AI Looks Like Software. It Costs Like a Factory.

An answer arrives on a screen weighing nothing. It took a fraction of a second and it looks like every other thing a computer has ever done for free.

Somewhere else, a building drew power to produce it. Chips ran, heat left through a cooling tower, and a meter moved.

That gap — between where the product appears and where the cost occurs — is the whole reason the largest technology companies on earth are now spending like heavy industry. It is also why their quarterly reports have started to read less like software statements and more like something a steel mill would file.

The Copy That Was Free

Traditional software has one defining economic feature: the second copy costs almost nothing.

Building the first version is expensive. After that, distributing it to one more user is close to free. That single property is why software companies could grow revenue far faster than costs for decades, and why the market learned to pay for them the way it does.

AI does not inherit that property.

Every answer is new work. Summarising a document, generating an image, reviewing a block of code, answering a search query, deciding an autonomous agent's next step — each one requires computation that has not happened before and will not be reused.

The millionth user of a traditional software product added almost no cost. The millionth user of an AI product adds a million more computations.

The screen still looks like software. The cost structure underneath has moved somewhere else.

Three Clocks

Inside an AI business, three things run at different speeds, and confusing them is where most of the argument about this spending comes from.

Revenue follows paid work. It moves when a customer does something they are being billed for.

Operating cost follows computation. It moves when anyone uses the system, paying or not — a free tier, an internal feature, a background process, a trial.

Depreciation follows time. It moves whether or not a single customer arrives. A chip loses economic value sitting idle. A building consumes power and maintenance with half its racks empty. Capacity secured two years early depreciates for two years before it earns anything.

A traditional software company has essentially one clock. An AI company has three, and only the first one is under the customer's control.

When computation and depreciation grow faster than paid work, the factory stops being an advantage and becomes an obligation. That is the entire question underneath the current spending, and it does not require a view on any company to state.

The Unit Gets Cheaper and the Bill Gets Bigger

The most common framing — more usage means more cost — is directionally right and misses the more interesting half.

The cost of a single unit is collapsing.

Stanford's 2025 AI Index tracked the price of querying a model at roughly GPT-3.5-level performance. In November 2022 it cost about $20.00 per million tokens. By October 2024 it was $0.07 — a reduction of more than 280 times in under two years. The same report found hardware costs falling around 30% a year and energy efficiency improving about 40% a year.

That is one of the steepest cost declines in the history of any industrial input.

And total spending went up anyway.

The reason is not complicated. When something becomes dramatically cheaper, organisations do not bank the savings. They use more of it. Cheaper computation means AI gets put into more products, given longer context to read, allowed to reason for longer, asked to generate several candidate answers, and then asked to have another model check the result.

Autonomous agents removed the last natural ceiling. A person can type maybe ten prompts in a working session — human attention was the throttle. An agent handling one task may call a model hundreds of times: read the files, write the code, run it, fail, retry, ask a second model to review. Nobody is watching, and nothing is waiting for a human to type.

The individual token is not getting more expensive. The number of tokens is growing faster than the price is falling. Every projection in this story sits on that one sentence.

The Bottleneck Moved From Chips to Power

For two years the scarce thing was accelerators. It is becoming electricity.

The International Energy Agency reported that data-centre electricity demand rose 17% in 2025, with AI-focused facilities growing far faster — about 50% in that year alone. Looking to 2030, the IEA expects total data-centre consumption to roughly double, while power use at AI-focused data centres triples. AI accounted for something like 5–15% of data-centre power in recent years; the IEA's range for 2030 is 35–50%.

In the United States specifically, Lawrence Berkeley National Laboratory estimates data centres could reach 11.8% of total national electricity by 2030, within a range of 9.5% to 15.3%.

The practical consequence is a change in what is hard to get. More accelerators can be ordered. A grid connection cannot. Transformers, turbines, transmission capacity and permits run on timelines measured in years and set by people who are not in the technology business.

Which reframes what these companies are actually buying. A data centre announcement reads like a purchase of servers. A meaningful part of it is a reservation — land, power, cooling, permits, a place in a future grid — made years before the equipment that will sit there has been designed.

A receding row of pale ceramic insulator stacks on steel substation gantries at night, lit by a single sodium lamp

Two Ways to Answer the Same Question

Microsoft and Alphabet are both spending at a scale that has no precedent in software, and they are answering a different question with it.

In the quarter ended June 2026, Microsoft reported $35.8 billion in cash paid for property and equipment, with total capital expenditure around $41 billion once finance leases are included. It brought 88 data centres online across the fiscal year. Alphabet spent $44.9 billion in its June quarter, roughly double the year before, and that exceeded its operating cash flow — producing negative free cash flow of about $5.9 billion. It then raised full-year 2026 capital expenditure guidance to $195–205 billion, up from $180–190 billion.

The interesting difference is not the amount. It is what each is trying to fill.

Microsoft's capacity is largely rented to outside customers who are already paying it for something else — identity, productivity software, developer tools, security, cloud contracts. The distance between a GPU cycle and an invoice is short, because the billing relationship already exists.

Alphabet's capacity has to be divided. The same accelerators serve external cloud customers, consumer products, search results, advertising models, recommendation systems, and internal training. Some of those produce a direct invoice. Some improve a product that earns money a completely different way. Some are research.

Neither arrangement is better on its face. One converts to revenue more legibly; the other can earn across several businesses from a single build. What they share is the third clock — and the third clock does not care which strategy is more elegant.

A quarter is a snapshot, not a verdict. These figures describe one three-month period, under accounting definitions that differ between the two companies, in the middle of a build that both have said will run for years.

