The Bill for Not Falling Behind: The Capital Half-Life of AI Hardware

The Bill for Not Falling Behind: The Capital Half-Life of AI Hardware

Tópico: The Bill for Not Falling Behind: The Capital Half-Life of AI Hardware
Categoria: Tutoriais | Programação & Tecnologia
Idioma Principal: Português (Conteúdo de Tecnologia)

Descrição do Conteúdo / Informações:
-------------------------------------------------------------------------
Prefer reading in Arabic? Read the Arabic version on Substack

Every quarter, the big technology companies tell us how much they spent. They don't tell us how much of that money bought back capacity they already had. The filings we reviewed don't separate the two.

In 2025, Amazon changed its mind about how long some of its servers would last.

It had raised the estimated useful life from five years to six. Then it moved a subset back down to five. The reason it gave in its own filing was blunt: technology was moving faster, "particularly in the area of artificial intelligence and machine learning." Shortly before that, the company had booked roughly $920 million in accelerated depreciation and charges tied to retiring equipment early. (SEC)

At a small company, an accountant would note this and move on. At a company building one of the largest compute fleets on the planet, a single year on the life of a server moves serious money.

And the striking part is that this decision did not come from a company losing customers. In the second quarter of 2026, AWS reported $42.2 billion in revenue and $16.6 billion in operating income, and Amazon said its AI business inside AWS had passed a $25 billion annualized revenue run-rate. (SEC)

So the demand is there, and the revenue is real.

And yet the company's trailing-twelve-month free cash flow fell to negative $7.6 billion. The reason Amazon itself gave: a $66.1 billion increase in purchases of property and equipment, which "primarily reflects investments in artificial intelligence infrastructure." (SEC)

Both of the ready-made stories collapse here — the bubble and the gold mine. AI can generate money and consume capital in the same breath.

Which opens a sharper question. When a company writes the next hundred-billion-dollar check, how much of it adds new computing capacity to the world, and how much of it buys back capacity that is starting to slip?

Obsolescence Reaches the Books

Before Amazon's reversal, the traffic was all going the other way.

Microsoft extended the estimated useful life of its server and network equipment from four years to six, effective with its fiscal 2023, which raised operating income by $3.7 billion and net income by $3.0 billion. (SEC) In January 2023, Alphabet raised the life of its servers from four years to six and certain network equipment from five to six, cutting depreciation expense by $3.9 billion and lifting net income by $3.0 billion. (SEC) And in 2025, Meta extended most of its servers to five and a half years, reducing depreciation by about $2.9 billion. (SEC)1

These extensions may well reflect better hardware management and a genuine ability to keep machines useful for longer. But they also tell you something about the nature of the number. An accounting life is a management estimate, and one extra year does not just change the page — it changes the profit printed on it.

Which is what made Amazon's reversal worth noticing. In the same filing, the company estimated the decision would reduce its 2025 operating income by about $0.7 billion, with accelerated depreciation taking another $0.6 billion. (SEC) When a company shortens the life of its own assets knowing exactly what that will cost it, it isn't flattering anyone.

The pace of technical change had moved out of conference panels and into a line item.

Ninety-One Billion, Undivided

For three decades, software was the light industry: lines written once and sold millions of times, at high margins and enormous scale, with no factory and no smokestack. Then the rule flipped.

Microsoft gives us the clearest window onto the fast-moving part of the new spending.

In four years, its cash outlay for property and equipment jumped from about $23.9 billion to $115.9 billion. (SEC) The same trajectory shows up at Alphabet, Meta and Amazon, though fiscal years and spending definitions differ.2 As for the number the headlines like — roughly $700 billion when you add up the big companies' guidance — that describes plans that can still be revised, not money paid.

Add up what Microsoft's management called capital expenditure across its fiscal 2026 quarters and you reach about $145.3 billion. From the company's own description of the mix, roughly $91.1 billion of it — about 63% — went into short-lived assets, led by processors, GPUs and network gear.3

Ninety-one billion dollars in twelve months, from one company, on equipment whose clock runs faster than the building's.

Timing complicates the arithmetic further. At the end of 2025, Alphabet held about $78.6 billion in assets not yet in service. By June 2026 that figure was $122.8 billion — around 40% of its $304.3 billion in total gross property and equipment, of which technical infrastructure accounts for $247.2 billion. (SEC) A single asset has three separate moments: the day it is paid for, the day it starts working, and the day the next generation starts pressing on its value. Months can separate them.

So you cannot take one year of spending, add five years, and announce the date the replacement bill comes due.

And here stands the question no published document answers: how much of those billions expanded the fleet, and how much replaced what was already there?

That is the missing number. The rest of this essay examines the forces that make it bigger or smaller, until we reach what the figures actually support — which is more than it first appears.

