building data centers,
and not necessarily in the direction you might expect.
The demand is real, enormous, historically unprecedented demand for computing capacity.
But the demand comes with requirements that are dramatically more expensive to meet than anything the industry has dealt with before.
A traditional server cabinet in a standard enterprise data center draws somewhere between 5 and 10 kilowatts of power.
A cabinet full of NVIDIA H100 graphics processing units,
the chips used to train large AI models draws 40 to 60 kilowatts,
and newer architecture is pushing past 100 kilowatts per cabinet.
That is not just more power. That is a completely different class of infrastructure problem.
At 40 kilowatts per cabinet, air cooling stops working.
Air simply cannot absorb and remove heat fast enough.
You need liquid cooling, pipes running to each server,
carrying chilled water or a specialized coolant directly to the chips themselves.
The infrastructure for this is expensive to install, expensive to maintain,
and requires modifications to the building structure,
because high-density AI cabinets weigh far more than standard servers,
and standard raised data center floors were not built to hold them.
The result is that building a data center specifically designed for AI workloads
costs dramatically more per megawatt than building a standard enterprise facility.
We are talking about $14 to $20 million per megawatt for an AI-optimized building,
compared to $8 to $12 million for a conventional one.
But here is the other side of the equation.
The revenue per kilowatt for AI-grade capacity is also dramatically higher.
Operators have been signing AI workload contracts at prices significantly above standard wholesale rates
because the customers, the major AI companies, need this infrastructure immediately
and cannot build it fast enough themselves.
The question that nobody can answer with certainty right now
is whether the higher construction cost is justified by the higher revenue,
and for how long.
AI hardware generations are moving fast.
The GPU that costs $400 million today
and requires a specific cooling configuration might be obsolete in three years,
replaced by something that needs a completely different infrastructure setup.
Building a specialized AI facility today
means betting that the revenue you earn in the next few years
will more than cover the extra construction cost compared to a standard building.
Most large operators are making that bet.
But it is a bet.
Let us run the numbers on what a mid-sized data center business actually looks like.
Financially, take a facility with 20 megawatts of customer-facing power capacity.
You built it for roughly $200 million.
You financed 70% of that, $140 million,
with debt at 7% per year,
which costs you $9,800,000 per year in interest.
You financed the remaining $60 million with your own equity.
If you can fill the building with wholesale co-location customers
at an average of $100 per kilowatt per month,
20 megawatts of capacity would generate $24 million per year in power-related revenue.
Add connectivity services, cross-connect fees,
which are the physical cables that connect one customer's network to another inside your building
and generate gross margins above 90%,
and managed service fees,
and a well-run facility might generate $28 to $32 million
in total annual revenue at full occupancy, now deduct operating costs.
Your electricity bill for 20 megawatts at a PU of 1.3
and an electricity rate of 7 cents per kilowatt hour
runs to approximately $16 million per year.
Staff, maintenance, insurance, and property costs add another $3 to $4 million.
So you are spending $19 to $20 million running the building
and taking in $30 million,
leaving $11 to $12 million in operating income before debt service.
Subtract the $9,800,000 in annual interest
and you are left with roughly $1 to $2 million of net cash flow per year
on a $200 million investment.
That is a thin margin.
It gets better as the debt gets paid down
and dramatically better if you can refinance into lower-cost debt
when the building is fully occupied and generating stable cash flow.
But it illustrates why utilization is everything in this business.
Drop from 100% occupancy to 70% occupancy
and that $11 million in operating income shrinks to roughly $7 million,
which barely covers your interest payments.
Drop to 50% occupancy and you have a problem.
The single most valuable thing in the data center business
is a full building on a long lease with a credit-worthy customer.
Everything else—the power efficiency, the cooling technology,
the tier classification, the fiber diversity—
exists in service of that one goal, speaking of reliability.
The data center industry has a formal rating system
developed by a research organization called the Uptime Institute.
The tiers go from 1 to 4,
describing how much redundancy exists in power and cooling.
A tier 1 facility has no backup systems.
If any component fails, the building might go down.
Nobody running critical business operations buys space there.
A tier 4 facility has complete redundancy across every system.
Two separate utility feeds,
two independent power paths through the building,
two independent cooling systems.
Any single failure, anywhere, causes zero downtime.
This is what major banks, cloud providers, and government agencies require.
The cost difference between tier 3 and tier 4
adds $5 to $8 million per megawatt of capacity.
On a 20-megawatt building, that is an additional $100 to $160 million.
You had better be able to charge for that.
And the reality is that many customers who demand tier 4 on paper
quietly accept tier 3 when they see the price difference.
Operators who have built out full tier 4 capacity
and failed to fill it at tier 4 pricing rates
have paid dearly for that miscalculation.
