AI Data Center Economics: What It Costs, and Is AI a Bubble?

What You'll Learn

By the end of this article, you will understand:

• What an AI data center physically is, starting from zero, and why the industry measures it in megawatts instead of square feet
• Exactly where the money goes when someone announces a "$10 billion AI data center", line by line
• Why the electricity bill, despite all the headlines, is not the scariest number on the page
• How one accounting assumption, buried in a footnote, swings the entire investment case
• The three completely different ways a data center earns money, and the break-even arithmetic for each
• Why the price of AI is collapsing and rising at the same time, and what that does to the asset
• Which layer of this industry actually keeps the profit, and which layers just pass money along
• How India's data center economics genuinely differ from America's, and which listed companies give you exposure
• The full bull and bear case on whether this is a bubble, with a checklist of what to actually watch

This article is about the money. If you want the layer-by-layer map of which company builds which component, we covered that separately in what really powers ChatGPT. This one follows the rupee and the dollar.

---

An AI Data Center Is Mostly Not a Building

Somewhere in Texas, or Andhra Pradesh, or rural Ireland, there is a low grey structure with no windows. It looks like a warehouse. It cost more than a nuclear power plant.

That comparison is not rhetorical. The economics of an AI data center are unlike any other large construction project, and the numbers attached to single sites are now genuinely difficult to hold in your head. Meta's Hyperion campus in Richland Parish, Louisiana is reported at roughly $27 billion for a site targeting up to 5 gigawatts. Closer to home, Google has committed approximately $15 billion over five years to an AI hub at Visakhapatnam, built in partnership with AdaniConneX and Bharti Airtel, its largest such hub anywhere outside the United States.

For scale, India's entire planned data center buildout to FY30 is estimated by CareEdge Ratings at ₹1.5 lakh crore of infrastructure spending, and roughly ₹3 to 4 lakh crore once you include the computing equipment that goes inside.

Put those next to each other and the scale of the American buildout becomes hard to ignore.

Google and AdaniConneX, Visakhapatnamsingle hub, over 5 years~$15bn

Meta Hyperion, Louisianasingle campus, up to 5 GW~$27bn

India, entire national buildout to FY30infrastructure plus tenant IT equipment$34bn to $45bn

A single campus in Louisiana costs roughly the same as everything India plans to build, nationally, over five years. That is not a comment on India. It is the clearest available measure of how much capital the American AI buildout is absorbing.

Here is the first thing that surprises people. The building is the cheap part.

This single fact reorganises how you should think about the entire sector. A data center is not a real estate business wearing a technology costume. It is a technology business wearing a real estate costume. That distinction determines who profits, how long the profits last, and what could go wrong.

Let us build the whole picture from the ground up.

---

Part 1: What an AI Data Center Actually Is

Start With One Computer

You know what a computer is. It has a processor that does the thinking, memory that holds what it is working on, and storage that holds everything else. It draws maybe 100 watts from the wall.

A watt is not an amount of electricity used up. It is a rate, like a speedometer reading, not an odometer reading. To turn a rate into an actual amount used, and a cost, you need to know how long it ran for too.

The formula is universal. What is not universal is the price of each kWh, which is set locally and varies enormously. Here is what that same 1 kWh would actually cost a household in two very different markets:

Here is where 100 watts, as a rate, sits next to things you already own. These are ballpark figures, meant to build intuition rather than to be exact.

A hundred watts is a laptop and a half, or a tenth of a kettle. It is a small number. The reason it matters is what happens when you multiply it by a very large number of computers, which is exactly what a data center is, and it is why the industry stopped measuring these buildings in square feet and started measuring them in watts, as the next section explains.

A server is the same idea, built for a rack rather than a desk. Instead of a screen and keyboard, it has network ports. Instead of one processor, it may have several. It lives in a metal frame called a rack, which is roughly the size of a tall refrigerator and holds many servers stacked vertically.

A traditional data center is thousands of those racks in a room, keeping websites, email and banking systems running. This has existed since the 1990s and it is a mature, unglamorous, moderately profitable business.

Now Add the GPU

A GPU, or graphics processing unit, was originally designed to draw video game images. It turns out that the mathematics of drawing images and the mathematics of training artificial intelligence are almost the same: enormous numbers of simple calculations, all happening at once. A regular processor is a brilliant professor doing one hard problem at a time. A GPU is ten thousand school students doing simple arithmetic simultaneously.

AI needs the ten thousand students.

Training TrainingThe one-off process of building an AI model by feeding it enormous quantities of data until it learns patterns. Takes weeks or months and consumes vast computing power. Think of it as building the factory.See all terms in the glossary is the process of building an AI model, feeding it enormous quantities of text or images until it learns patterns. It is done once per model, takes weeks or months, and consumes staggering amounts of computing power. InferenceInferenceWhat happens every time someone actually uses an AI model: you ask a question, it answers. Cheap per query but done billions of times. Think of it as running the factory.See all terms in the glossary is what happens every time you actually use the model, when you type a question and it answers. Training is the factory being built. Inference is the factory running.

Both need GPUs. This is why the AI data center exists.

Why the Industry Talks in Megawatts

Here is where AI facilities break from the traditional model, and it is the single most important physical fact in the sector.

A rack of ordinary servers draws around 5 to 10 kilowatts of power. A rack packed with modern AI accelerators can draw 100 to 130 kilowatts or more. That is more than ten times the power in the same physical footprint, and every watt that goes in comes back out as heat that must be removed.

📌 note

A chip does no mechanical work that leaves the room, unlike a fan or a motor. So almost all the electricity it consumes has nowhere else to go but out as heat. A rack pulling 120 kilowatts is, for practical purposes, a 120-kilowatt heater. That is why a data center's cooling capacity has to roughly match its electrical capacity, not just top it up.

tap to close📌

A chip does no mechanical work that leaves the room, unlike a fan or a motor. So almost all the electricity it consumes has nowhere else to go but out as heat. A rack pulling 120 kilowatts is, for practical purposes, a 120-kilowatt heater. That is why a data center's cooling capacity has to roughly match its electrical capacity, not just top it up.