The Old Chip Doesn't Die

The strongest argument against the idea of permanent replacement is simple: a chip doesn't end the day its successor arrives. It moves down a rung.

A processor that has dropped out of frontier training moves to running models after training — inference — or serves smaller models and customers who care more about price than about the latest performance. The market shows this second life is real.

In September 2026, the hourly price of an A100 with 80GB ran from about $0.45 at the bottom of the surplus-capacity marketplaces — where individuals and small operators sell unused cycles at volatile prices — up to $5.03 on Google Cloud. (Thunder Compute)4

That range is the story, not either end of it.

Take a unit bought for $20,000, running 80% of the time over five years. The capital share alone comes to about $0.57 per hour used, before power, cooling, networking and operations. At $2.70 an hour — CoreWeave's on-demand price — revenue covers that easily. At $1.09, the cheapest on-demand price published under a provider's name, it covers it with comfortable margin. At the floor of the surplus market, where offers fall to $0.45, it doesn't cover it at all.

Now change one variable. If the acquisition price were $8,000 instead of $20,000, the capital share drops to about $0.23 an hour, and the result flips at nearly all of those prices.

These are illustrative figures, not a company's books; the large fleets don't publish what they actually pay. But they are enough to separate two things people routinely merge: that a machine still runs is one fact, that it has earned back its price is another. That it will fund its own successor is a third, and it doesn't follow from the first two.

There is a finer distinction too. A machine can fail to recover its original cost and still be worth running today. The purchase price is spent and gone; the only live question is whether its revenue covers its power and its floor space.

Except that calculation runs into a last question: what would the company have earned by putting a newer machine in the same spot?

And that spot, it turns out, is not a neutral box.

The Building Is Fine. The Hall Isn't.

When power is the binding constraint, swapping old hardware for something that yields more value per megawatt looks obvious. The obvious move assumes the hall will take the new generation the way a warehouse takes a box.

In one NVIDIA reference design, four DGX H100 systems in a single rack draw about 40.8 kilowatts. A GB200 NVL72 system reaches roughly 120 kilowatts per rack.

Three times the power density in the same footprint.

That jump is not executed by pulling one box out and sliding another in. You cannot cool 120 kilowatts with air. It needs liquid running to the chip, pumps, and heat exchangers. It needs new busbars inside the rack and new power distribution units, and sometimes new feed all the way back to the building.

The result is a paradox that cuts both ways. A hall that cannot host the new generation gives the old hardware a kind of immunity, because keeping it running is cheaper than gutting the room. At the same time, the company may be forced to reinvest in the hall long before the building itself grows old. The concrete may serve a quarter of a century; the hall's electrical and mechanical arteries can reach the end of their economic life before the building is halfway through its own.

What does that cost? Companies don't disclose a separate figure; it disappears into capital project lines. The closest published quantitative signal is a study by STL Partners, commissioned by the cooling company Airedale. It estimates converting cooling to liquid at roughly $2 million per megawatt, against more than $11 million per megawatt for a new facility built liquid-cooled from the start. (STL Partners)

The two numbers do not divide into each other. The first covers cooling alone, while upgrading a hall extends to power, rack density and construction work. The second is a floor, not a point. And the study itself warns that the simple figure excludes revenue lost while the hall is out of service. (STL Partners)

What remains is that facility life is a physical fact, and its price is still hidden. It is the silent difference between a long-lived building and a plant that demands recapitalization before its midpoint.

Who Carries the Risk?

Before we size the next cycle, there's a prior question: who can absorb it at all?

The scale of replacement is unknown. How all of this is financed shows up more clearly — and it reveals that the four companies are not one bloc.

Microsoft is still building out of its own cash. In fiscal 2026 it spent about $115.9 billion on property and equipment against $182.9 billion in operating cash flow. It issued no new debt during the year, and repaid about $3 billion. (SEC)

Meta went to the bond market, raising $24.91 billion net — after a year in which it issued no long-term debt at all. (SEC)

Amazon went to the same market with more weight: roughly $67 billion in long-term debt issuance in the first half of 2026 alone, alongside an accelerating build. (SEC)

Then there is Alphabet, which is the case worth pausing on.

In June 2026 the company raised about $49.6 billion net from common stock and mandatory convertible preferred, alongside $56.2 billion in debt issuance in the first half. (SEC)5 A company that used to buy its own shares back is now selling new ones to fund infrastructure — and it spent nothing at all on buybacks in the first half of 2026.

That shift alone says more about the size of the bill than its figures do. The four are distributing the weight of risk to different parties: Microsoft pays out of current profit, Meta and Amazon move part of the load to creditors, and Alphabet moves it to shareholders.

And the larger replacement's share of spending becomes, the more that distinction matters. Financing new expansion is one thing; financing a recurring cycle to hold your position is another entirely. The first buys growth. The second buys staying put.

On top of current financing, there is another bill already booked to the future.