The Clock You Can Change With a Sentence

One detail from Microsoft's disclosures is worth pulling out, because it shows how much of this is accounting rather than physics.

Effective at the start of fiscal 2027, Microsoft is extending the estimated useful life of its data centres and office buildings from 15 years to 25 years. Management described it as reflecting operating history and expected use, and said the direct effect on operating income would be small.

The larger effect is elsewhere. The change shifts more future data-centre leases from finance leases to operating leases — and finance leases count as capital expenditure while operating leases do not. Reported capital expenditure for calendar 2026 therefore moves from roughly $190 billion to roughly $175 billion, with the underlying investment plan unchanged.

Nothing about the buildings changed. Nothing about the spending changed. A $15 billion difference in the most-quoted number in this entire story came from a revised estimate and a lease classification.

This is not an accusation. Extending a useful life is defensible if the buildings genuinely last longer, and buildings plausibly do outlive several generations of the equipment inside them. The point is narrower and more useful: the headline capital expenditure figure is not a physical measurement. Anyone comparing these companies on that single number is comparing two accounting presentations at least as much as two strategies.

What is harder to move is the split between long-lived assets and short-lived ones. Microsoft said roughly two-thirds of its quarterly spending went to shorter-lived assets — mainly CPUs and GPUs. Land and steel can be depreciated over decades. Accelerators cannot.

What This Looks Like From Korea

There is a version of this story that reads completely differently from here, and it is worth naming because it changes which risk a Korean saver is actually carrying.

Most coverage written elsewhere treats AI infrastructure spending as cost — capital leaving, margins compressing, free cash flow turning negative. From Korea, a large share of that same spending is revenue. It arrives as orders for high-bandwidth memory, and it lands in the companies that dominate the local index.

The same event, read from two sides of the invoice.

That is a genuine advantage in understanding, and it comes with a specific hazard. When the factory boom is your customer, the temptation is to treat it as a structural change rather than a cycle. Memory has never behaved that way. High margins invite capacity, capacity arrives late and all at once, and the wheel turns — the pattern underneath what an ADR listing does and does not tell you.

There is a second layer. Because a small number of names carry an outsized share of the Korean index, an investor here holding a broad domestic fund may already have concentrated exposure to AI capital spending without having chosen it — the problem of an index that is doing less diversifying than its name suggests.

So the local question is not whether AI spending continues. It is whether you know how much of your savings is already positioned as a supplier to it.

Where This Stops Being Useful

It does not say whether this spending will pay off. Three clocks describe the structure. Whether revenue outruns computation and depreciation over a decade is not knowable from any quarter, including this one.

It does not pick a company. Microsoft's conversion is more legible today. Legibility is not the same as return, and Alphabet's approach earns across several businesses in ways a single quarter cannot capture. Both are still building.

And the figures here have a shelf life. Quarterly results, capital expenditure guidance and lease classifications get restated, revised and re-presented. The institutional projections — inference costs, electricity demand — move slowly. The company numbers do not. Read them as of the dates attached, not as standing facts.

Where You Are Standing

You are not choosing between two data-centre strategies. You are probably holding both companies inside a fund, alongside the suppliers, without having selected any of it.

Watch conversion, not spending. The headline capital expenditure number is partly an accounting presentation. What it produced — revenue from new capacity, utilisation of what was built, how quickly capacity fills — is the part that says whether the factory works.

Ask who the customers are. Infrastructure demand concentrated among a handful of AI laboratories can move to another provider or be renegotiated. Demand spread across banks, hospitals, governments and ordinary corporate work grows more slowly and tends to stay.

Separate the two kinds of asset. Land and buildings depreciate over decades. Accelerators do not. A single capital expenditure figure hides that split, and the split is most of the risk.

And treat the leaders as temporary. Whichever company looks structurally advantaged right now, that impression has a poor historical record — the throne changes hands more reliably than anything else in this market. The infrastructure may outlast the company that built it. That has happened before, with railways and with fibre.

A single open laptop on a bare wooden desk in a dark room, its blank screen the only light in the frame

A Closing Note

The early part of this competition was easy to watch. A model was released, a benchmark score improved, people compared the answers.

The next part happens in power contracts, lease classifications, depreciation schedules and billing systems. None of it is quotable, and all of it decides whether the most expensive construction project in corporate history becomes a moat or a warehouse of aging chips.

The question that eventually matters is not how large the model is or how much was spent. It is narrower and, for once, arithmetic: how much customer value does one dollar of electricity and depreciation produce?

An answer on a screen weighs nothing. Behind it, the meter has not stopped once.

Data as of 2026-08-11. Sources: Stanford HAI AI Index Report 2025 (inference cost, hardware and energy-efficiency trends); International Energy Agency (data-centre electricity demand and 2030 projections); Lawrence Berkeley National Laboratory, US data-centre energy usage; Microsoft FY2026 Q4 and Alphabet Q2 2026 results and guidance as reported. Company figures reflect one reporting period under differing definitions and are not directly comparable. This post is observation, not investment advice.

SK Hynix's ADR Listing — The Split That Isn't a Split

When Two Companies Are Half Your Index

Microsoft's $190 Billion Capex — The Bet Nobody Is Sizing

Alphabet's Q1 2026 — A Quiet Bet on Becoming Furniture

Visuals on this post are AI-generated. The author works with AI as a research and drafting assistant; topics, judgments, and final edits are the author's own. This post is observation, not investment advice. See full Disclaimer for details.

Search This Blog

IJin Insight