A 100-Year Bond. A Five-Year Machine.一百年的债，五年的机器

Note: Educational content only — research, not investment advice. The author is not a registered investment adviser.

Two dates

In 2026 Alphabet sold sterling bonds with maturities running from 2029 to 2126. One tranche is a century bond: 6.125% Senior Notes due 2126, listed in its quarterly filing.

In the same year, Microsoft's annual report put the useful life of its servers and network equipment at two to six years. Amazon had already shortened the life of some of its servers from six years to five, and its annual report gave the reason: artificial intelligence is speeding up technological change. Microsoft's finance chief has said roughly two-thirds of its capital spending goes to short-lived assets like GPUs and CPUs.

Money is fungible. No bond is tied to a particular server. But put those two dates side by side, 2126 and roughly 2031, and you are looking at the most important question about the AI boom that almost nobody asks out loud. Not "is AI real?" It is. The question is whether the way it is being paid for matches how long the things it buys will last.

Here is the number that made us start asking. In the first half of 2026, the four biggest spenders on AI infrastructure (Microsoft, Alphabet, Amazon and Meta) spent about 91 cents of every dollar of operating cash flow on capital expenditure: chips, servers, data centers. A year earlier it was 66 cents. In the second quarter alone it was 96.

Our thesis. The technology is real, and the companies at the center are still strong. What has changed is the shape of the financing. Over the past year the AI build-out has moved from cash to credit, from balance sheets to guarantees, and from short commitments to very long ones, at the same moment the price of money rose to its highest real level since 2008. That combination doesn't tell you when anything breaks, and we won't pretend to time it. It tells you something more specific. If this cycle turns, the damage will not follow who is closest to AI. It will follow who is carrying fast-aging assets on money that may have to be repaid or rolled over before those assets pay for themselves.

A useful line from our research files: whether a boom ends badly isn't a prophecy question; it's an accounting question. So here is the accounting.

1. The cash cushion thinned first

For most of this boom, the biggest spenders paid with their own money. That is what made the "this isn't 2000" argument persuasive: these are some of the most profitable companies in history, funding investment out of cash flow.

That cushion has thinned fast. Combined free cash flow at the four was $40.2 billion in the second quarter of 2025. In the second quarter of 2026 it was $6.7 billion. Alphabet's free cash flow for the quarter was negative, about −$5.9 billion, which CNBC described as the first negative quarter on record. Amazon's trailing twelve-month free cash flow was also negative. Alphabet and Meta bought back no stock in the first half of the year, after buying back $28.3 billion and $22.9 billion a year earlier.

Guidance points the same way. Combined 2026 capital-spending guidance is $720 to $745 billion, about 111% of the four companies' operating cash flow over the past four quarters.

When spending exceeds cash flow, the difference has to come from somewhere. In 2026 at least three of the four issued about $169 billion of new long-term debt between them. Alphabet's long-term debt, at face value, went from $49.1 billion at the end of 2025 to $101.1 billion by the end of June. Alphabet also raised about $50 billion of equity, which, in the framework we use below, actually makes it more stable, not less.

Here is the fair counterpoint, and it's a strong one. These companies started with very little debt. Borrowing more can be read as normalizing an unusually conservative capital structure, not as distress. The four still generated $322.5 billion of operating cash flow in the first half. Alphabet's new debt is rated Aa2/AA+. Nobody serious thinks these balance sheets are about to break.

But this is where it helps to borrow a ruler from the economist Hyman Minsky. He sorted borrowers into three postures. *Hedge* borrowers can pay interest and principal from their cash flow. *Speculative* borrowers can cover the interest but need to roll over the principal. *Ponzi* borrowers, a technical term in his framework and not an accusation, can cover neither and depend on asset prices rising or new money arriving. By Minsky's definitions, the big four are still hedge borrowers. The migration is happening at the margin: in the newer AI cloud providers and the AI labs whose commitments far exceed their own cash flow.

2. Suppliers are becoming guarantors

The margin is where the most important change of 2026 lives, and it is buried in filings few people read.

In an 8-K filed on August 17, Nvidia disclosed residual value guarantees on data centers in Ohio, about 4.25 gigawatts of capacity, leased to an OpenAI affiliate by a landlord called SB Energy. Nvidia's cumulative payment obligation is capped at $105 billion. Payments are not expected to be possible before about 2028; the guarantee ends if OpenAI achieves a "satisfactory credit rating"; and OpenAI has agreed to reimburse Nvidia for anything it pays. Nvidia also has an option to support roughly 3.8 more gigawatts. Its supply and capacity commitments rose from $119 billion to $279 billion, and it has about $36 billion of commitments to buy back cloud capacity from customers.

Others are doing versions of the same thing. Meta has a residual value guarantee of about $28 billion in force on its Hyperion data-center joint venture, and has proposed up to about $13 billion more for a second campus in El Paso. Amazon is investing $50 billion in OpenAI preferred stock, while OpenAI expands its commitment to Amazon's cloud by $100 billion over eight years; Anthropic's commitment to the same cloud grows by more than $100 billion over ten. Microsoft said in January that about 45% of its commercial backlog, then $625 billion, came from OpenAI.

Our research notes compress the pattern into one uncomfortable sentence: your revenue is guaranteed by yourself. A supplier invests in a customer or guarantees its lease; the customer uses that support to buy the supplier's products; the supplier books the sale.