Alphabet's filings show $85.2 billion in undiscounted lease payments for leases that have not yet commenced. (SEC) Meta disclosed in June $349.3 billion in non-cancelable contractual obligations, of which $279 billion are operating and finance leases not yet commenced, then added roughly $68 billion more in contracts after quarter end. (SEC)

The next build round is no longer a plan to be reviewed in a meeting. It is an obligation that comes due whatever the market does.

The Second Layer of the Bill

There is a layer of this bill that does not appear in the four companies' capital spending at all: the model labs themselves. They don't own most of the data centers they run in, but they have started reserving compute under contracts measured in tens and hundreds of billions.

OpenAI closed a $122 billion funding round on 31 March 2026 at a valuation near $852 billion. (OpenAI) (CNBC) Press coverage of the round puts its revenue at roughly $2 billion a month.6 In return it has spread its suppliers wide: a $250 billion commitment to Azure services (Microsoft), an AWS agreement that started at $38 billion and then expanded by another $100 billion over eight years, plus Oracle, CoreWeave and Google Cloud.

Anthropic follows the same path even more plainly. In April 2026 it committed more than $100 billion over ten years to AWS services in exchange for up to five gigawatts of capacity, and says it is already running more than a million Trainium2 chips. (Anthropic) The same month it signed for multiple gigawatts of next-generation TPU capacity with Google and Broadcom, starting in 2027. (Anthropic) Then in May it raised $65 billion at a $965 billion valuation, announcing that its run-rate revenue had crossed $47 billion. (Anthropic)

The picture is completed by noting where the capital behind these commitments comes from. The companies building the clouds are often the ones funding their largest customers. Amazon committed $50 billion to OpenAI's latest round, $15 billion of it up front and the rest conditional on later milestones. It also invested $5 billion in Anthropic with a promise of up to $20 billion more. (Anthropic) NVIDIA invested $2 billion in CoreWeave, with a contractual commitment to absorb its unsold capacity up to a $6.3 billion cap.7 And Microsoft holds an ownership stake in OpenAI of roughly 27%. (Microsoft)

None of this proves phantom revenue circling a closed loop. The relationships are not carbon copies of one another, and the available figures do not measure how much of the labs' cloud spending originated in the pocket of the company selling them the cloud.

But it does establish something narrower and sufficient: part of the demand that justifies the build is not independent demand arising from the broader market. And the practical question it leaves is this — who is left holding the asset, the debt and the lease if the customer is late?

Here the bill splits into two different bets. The infrastructure owner is betting the asset stays economic for its full life. The model lab is betting that its future revenue will justify the compute it reserved in advance, before it knows what the generations consuming that compute will look like. If either side is wrong, the risk does not disappear. It moves somewhere else along the chain.

Why They Accept It

Whoever carries the risk needs a reason to carry it. Why do companies this size board a train this expensive?

The answer is a cold defensive calculation. AI is not being sold today as a standalone chatbot; it is being embedded inside software that runs the working lives of hundreds of millions of people. Microsoft has said Microsoft 365 Copilot passed 30 million paid seats, that GitHub Copilot reached 50 million developers, and describes its work platform as connected to more than 17 exabytes of enterprise data, mail and meetings. (Microsoft) (Microsoft) Google has said its AI-powered search mode passed a billion monthly active users, and that AI Overviews serve more than 2.5 billion. (blog.google)

A company that already owns a platform can absorb weak direct returns for years, because what it is buying is not only a new revenue line. It is cover for an older and far larger business that it cannot afford to leave exposed.

There is a second attempt to slow the cycle itself: designing the chip in-house. When Amazon develops Trainium (SEC) and Google expands its TPU deployments — Anthropic has announced an expansion of up to a million of them (Anthropic) — the goal is not just to escape NVIDIA's margins. An operator that designs its own silicon gains some control over how fast its fleet ages, and over what the next generation costs.

And this can work in both directions. A cheaper chip makes every cycle lighter; a large performance jump pushes the previous generation out faster. The direction isn't determined by the name on the chip. It shows up in price, performance and actual use.

The State Pulls Up a Chair

Once the requirements for power, concrete and liquidity reached this scale, the race moved out of boardrooms and into politics.

In July 2025, America's AI Action Plan described building data centers and power networks as a condition of American "dominance" against the speed of China's buildout. (The White House) In April 2026, authority under the Defense Production Act of 1950 was used to classify large-scale energy infrastructure as a national-security resource, opening the way to early financing and exceptional treatment. (The White House)

Then the bill reached ordinary pockets. As worries mounted that household electricity bills would rise because of data center consumption, the government sponsored the Ratepayer Protection Pledge in March 2026, then expanded it in July. Seven companies signed it, among them Amazon, Google, Meta and Microsoft, committing to bear the cost of the additional generation and transmission their own projects require. (The White House) (The White House)

The pledge is voluntary and not legally binding, and the signatories are the same companies this essay is about. Which is precisely its significance: the cost of building this era has reached a point where deciding who pays the final bill is a political question argued at the White House.