Two separate things stack together to produce that gap, and it is worth seeing both.

1 Each AI chip is simply hungrier than a normal one. A processor built for everyday computing spends much of its time waiting, for a disk, a network request, a person to click something, so it is designed to sip power at idle and burst only when needed. A GPU built for AI has no idle moments by design: it packs thousands of small cores meant to run flat out, together, for as long as a training job lasts. That difference shows up directly in the spec sheet.

A single modern AI GPU alone can draw two to three times what an entire CPU socket draws, before you even get to the second factor.

2 Then you pack far more of them into the same rack. NVIDIA's own reference design, the GB200 NVL72, crams 72 GPUs into a single rack, drawing about 120 kilowatts, which is where the "100 to 130 kilowatts" figure above actually comes from. Operators pack them this tightly on purpose: as Part 2 explains, these chips must sit close together to talk to each other fast enough, so spreading them across more, emptier racks is not really an option.

Once rack densityRack DensityThe power drawn by a single rack of equipment. Ordinary servers draw 5 to 10 kilowatts per rack; AI racks draw 100 kilowatts or more. Above roughly 30 kW, air cooling stops working and liquid cooling becomes mandatory.See all terms in the glossary passes roughly 30 kilowatts, blowing cold air at the equipment stops working. Air simply cannot carry heat away fast enough. The industry has therefore moved to liquid coolingLiquid CoolingCirculating coolant directly across chips through pipes instead of blowing cold air at them. Necessary above about 30 kW per rack, and reported to cut energy overhead by up to 30%. It is why older data centers cannot simply be filled with AI hardware.See all terms in the glossary, circulating coolant directly across the chips through pipes. This is more efficient, but it means the building itself must be designed differently, which is why older data centers cannot simply be filled with AI hardware.

The Three Kinds of Owner

Before we price anything, understand who is doing the building, because their economics are not the same.

A hyperscalerHyperscalerOne of the handful of giant technology companies running global cloud platforms (Microsoft, Amazon, Google, Meta). They build data centers primarily for their own use and rent out the surplus.See all terms in the glossary is one of the handful of enormous technology companies operating global cloud platforms. A colocationColocation (Colo)A specialist landlord model for computers. The operator builds the powered, cooled shell and leases it; the tenant brings their own servers. Revenue comes from long contracts, typically 5 to 15 years.See all terms in the glossary operator is essentially a specialist landlord for computers. A neocloudNeocloudA newer breed of company that buys GPUs outright and rents computing capacity by the hour, competing with hyperscalers on price and availability. Examples include CoreWeave, Nebius and E2E Networks.See all terms in the glossary is a newer breed of company that buys GPUs and rents them out by the hour, competing with the hyperscalers on price and availability.

They face the same physics and completely different financial risks. Hold that thought.

---

Part 2: AI Data Center Cost Breakdown, Line by Line

Let us price one large facility properly. The research group Epoch AI published a detailed teardown of a hypothetical 1 gigawatt AI data center built to current US hyperscaler standards, and it is the clearest public breakdown available. Their assumptions: NVIDIA GB200 NVL72 systems, US electricity at 8.34 cents per kilowatt-hour, a PUEPUE (Power Usage Effectiveness)Total facility electricity divided by the electricity reaching the computing equipment. 1.0 would be perfect; 2.0 means one watt wasted on cooling and overhead for every useful watt. Best global operators run 1.1 to 1.2; conventional Indian facilities run 1.5 to 1.9.See all terms in the glossary of 1.14 (drawn from Lawrence Berkeley National Laboratory's research on AI data center energy use), 71% utilisation (an average of four separate estimates collected by researcher Tyler Norris), a five-year life for the computing equipment and fourteen years for the building.

Servers and GPUs56% of the total$21.2bn

Facility and power30%$11.4bn

Networking13%$4.9bn

Landunder 1%$0.17bn

Utility worksunder 1%$0.16bn

Total: about $37.9 billion. Source: Epoch AI cost teardown (link in Sources).

Read that chart slowly, because it contains the whole thesis.

Land is under half a percent. For all the talk of data centers as a property play, the dirt is a rounding error. Location matters enormously, but it matters for access to power and fibre, not for the value of the land itself.

The building and its electrical systems are 30%. This is the transformers, switchgear, backup generators, uninterruptible power supplies, chillers and the shell itself. It is real money, roughly $11.4 billion, and it is what a colocation landlord actually owns.

Servers and networking together are 69%. More than two-thirds of an "AI data center" is computing equipment. And within that, the dominant cost is the accelerator chips themselves.

The $4.9 Billion of Cable Nobody Talks About

The networking line deserves a moment, because at 13% of capital cost it is larger than the land, the utility connection and the annual power bill put together, and almost nobody explains why.

In a traditional data center, servers mostly work independently. One machine serves one website. Networking is important but modest.

Training a large AI model is the opposite. The work is split across thousands of GPUs that must stay in lockstep, exchanging intermediate results constantly. If one GPU waits, they all wait. A cluster is only as fast as the connection between its slowest pair of chips, which means the network is not a convenience layer. It is part of the computer.

This forces expensive choices: very high-speed optical transceivers, specialised switching fabrics, and enormous quantities of fibre inside the building. The cost scales worse than linearly, because connecting twice as many GPUs to each other takes more than twice as many connections.

Here is why, using an old party trick. If ten people at a party each shake hands with everyone else exactly once, that is not ten handshakes. It is 45, because each of the ten people shakes nine hands, and dividing by two so you do not count each handshake twice gives 10 × 9 ÷ 2 = 45. Double the party to twenty people and handshakes do not double to 90. They nearly quadruple, to 190.