Now let us talk about what happens when the model breaks.
In the mid-2010s,
a wave of developers built facilities in secondary markets,
cities with cheap land and available power,
but limited fiber and a small local customer base.
They financed at low interest rates
and modeled their lease-up timelines
based on what was happening in primary markets
like Northern Virginia and Silicon Valley.
The demand did not follow.
Buildings sat at 40% or 50% occupancy for years.
When rates rose, debt costs climbed while revenue stayed flat.
Some facilities were sold at significant discounts to construction costs.
Some were foreclosed on.
A few sat partially empty for nearly a decade.
The lesson is not that data centers are bad investments.
The lesson is that a data center is a capital-intensive illiquid asset
with enormous fixed costs.
If any link breaks, tenants arrive slower than modeled,
power costs rise, interest rates climb,
competing supply hits the market
at the same moment you are trying to lease up.
The economics can move from attractive
to devastating with surprising speed.
Before we get to where the whole industry is heading,
here is something worth having in your back pocket.
If this kind of real-world breakdown
of how businesses actually work appeals to you,
there is a detailed business guide available in the description
and the first pinned comment.
It covers the core financial frameworks
that apply across industries like this one.
The current moment in the data center industry
is unlike anything the sector has experienced before.
The demand being driven by artificial intelligence
is not a normal demand cycle.
The hyperscalers, Amazon, Microsoft, Google, Meta,
have collectively committed to spending hundreds of billions of dollars
on computing infrastructure over the next several years.
Microsoft alone announced plans to spend $80 billion
on data center construction in a single fiscal year.
At those spending levels,
the bottlenecks have shifted from customer demand to physical supply.
The constraints right now are electricity and people.
Electrical grid capacity is the hardest constraint.
In Northern Virginia, some developers report
waiting queues of five to seven years for new utility connections,
which means a company that decides today
to build a new data center in the primary market
cannot expect to have power flowing through it
until well into the next decade.
This has pushed development into secondary and tertiary markets,
Midwest cities, parts of the Southeast,
rural areas near hydropower.
Anywhere that power can be secured more quickly,
the equipment shortage is also real.
Large power transformers have lead times
of two to three years from major manufacturers.
Generator sets, switchgear, and liquid cooling systems
are all backlogged.
Developers who signed leases with hyperscale customers
and then could not build fast enough
because equipment was stuck in a manufacturing queue
have found themselves contractually committed
to delivering power they cannot yet produce.
This is not a hypothetical.
It has happened repeatedly.
What this means for the economics is complicated.
On the positive side, supply constraints
allow existing data center operators
with available capacity to charge premium rates.
Companies that built and filled facilities
before the AI demand wave hit
are generating returns that looked impossible
when they signed the original leases.
On the negative side, anyone building now
is doing so at construction costs
dramatically higher than historical norms,
both because of inflation in materials and labor
and because the AI-specific requirements
push the build cost higher per megawatt.
The fundamental financial equation has not changed.
Revenue in minus electricity
minus maintenance minus staff minus debt service
minus a reserve for replacing equipment,
that is what you actually make.
Every variable can move against you at once.
Electricity prices rise.
Customers sign slower than modeled.
Construction runs over budget.
Interest rates climb.
The equipment you spent years paying off
becomes obsolete before it is fully depreciated.
None of these individually destroys a well-run business.
All of them together is how a data center
goes from looking like a gold mine
to becoming a problem that gets sold at a loss.
The people who win in this business consistently
are not the ones who built the biggest buildings.
They are the ones who understood
that the building is just the container
and that the real business is securing power
that others cannot access,
building reliability that customers cannot do without,
signing contracts that lock in revenue
before the ink on the building permits is dry,
and then using those contracts
to finance the next building
at a lower cost than their competitors.
It is a business about controlling scarce infrastructure.
The scarcest thing in the entire digital economy
right now is not computing hardware.
The hardware can be bought if you can afford it.
The scarcest thing is reliable electricity
delivered at scale with a commitment from a utility
that it will be there tomorrow
and the day after that for the next 20 years,
as artificial intelligence continues consuming more power
than any previous computing era.
And the projections suggest
that AI data centers could consume
somewhere between 8 and 21 percent
of total U.S. electricity by 2030,
according to analysis from Goldman Sachs.
The data center operator
who secured their grid connection agreements five years ago
is sitting on something more valuable
than they may have understood when they signed the paperwork.
That is the real business.
Not servers. Not buildings.
Not fiber cables.
Megawatts.
And who controls the megawatts
in the era of artificial intelligence
may turn out to be one of the most
consequential economic questions of the decade.
If you found this video valuable,
please like it and subscribe to the channel
if you have not already.
It takes a few seconds, and it means a great deal.
Thanks for watching.