What We Actually Know

By this point the picture is clear enough: short-lived spending is enormous, obsolescence has entered the books, and the hall itself may need recapitalizing. Working the other way, the secondary market and custom silicon extend the economic life of capacity.

The decisive number the filings don't separate is the ratio of growth to replacement.

But the direction of that ratio can be read.

The Wave Hasn't Arrived

The equipment leaving service this year was not bought this year. Assets now reaching accounting lives of five to six years date mostly from before the current spending surge.

And even if we take 2022 as a deliberately late reference point, Microsoft's cash outlay on property and equipment was only $23.9 billion, against $115.9 billion in 2026.

The pressure we see today comes from a fleet far smaller than the one now being built. The large wave of assets built during the AI surge has not yet reached the usual replacement years.

What this does not allow is a precise estimate of today's replacement share. A single year's retirements don't come from a single purchase year, the asset mix has changed, some equipment retires before its accounting life and some after, and a large part of today's spending has not entered service at all.

But the direction doesn't need that precision: the mass of assets heading for a second cycle is growing, while many future build rounds have already become contractual obligations.

The Real Race

One objection remains, and it is the smartest thing that can be said against everything above.

When you replace an A100 with a newer-generation system, you spend money classified as "replacement," but you do not get the same capacity back. You get a multiple of it. The dollar is replacement; the computing power that results is expansion.

Which reveals that the word "replacement" has been misleading us the whole way. The right question is not how many dollars go into renewing the fleet. It is this:

Does computing capacity per dollar — after the cost of the server, the power and the facility retrofit — improve fast enough to offset the growth in the mass of assets coming due for replacement?

That is two things racing, not one unknown. And every chapter of this essay was in fact a force pushing one of the runners.

The secondary market for older hardware delays the replacement date. Custom silicon — Trainium and TPU — tries to lower the price of new capacity. Both work in the dollar's favor.

Against them, the fleet grows, and so does the mass coming due for renewal. And a hall that needs liquid cooling and new power distribution makes replacing a processor drag part of the building along with it, raising the price of new capacity instead of lowering it.

Then financing decides who survives a bad round, and demand and revenue decide whether the round was worth running at all.

Six forces, one equation, and a result that is not yet settled.

The Bill for Not Falling Behind

Which curve gets there first? Published disclosures do not answer that today.

But what the companies are doing is written in front of us. Designing their own chips, widening their funding base, reserving data centers that have not begun operating, negotiating over the cost of electricity: different tools performing different jobs — lowering the cost of the next tranche of capacity, securing it early, and spreading the risk of financing it across multiple parties.

Taken together, they show companies preparing for a capital cycle lasting years, not for a single buying round.

That is what the numbers support: an industry that has entered a race between two curves, and is acting in public like one that knows the race is close.

Every quarter, the companies tell us how much they spent to move forward.

What they have not yet written on a line of its own is how much they will pay not to fall behind.

•
Useful-life changes for Microsoft and Alphabet servers were matched against original filings; the effect of Meta's extension on depreciation ($2.9 billion) rests on a secondary source. ↩

•
Cash figures for property and equipment do not add into a single valid total: Microsoft's fiscal year ends in June, and definitions of capital expenditure differ between companies. The trend within each company is what the argument rests on. Alphabet went from about $31.5 billion in 2022 to $91.4 billion in 2025; Meta from $31.2 billion to $69.7 billion; Amazon from $58.3 billion to $128.3 billion over roughly the same period. ↩

•
Microsoft's management measure of $145.3 billion includes finance leases, so it does not carry over directly to the $115.9 billion cash flow figure. ↩

•
September 2026 prices per unit-hour, by provider name: Thunder Compute at $1.09, Runpod at $1.59, Crusoe at $2.00, CoreWeave at $2.70, Lambda at $2.79, AWS at $3.43, and Google Cloud at $5.03. On the Vast.ai surplus-capacity marketplace, individual providers' offers range from $0.45 to $0.99 — which is not an average of institutional contracts and does not guarantee availability. (Thunder Compute) ↩

•
Comparing financing across these companies is a comparison of direction, not of ratios: Microsoft here is a full fiscal year, while the other three are half-year data in the places cited. ↩

•
OpenAI's own page for the round did not open for us during review (the server refused access), so the round's figures were matched against consistent coverage from Bloomberg, CNBC and SiliconANGLE, and the company page's title matches the round's value. The monthly revenue figure comes from that coverage, not from a disclosure we read ourselves. ↩

•
NVIDIA's investment in CoreWeave and the capacity-absorption cap come from the two companies' announcements and consistent press coverage; we did not match them against a regulatory filing. ↩