Real clusters do not literally wire every GPU straight to every other one. They use switches, the same way a phone network uses exchanges instead of a wire from every phone to every other phone. But the switches exist precisely to manage this same explosion of possible pairs, and the total switching capacity, cabling and bandwidth needed still grows faster than the GPU count itself. That is the "worse than linearly" in plain terms: add chips, and the plumbing between them grows even faster than the chips did.

What This Means Per Megawatt

Divide through and you get roughly $38 million per megawatt all-in, of which about $21 million is IT equipment and about $12 million is the facility.

That facility number is the one you will see quoted in construction research. JLL's 2026 Global Data Center Outlook puts the global benchmark for a standard build at about $11.3 million per megawatt, up from $7.7 million in 2020, a rise of nearly 47% in six years. AI-ready facilities with liquid cooling and higher power density run materially above that, with industry cost indices placing the liquid cooling premium alone at roughly $2.7 million to $3.7 million per megawatt. Costs have been climbing: steel, copper, transformers and skilled electrical labour are all in short supply.

The Same Building in India

CareEdge Ratings, in a March 2026 sector report, puts Indian construction costs at roughly 30% to 40% lower than China and the United States, driven mainly by cheaper land and labour. Their cost split for an Indian facility is instructive:

Notice what is missing from that table. It covers the facility, not the computing equipment, because in India most operators are colocation landlords. The tenant brings the servers. CareEdge estimates tenants spend 1.5 to 2 times the landlord's investment on IT equipment, which is how ₹1.5 lakh crore of Indian infrastructure capex turns into ₹3 to 4 lakh crore of total investment.

We will return to India properly in Part 9.

---

Part 3: Data Center Energy Consumption, and Why the Power Bill Is Not the Scary Part

Every article about AI leads with energy. Let us actually do the arithmetic.

Our 1 GW facility, running at 71% utilisation on 8.34 cent electricity, spends about $594 million a year on power. That is a genuinely enormous number. It is also, as we are about to see, not the number that decides whether the investment works.

Energy65% of opex$594m

Taxes16%$143m

Maintenance13%$120m

Labour4%$40m

Utility works2%$20m

Total: about $907 million a year. Source: Epoch AI.

Note the labour line. A facility representing $38 billion of capital employs enough people to cost $40 million a year in wages.

📌 note

Data centers create very few permanent jobs relative to their capital. That is an underappreciated political risk wherever a local government hands out incentives expecting employment in return.

tap to close📌

Data centers create very few permanent jobs relative to their capital. That is an underappreciated political risk wherever a local government hands out incentives expecting employment in return.

Data centers are extraordinarily capital-intensive businesses, and it shows up nowhere more clearly than in that one line.

Understanding PUE

You will see one acronym constantly in this industry, so let us define it precisely.

PUEPUE (Power Usage Effectiveness)Total facility electricity divided by the electricity reaching the computing equipment. 1.0 would be perfect; 2.0 means one watt wasted on cooling and overhead for every useful watt. Best global operators run 1.1 to 1.2; conventional Indian facilities run 1.5 to 1.9.See all terms in the glossary, or Power Usage Effectiveness, is total facility electricity divided by the electricity that actually reaches the computing equipment. A PUE of 1.0 would be perfect: every watt goes to computing and none is wasted on cooling, lighting or conversion losses. A PUE of 2.0 means you burn one watt on overhead for every watt of useful computing.

The comparison between markets is revealing:

That Indian range matters. India enjoys electricity tariffs roughly 50% below US levels, which sounds like a decisive advantage. But a PUE of 1.7 against 1.14 means an Indian facility burns about 49% more total electricity for the same useful computing. A large chunk of the tariff advantage is consumed by the hotter climate and older cooling technology before it reaches the bottom line. Newer Indian builds with liquid cooling close much of this gap, which is exactly why the newest campuses are being designed very differently from the ones built five years ago.

The Reveal

Now put the two halves together. Upfront capital of $37.9 billion. Annual operating cost of $907 million.

Spread the capital over its useful life and Epoch AI arrives at an annualised total cost of ownership of roughly $8.5 billion a year. Of that, about $5.0 billion is the servers alone.

The electricity bill is $594 million a year. The depreciation on the chips is roughly $5 billion a year. Power is about 12% of the true annual cost of running an AI data center. The hardware wearing out is about 60%.

This is the most important sentence in the article. Energy dominates the headlines because it is visible, physical and politically charged. But if you are trying to work out whether these investments pay off, you should be watching the depreciation schedule far more closely than the power contract.

Which brings us to the number that decides everything.

---

Part 4: GPU Depreciation, the Footnote That Swings the Thesis

What Depreciation Actually Is

If a company buys a machine for ₹100 crore that will last ten years, it does not record a ₹100 crore expense in year one. It spreads the cost across the years the machine is useful, recording ₹10 crore of depreciation each year. The number of years chosen is called the useful lifeUseful LifeThe number of years a company assumes an asset will remain productive, which sets its annual depreciation charge. For AI hardware this estimate is contested: Microsoft and Alphabet extended server lives to six years while Amazon shortened a subset to five.See all terms in the glossary, and it is an accounting estimate made by management.

For most industries this is a boring, stable assumption. For AI data centers it is the whole ballgame.

Why It Matters So Much Here

Look at what happens to our 1 GW facility when you change only the assumed life of the computing equipment, holding everything else constant. Before looking at the chart, it is worth actually doing the arithmetic, using nothing but numbers already on this page, so the totals below do not feel like they appeared from nowhere.

Split the $37.9 billion of capex from Part 2 into the two piles that behave differently. Servers, GPUs and networking, 56% plus 13%, were $26.1 billion, and that is the pile whose life is being tested here. The facility, land and utility works, the remaining 31%, were $11.7 billion, and its 14-year life stays fixed throughout this exercise.

With that mechanism in view, here is the full sensitivity Epoch AI actually publishes:

3-year hardware lifeAggressive obsolescence assumption$12.0bn

5-year hardware lifeThe base case$8.5bn

7-year hardware lifeOptimistic longevity assumption$7.0bn

Same facility, same electricity, same everything. Only the assumed useful life of the IT equipment changes. Source: Epoch AI sensitivity analysis (link in Sources).

The annual cost of the identical building swings by about 70%, from $7 billion to $12 billion, purely on an accounting estimate. No physical fact changed. Nobody bought a different chip or signed a different power contract. One assumption moved.

The Disagreement Is Public and Real

Here is what makes this genuinely interesting rather than merely technical. The largest, best-informed buyers of AI hardware on earth do not agree with each other, and they have moved in opposite directions in public filings.

The Question Behind the Question

Strip away the accounting and the real question is this: when a new generation of chips arrives that is three times faster per watt, what is the old generation worth?

If the answer is "still useful for years, just for cheaper work", then six-year lives are honest and the economics hold. If the answer is "it costs more in electricity and cooling than the work it produces is worth", then the hardware is effectively scrap the moment it becomes uncompetitive, and a large amount of reported profit across the sector is borrowed from the future.

Both answers are defensible today. Watch the second-hand market for older accelerators and the disclosed utilisation ratesUtilisation RateThe share of time computing hardware is actually being used and paid for. Because the cost of a GPU is fixed once bought, utilisation determines whether a GPU rental business earns a high margin or loses money.See all terms in the glossary on older fleets. Those are the evidence that will settle it.

---

Part 5: How AI Data Centers Actually Make Money

We have priced the asset. Now, how does it generate a return? Three models, three completely different risk profiles.

Model 1: The Hyperscaler Self-Build

Microsoft, Amazon, Google and Meta build for their own use. There is no rent, no tenant and no lease. The data center is a cost center that makes a product possible, and the return shows up somewhere else entirely, in cloud revenue, advertising revenue or software subscriptions.

This is the hardest model to evaluate from outside, and the disclosure makes it harder still. Microsoft does not break out AI revenue as a separate line at all. What it does report for the fourth quarter of fiscal 2026 is Microsoft Cloud revenue of $59.3 billion, up 27%, Azure and other cloud services up 43%, and Azure passing $100 billion of annual revenue for the first time. Those are real, collected revenues. But AI workloads and ordinary cloud workloads run on overlapping infrastructure, so there is no honest way to divide a return figure out of them.

Be wary of the AI run-rate numbers that circulate in secondary coverage. Several widely repeated figures do not appear in any hyperscaler's own filings.

What you can watch is the ratio of capital expenditure to operating cash flow. Historically the large technology platforms spent around 40% of operating cash flow on capex. In 2026, collective capex guidance across the four largest reached roughly $660 billion to $725 billion depending on whose tally you use, against about $410 billion in 2025. That is approaching, and by some estimates reaching, essentially all of their operating cash flow.

Model 2: The Colocation Landlord

This model is straightforward and genuinely resembles real estate. You build the powered, cooled shell. A tenant signs a long lease, brings their own servers, and pays you for space, power and cooling.

The economics are annuity-like. CareEdge Ratings reports Indian colocation contracts typically run five to fifteen years with strong counterparties, and that Indian operators have sustained EBITDA margins of 40% to 43% from FY22 to FY25, with capacity absorption above 90% throughout and reaching 96.1% in December 2024.

Two features define the risk here:

Take-or-payTake-or-Pay ContractAn agreement where the tenant pays for contracted capacity whether or not they use it. It is what makes data center revenue reliable, which means the credit quality of the counterparty matters more than almost anything else.See all terms in the glossary contracts. The tenant commits to pay for contracted capacity whether or not they use it. This is what makes the revenue reliable, and it means the credit quality of the tenant matters more than almost anything else. A fifteen-year take-or-pay contract with a well-capitalised hyperscaler is a bond. The same contract with a thinly funded AI startup is an option on that startup's survival.

Pre-leasing. Serious operators sign tenants before construction completes. Capacity built without a signed tenant is speculative, and speculative capacity is where losses live in every property cycle ever recorded.

The catch is leverage. CareEdge notes that Indian data center capex is typically financed with 70% to 80% debt, that cash flows take two to three years to materialise, and that interest coverage has been modest at two to three times. That is a normal infrastructure profile, and it is fine while demand holds. It is unforgiving if demand pauses.

Who Actually Does This, and Who Their Tenants Are

Naming a colocation landlord is easy. Naming its tenants usually is not, because most leases carry confidentiality clauses that keep both the tenant's identity and the exact terms out of public view. What follows below are the exceptions: relationships confirmed by an official announcement from at least one side of the deal, not inferred from industry chatter.

Model 3: The Neocloud

The newest and most exposed model. A neocloud buys GPUs outright and rents computing capacity by the hour. Two questions usually come up here, and they are worth answering properly before the arithmetic, because the model only makes sense once you see where it sits relative to the other two.

Does a neocloud own the building? Usually not, at least not at first. A neocloud's balance sheet is built almost entirely around chips, not real estate. It typically signs a lease with a Model 2 colocation landlord for powered, cooled shell space, the same product a hyperscaler or an enterprise might rent, and then fills that leased space with GPUs and servers it owns and financed itself. In effect, a neocloud is simultaneously a tenant of Model 2 and a landlord, of a different kind, to its own customers: it does not sublease space, it subleases computing power by the hour.

CoreWeave is the clearest public example. Through 2025 its capacity came almost entirely from long-term leases with data center owners including Core Scientific, Digital Realty, Applied Digital and others, no single landlord holding more than 17% of its total. Its own announcement of acquiring Core Scientific outright in 2025 put a number on what that leasing had been costing: "over $10 billion of cumulative future lease overhead to be paid for existing contractual sites over the next 12 years." CoreWeave's CEO described the acquisition as "verticalizing the ownership" of that infrastructure, in his words, precisely to escape lease payments and gain direct control. That single sentence tells you the model: leasing the shell is the default, owning it is the exception large neoclouds graduate into once they are big enough to justify it.

How is this different from a hyperscaler "renting the surplus"? They can look similar from the outside, both are, in the end, someone paying by usage to run workloads on someone else's chips, so the real differences are in ownership depth and what is actually being sold.

With that distinction in view, here is why a neocloud's economics are so much tighter than either of the other two models. Let us do the arithmetic explicitly, because it is the clearest window into the whole industry.

Independent analysis puts the fully burdened cost of owning and operating an NVIDIA H100 at roughly $1.45 per GPU-hour at 70% utilisation, split into about $1.12 of hardware amortisation and about $0.33 of energy and facility costs.

Now compare that to what providers actually charge, as of August 2026:

At first glance this looks like a wonderful business. A $1.45 cost against a $2.15 to $6.16 price implies gross margins between 33% and 76%.

This is why the neocloud model is simultaneously the most exciting and the most fragile. It converts a fixed, depreciating, five-year capital commitment into revenue that is priced by the hour in a competitive market. If demand stays ahead of supply, the returns are excellent. If supply catches up, the cost is already sunk and the price is set by whoever is most desperate to fill capacity.

---

Part 6: AI Inference Costs and the Deflation Squeeze

Here is the mechanism that connects everything above to the actual product, and it is the part most coverage misses entirely.

AI Is Getting Radically Cheaper

AI models are sold by the tokenTokenThe unit AI models are billed by, roughly a word fragment. Prices for a given level of capability have fallen about 95% in two years, while frontier model pricing has risen, so the market has split in two.See all terms in the glossary, roughly a word fragment. Providers charge per million tokens processed. And the price of a given level of capability has collapsed.

Epoch AI has tracked this carefully: score AI models on a fixed benchmark, then track what it costs, over time, to buy that same score from whichever model is cheapest. Their finding is more nuanced than the headline version. Measured across six different benchmarks over three years, that cost fell somewhere between 9 times and 900 times cheaper every year, and the two ends of that span are not a rounding error, they reflect real differences in how fast each individual benchmark moved. One concrete point from inside that range: matching GPT-4's score on a set of demanding, PhD-level science questions got roughly 40 times cheaper with each year that passed.

Two honest caveats belong here. First, that wide span is exactly why no single number can honestly describe "the price of AI" in general, only the price of a specific capability. Second, the pace itself is new: most of the steepest drops Epoch measured landed in the final year of their three-year window, which is not enough runway yet to know whether a decline this fast is sustainable or a temporary catch-up.

DEFLATION

There is no precedent for that rate of price decline in enterprise software. It is a genuine engineering triumph, driven by better model architectures, better chips and vastly better inference optimisation.

And More Expensive at the Same Time

The confusing part is that the frontierFrontier (AI)The single most advanced capability level available at any given moment, the current cutting edge of what AI models can do. It is a moving target: a capability at the frontier today typically becomes cheap, widely available commodity capability within about eighteen months, as newer models push the boundary further out.See all terms in the glossary moved in the opposite direction. The most capable reasoning models in 2026 command single-digit to low-tens of dollars per million input tokens, and top-end pricing has risen as models spend far more computing power per query reasoning through problems. Because these prices change frequently, check the providers directly (OpenAI, Anthropic) or a current comparison rather than trusting any figure printed in an article, including this one.

What This Does to a Five-Year Asset

Now connect it back to the depreciation question, because this is where the two threads meet.

You have bought hardware that sits on your balance sheet for five or six years. The work that hardware does is repriced downward at rates measured in multiples per year, not percentages per decade. Your revenue per GPU-hour is therefore under continuous, severe pressure from the moment you switch the machine on.

The asset depreciates on a five-year schedule. The price of what it produces has been falling by multiples every single year. Those two clocks run at wildly different speeds, and the gap between them is where the entire risk of this industry sits.

There are two honest ways to read this.

The optimistic reading: falling prices expand the market faster than they compress margins. When something gets 200 times cheaper, people do vastly more of it. Cheap AI gets embedded in every application, total token volume explodes, and the total revenue pool grows even as the price per unit collapses. This is exactly what happened with bandwidth, storage and computing power in every previous technology cycle, and each time the volume growth won.

The pessimistic reading: the volume growth has to be extraordinary to offset price declines of that magnitude, and much of the new volume runs on newer, cheaper hardware rather than the fleet you already bought. Your specific installed base of two-year-old chips faces the price collapse without capturing the volume growth.

Both readings have historical support. The honest position is that this is unresolved, and that the resolution will differ by operator depending on how well they redeploy older hardware.

Part 7: Data Center Power Constraints, the Grid Bottleneck That Actually Bites

We established that electricity is only about 12% of the true cost. So why does energy dominate every conversation about this industry?

Because cost and availability are different problems. Power is cheap. Power is also, right now, nearly impossible to get quickly.

The Numbers, Honestly Presented

📌 note

Gartner projects that AI-optimised servers will consume more electricity than all conventional data center hardware combined by 2027.

tap to close📌

Gartner projects that AI-optimised servers will consume more electricity than all conventional data center hardware combined by 2027.

The composition of the load, not just its size, is changing.

The Context Most Coverage Omits

Now, a piece of honesty that rarely survives the headline stage. The IEA projects data center electricity demand to grow by about 530 TWh between 2024 and 2030. Over the same period:

Industry+1,936 TWh

Electric vehicles+838 TWh

Air conditioning+651 TWh

Data centers+530 TWh

Data centers are growing fast, but they are not the largest source of new electricity demand. Source: IEA (link in Sources).

Where the Real Bottleneck Is

The binding constraint is not generation. It is the equipment that connects generation to a building, and the queue to use it.

Transformers. The large units that step high-voltage transmission down to usable levels had lead times of roughly 24 to 30 months before 2020. Industry reporting through 2025 and 2026 places current lead times at two to four years, with some high-capacity units quoted as long as five. There are few manufacturers, capacity took decades to build, and nobody wants to add capacity for a demand spike that might not persist.

Interconnection queuesInterconnection QueueThe waiting list to connect a new facility to the electricity grid. Waits of four to seven years are routine in major data center hubs, and up to ten in constrained regions. This, not chips or money, is the sector's binding constraint.See all terms in the glossary. Before a facility can draw grid power, the utility must study the impact and approve the connection. In established data center hubs such as Northern Virginia, Dublin, Singapore and Amsterdam, waits of four to seven years are now routine, and in constrained regions the wait to energisation has been quoted at up to ten years.

The consequence is visible. Industry estimates suggest that 30% to 50% of US data center projects planned for 2026 will be delayed or cancelled, overwhelmingly for power reasons rather than demand reasons. That range is an estimate rather than a measured figure, and it is one of the softer numbers in this article, so treat it as an order of magnitude. Microsoft's commercial remaining performance obligations, meaning contracted business it has sold but not yet delivered, stood at $678 billion at the end of fiscal 2026, up 84% year on year, and management has said it expects to remain capacity constrained.

The Workarounds

Because waiting seven years is not an option, operators are going around the grid:

Water

One line that deserves attention despite its small cost. Water appears at only $6 million a year in Epoch AI's model, under 1% of operating cost. Financially it is negligible. Politically it is not. Evaporative cooling consumes large volumes of fresh water, and CareEdge specifically flags water usage efficiency as a constraint for Indian expansion given urban water scarcity. A cost line that is trivial on the income statement can still stop a project from being approved.

---

Part 8: Who Profits From AI Data Centers

We now know what a data center costs and how it earns. The investor's question is different: of every dollar spent, who actually keeps a profit?

Follow one dollar of AI capital expenditure down the chain.

NVIDIA's fiscal 2026 results make the concentration concrete: $215.9 billion of total revenue, up 65%, of which $193.7 billion came from the data center segment, up 68% and roughly 90% of the company. Fourth-quarter GAAP gross margin was 75.0%.

The obvious follow-up is whether that concentration is durable. NVIDIA's position rests on more than the silicon: it rests on CUDA, the software layer that a generation of AI researchers learned on and that most AI code is written against. Switching hardware means rewriting and revalidating software, and that is a switching cost, not a performance gap. We examined how durable that arrangement looks in the NVIDIA analysis.

📌 note

A useful habit: whenever you see a company described as an "AI beneficiary", find its gross margin. If it is under 20%, the company is participating in the volume, not the profit. Both can be good investments, but they are entirely different bets.

tap to close📌

A useful habit: whenever you see a company described as an "AI beneficiary", find its gross margin. If it is under 20%, the company is participating in the volume, not the profit. Both can be good investments, but they are entirely different bets.

For the full layer-by-layer map of which companies operate where in this chain, including the less obvious layers such as cabling, cooling equipment and cybersecurity, see what really powers ChatGPT.

---

Part 9: India's Data Center Buildout, and the Listed Stocks Exposed to It

India is not a small version of the American story. The physics are identical and the financial structure is not.

The Scale, From a Rating Agency Rather Than a Press Release

From CareEdge Ratings, March 2026:

That 1.2 MW per million internet users, against a world average of 5.0, is the single most compelling number in the Indian story. India has achieved digital parity on usage, with 77% mobile penetration and 67% internet penetration against world averages of 70% and 73%, while holding roughly a fifth of the world's average data center capacity relative to its user base.

📌 note

Wireless data usage per Indian user roughly doubled from about 11 GB a month in 2020 to about 22 GB in 2025.

tap to close📌

Wireless data usage per Indian user roughly doubled from about 11 GB a month in 2020 to about 22 GB in 2025.

The gap between how much Indians use the internet and how much local infrastructure serves them is the entire investment thesis, and it does not depend on AI at all. AI accelerates it.

One Number That Cuts the Other Way

Before the enthusiasm runs away, hold this next to it. CareEdge, citing OECD data, puts cumulative global AI investment at close to $1 trillion between 2020 and 2025, with the United States and China accounting for roughly 75% of it. India's cumulative figure over the same period is around $20 billion, running at roughly $2 billion to $2.5 billion a year. On AI compute specifically, the gap is starker still: about $153.9 billion globally against roughly $0.5 billion in India.

The Indian government's AI Mission commits ₹10,372 crore over five years, which is meaningful policy but small against those numbers.

Why the Economics Differ

The Companies, Honestly Categorised

This is where Indian investors need the most care, because the theme is popular and the pure-play options are genuinely scarce.

Listed, with direct data center operations:

• Anant Raj (NSE: ANANTRAJ). A real estate developer that pivoted into data centers, with roughly 28 MW operational across Manesar and Panchkula, and a stated target of around ₹9,000 crore of data center revenue by FY32. The clearest listed operator play, though it remains a property company with a data center division rather than the reverse.
• Sify Technologies (NASDAQ: SIFY). Listed in the US rather than India. Reported about 138 MW of operational data center capacity, with new Delhi and Chennai campuses of 26 MW design capacity each and two Mumbai facilities of 52 MW each under construction.
• E2E Networks (NSE: E2E). A GPU cloud provider, the closest Indian equivalent to the neocloud model, and therefore carrying neocloud economics including the utilisation sensitivity described in Part 5.

Listed, with data center exposure inside a much larger business:

• Bharti Airtel, which owns Nxtra Data, one of India's largest operators. Data centers are a small fraction of Airtel's value.
• Adani Enterprises, through the AdaniConneX joint venture with EdgeConneX, which announced a large AI data center campus and associated green energy infrastructure at Visakhapatnam in partnership with Google.
• Reliance Industries, through Jio's infrastructure ambitions.

Listed suppliers and enablers:

• Netweb Technologies (NSE: NETWEB), a domestic server and high-performance computing manufacturer that has grown revenue rapidly on the back of domestic AI and HPC demand. Check the latest filed results rather than any growth rate quoted second-hand, including here.
• RailTel, Techno Electric & Engineering, KEC International and similar power and connectivity infrastructure firms that win data center contracts among other work.
• Transformer, switchgear, cable and cooling equipment manufacturers, which supply the 40% of Indian data center cost that is electrical systems and the 20% that is cooling.

Large, unlisted, and the ones to watch:

• CtrlS, Yotta, Nxtra and Princeton Digital Group operate at meaningful scale without listed equity. An IPO pipeline is widely discussed for 2026 and 2027.

Part 10: Is AI a Bubble? The Bull and Bear Cases

The honest answer is that credible, well-informed people disagree, and the evidence genuinely points both ways. Here is each case at full strength.

Part 11: What to Actually Watch as an Investor

Here is the practical output. These are observable, checkable signals, most of them free, that resolve the debate above over time.

In the financial statements

In the operating data

In the market

On how to translate all of this into a view on what a business is worth, rather than merely how it is doing, our guide to valuation covers the specific difficulty of valuing capital-intensive businesses whose earnings depend heavily on depreciation assumptions.

---

Sources

Every figure in this article is traceable. Where you can check the primary document yourself, here it is.

Cost and unit economics

• Epoch AI, total cost of ownership of a one-gigawatt AI data center. The source for the $37.9 billion capex breakdown, the $907 million operating cost, the $8.5 billion annualised total cost of ownership, and the three-year versus seven-year sensitivity. Assumptions stated there: GB200 NVL72 systems, 8.34 cents per kWh, five-year IT and fourteen-year facility lives. Epoch's own PUE assumption of 1.14 traces to Lawrence Berkeley National Laboratory's research on AI data center energy use, and its 71% utilisation assumption is an average of four estimates collected by researcher Tyler Norris, both credited on Epoch's page rather than original to Epoch.
• JLL 2026 Global Data Center Outlook. The $11.3 million per megawatt global construction benchmark, up from $7.7 million in 2020.
• Epoch AI, LLM inference price trends. The 9x to 900x per year range of price declines across benchmarks, and the 40x per year figure for GPT-4 level performance on PhD-level science questions.
• Silicon Analysts, GPU rental price decomposition. The $1.45 per GPU-hour fully burdened H100 cost.
• Nebius published pricing and CoreWeave pricing. The advertised GPU-hour rates, as at August 2026.
• CoreWeave's own announcement of its Core Scientific acquisition. Confirms a neocloud typically leases its real estate rather than owning it, the over $10 billion of cumulative lease overhead being eliminated, and the "verticalizing the ownership" quote.
• Karnataka Electricity Regulatory Commission, tariff orders and Asianet Newsable on the FY2026-27 true-up order. BESCOM's base domestic rate and the true-up surcharge.
• PG&E residential rate plan pricing, effective March 2026. The Tier 1 and Tier 2 per-kWh prices for San Francisco.
• NVIDIA's DGX GB200 NVL72 hardware documentation. The 72-GPU, 120 kilowatt rack specification. TRG Datacenters on H100 power draw and AMD's EPYC server processor line. Per-chip wattage for the CPU-versus-GPU comparison.

Company filings and disclosures

• NVIDIA fiscal 2026 fourth quarter and full year results. Revenue of $215.9 billion, data center revenue of $193.7 billion, Q4 GAAP gross margin of 75.0%.
• Microsoft fiscal 2026 fourth quarter results. Commercial remaining performance obligations of $678 billion up 84%, Microsoft Cloud revenue of $59.3 billion, Azure past $100 billion annually, and $115.9 billion of additions to property and equipment.
• Amazon quarterly filing, useful life change. The reduction of a subset of servers and networking equipment from six years to five, effective January 2025.
• The Register on Microsoft's 2022 server life extension. The move from four years to six.

Energy

• IEA, Energy and AI. The 945 TWh 2030 projection, AI's rising share of data center power, and regional concentration.
• Carbon Brief's charts on the IEA data. The comparison showing data centers adding less new demand to 2030 than electric vehicles or air conditioning.
• Gartner, June 2026. The 565 TWh 2026 forecast, up 26% from 447 TWh, and the 132 GW power demand figure.

India

• CareEdge Ratings, India's Data Centre Capacity to Grow 4x to ~4 GW by 2030, March 2026. Capacity trajectory, the ₹1.5 lakh crore investment estimate, absorption above 90%, EBITDA margins of 40% to 43%, PUE of 1.5 to 1.9, the 70% to 80% debt financing profile, and the cost comparison against China and the US.

Individual projects and Indian companies

• Google's announcement of its $15 billion India AI hub and Adani's release on the Visakhapatnam campus.
• CNBC on the scale and financing of US megaprojects, including Meta's Hyperion campus.
• Anant Raj's expansion to 28 MW and its FY32 revenue target.
• Sify Technologies quarterly earnings call. Operational capacity and campus pipeline.
• Airtel's own press release on the Nxtra-AdaniConneX-Google Visakhapatnam partnership and Google Cloud's groundbreaking announcement. Named landlord and tenant on the same deal, confirmed from both sides.
• AWS's own Direct Connect location announcements, which name specific Equinix buildings by ID. Digital Realty's newsroom on Oracle Cloud Infrastructure access. CtrlS's own website on serving "5 of 7" hyperscalers without naming them, the more typical case.

The bubble debate

• Bloomberg on AI circular deals. The circular financing arrangements.
• MIT's GenAI Divide study, as reported by Fortune. The finding that roughly 95% of enterprise AI pilots showed no measurable profit impact.
• CNBC on the GPU depreciation debate.
• CNBC on hyperscaler capex scrutiny. Context for the 2026 capex guidance range.

---

Frequently Asked Questions

How much does an AI data center cost to build?

Roughly $38 million per megawatt all-in for a modern US facility. Epoch AI's teardown of a 1 gigawatt AI data center puts total upfront capital at about $37.9 billion: $21.2 billion of servers and GPUs, $11.4 billion of facility and power infrastructure, $4.9 billion of networking, and under $350 million of land and utility works combined. The shell and electrical systems alone run about $11 million to $12 million per megawatt, rising to $15 million to $20 million for high-density AI-ready builds. Indian construction costs run roughly 30% to 40% below US levels.

How much electricity does an AI data center use?

A 1 GW facility running at 71% utilisation spends about $594 million a year on power, which is 65% of its operating costs. Globally, Gartner forecasts data center electricity consumption at 565 TWh in 2026, up 26% from 447 TWh in 2025, with the IEA projecting roughly 945 TWh by 2030, or about 3% of world electricity. AI's share of data center power is expected to rise from 5% to 15% recently to 35% to 50% by 2030.

Is the electricity bill the biggest cost of running an AI data center?

No, and this is the most commonly misunderstood point. On an annualised total cost of ownership basis of roughly $8.5 billion a year, power is about 12% while hardware depreciation is about 60%. Energy dominates the headlines because it is visible and politically charged, but the chips wearing out is the far larger economic cost.

Why does GPU depreciation matter so much to AI data center economics?

Because the assumed useful life of the hardware swings the annual cost by about 70%. The same 1 GW facility costs $12 billion a year at a three-year hardware life, $8.5 billion at five years, and $7 billion at seven. Microsoft and Alphabet extended server and network useful lives from four to six years, while Amazon shortened a subset to five, citing the pace of AI innovation. The largest buyers of this hardware publicly disagree about how long it lasts.

Is AI a bubble?

The evidence genuinely points both ways and this article deliberately reaches no verdict. The bear case cites 2026 hyperscaler capex of roughly $660 billion to $725 billion approaching all of their operating cash flow, possibly understated depreciation, upward of $800 billion of circular financing arrangements, and MIT's finding that about 95% of enterprise AI pilots showed no measurable profit impact. The bull case cites Microsoft's $678 billion of contracted, undelivered business, Azure passing $100 billion of annual revenue, NVIDIA data center revenue of $193.7 billion at 75% gross margin, and Indian colocation absorption above 90% for four years. The narrower answerable question is whether the revenue arrives before the hardware becomes obsolete.

Which companies actually profit from AI data centers?

Profit concentrates where there is no substitute. NVIDIA reported roughly 75% gross margin in Q4 FY2026, SK hynix holds above 70% of the high-bandwidth memory market at operating margins near 47%, and TSMC's advanced packaging reportedly earns near 80%. Further down the chain, chip assembly and test firms earn 15% to 22% and server assemblers earn single digits for genuinely hard work. Colocation landlords earn 40% to 43% EBITDA in India but carry heavy debt.

Which data center stocks are listed in India?

India has very few pure-plays. Listed operators include Anant Raj (about 28 MW operational), Sify Technologies (NASDAQ listed, about 138 MW) and E2E Networks (GPU cloud). Larger exposure sits inside conglomerates: Bharti Airtel owns Nxtra, Adani Enterprises holds the AdaniConneX joint venture, and Reliance operates through Jio. Suppliers such as Netweb Technologies, RailTel and Techno Electric benefit indirectly. Major operators CtrlS, Yotta, Nxtra and Princeton Digital Group remain unlisted. Before treating any of these as a pure play, check what share of revenue and operating profit actually comes from data centers.

---

Key Takeaways

1. A data center is a chip purchase wearing a real estate costume. Servers and networking are about 69% of the upfront cost of a large AI facility. Land is under 1%. Judge these projects as technology investments with a short asset life, not as property.

2. Electricity is about 12% of the true annual cost. Hardware depreciation is about 60%. Energy dominates the coverage because it is visible and political. Depreciation dominates the economics.

3. One accounting assumption swings annual costs by roughly 70%. Moving assumed hardware life from three years to seven takes the annualised cost of a 1 GW facility from $12 billion to $7 billion. Microsoft and Alphabet extended lives to six years. Amazon shortened a subset to five. Nobody knows yet who is right.

4. Utilisation is the whole business for GPU renters. An H100 costs roughly $1.45 per GPU-hour at 70% utilisation and above $2.90 at 30%, because the hardware cost is fixed whether the chip is busy or idle.

5. The price of AI capability falls about 90% every two years while the asset depreciates over five. Those two clocks running at different speeds is the central risk in the sector, and whether volume growth outruns price collapse is genuinely unresolved.

6. The bottleneck is transformers and interconnection queues, not money or chips. Lead times of two to five years and grid connection queues of four to ten years mean 30% to 50% of US projects planned for 2026 face delay or cancellation, for supply reasons rather than demand reasons.

7. The profit concentrates where there is no substitute. Accelerators at roughly 75% gross margin, high-bandwidth memory near 47% operating margin, advanced packaging reportedly near 80%. Server assembly earns single digits for genuinely hard work. In India, most listed exposure is indirect, so check what percentage of revenue and profit actually comes from data centers before treating a stock as a pure play.