There is a historical rhyme here, and it deserves care. In 2000, telecom equipment makers lent money to their customers so the customers could buy their equipment. Lucent's customer-financing commitments reached $8.1 billion. As its customers weakened, and some, like Winstar in April 2001, went bankrupt, Lucent's provisions for bad debts and customer financing rose from $66 million in fiscal 1999 to $505 million in fiscal 2000 and $2.249 billion in fiscal 2001. Telecom accounted for 55% of US high-yield default dollars in 2001, and defaulted telecom bonds traded around 16 cents on the dollar.

The differences matter. Nvidia's guarantee is conditional, nothing is payable for years, the customer must reimburse, and a data center can be re-let to someone else in a way that a bankrupt startup's custom network often could not. No guarantee has paid out. But the direction of travel is the thing to watch: the companies selling the shovels are now underwriting the people buying them.

3. The earnings mirror

There is a second loop, and it runs through earnings.

In the second quarter of 2026, S&P 500 earnings grew about 52% from a year earlier. Part of that came from non-operating gains: about $98 billion at Alphabet (its filing attributes them mainly to unrealized gains on SpaceX and another private company) and about $53 billion at Amazon (mainly from its investment in Anthropic). According to FactSet, strip out those two companies and the index's growth was 33.8%.

These are legitimate accounting. Fair-value gains are what the rules require. But they create a mirror. Private AI valuations rise; the cloud companies that own stakes book gains; index earnings rise; price-to-earnings ratios look lower; AI stocks look more reasonable; private valuations are supported. One down round in a large private company would run the same loop in reverse.

And the mirror is not a niche concern, because of what the index has become. The ten largest companies make up about 40.1% of the S&P 500. Seven of them sit on the AI capital-spending chain, as chipmakers or as cloud and platform spenders, and together they are about 29.6% of the index. Stocks and mutual funds now make up 34.8% of US household net worth, the highest share in the data series that starts in 1989.

One line from our research files says it plainly: ordinary investors think they're buying the US economy; to a large extent they're buying the AI capital-spending cycle.

4. Two clocks

Every infrastructure boom leaves something behind. The question is what, and for how long.

Use the Bureau of Economic Analysis's depreciation tables as a guide. Railway structures have a service life of about 54 years. Electric power structures, 40 to 45. Communications structures, 40. When the British railway mania peaked in 1847, investment reached about 7.3% of GDP. Roughly a third of the money invested was lost, almost all of it by ordinary shareholders. But because borrowing was capped at about a third of share capital, there were few bankruptcies, and the track was still there decades later, carrying the economy that came next. The fiber laid in the late 1990s also outlasted many of the companies that laid it.

AI is a two-speed build. Part of it is long-lived: data-center buildings (Microsoft now depreciates them over 25 years, Amazon up to 40), substations and grid connections. The other part, and by Microsoft's own account the larger part of spending, is chips and servers with lives of two to six years. If this cycle has to pass through a downturn before its payoff, the buildings and the power lines can wait for it. Most of the chips can't. They will be obsolete before the recovery arrives.

Our files estimate that cutting server lives from six years to four would reduce pretax profit by roughly $40 billion a year on 2026 spending alone. That's our own rough calculation, not a forecast. The point is that depreciation is a slow variable that markets tend to ignore until it isn't.

Then there is a third clock: the physical world. Chips can be added in six to twelve months. Gas turbines and large transformers take three to seven years. PJM, the largest power market in the US, has cleared its capacity auction at the price cap three times in a row ($329.17, $333.44 and $325.00 per megawatt-day). Without the cap, PJM's own simulation would have cleared at $554.72. Its reserve margin is about 14.4% against a 20% target.

So the build-out runs on three clocks at once. Debt measured in decades. Chips measured in years. Power measured in auction cycles and permitting queues. A line from our technology research captures where those clocks collide: the data hall waits for power, and the interest keeps running.

5. The price of money changed

None of this would matter much if money were still cheap. It isn't.

On September 16, 2026, the Federal Reserve raised its target range by a quarter point to 3.75–4.00%, by a 12–0 vote, its first hike since 2023. The statement said: "Productivity growth is strong, and capital investment is robust." The minutes, released in October, said most participants saw another increase as likely appropriate by year-end. The investment boom is now part of the reason money costs more; the minutes also noted that surging AI-related investment is adding to inflation pressure.

The move in long-term rates is unusual in its composition. From February 27 to October 5, the 10-year Treasury yield rose 134 basis points. 123 of those came from real yields (measured by inflation-protected Treasuries); inflation expectations added only 11. This is a real-rate shock, not an inflation scare. The 10-year real yield, at about 2.78%, is in its highest range since November 2008. The 30-year mortgage rate is 7.40%, up from 5.98% in February.

Washington feels it too. In fiscal 2026 the federal government spent $1.052 trillion on net interest, 19.5% of revenue, the highest share since 1940 and more than it spent on defense.

And stock valuations have very little cushion left against this. The S&P 500's forward earnings yield is about 5.25%, against a 10-year Treasury yield of about 5.10%, a gap of 0.15 percentage points. The Shiller CAPE ratio is about 40.7. In the history of the series, the only months above 40.5 have been January 1999 to September 2000, and August to October 2026. Margin debt reached $1.454 trillion in August, up 37% from a year earlier.

Minsky has a line for this combination, from a 1992 paper: if an economy with many speculative borrowers is in an inflationary state, and the authorities try to fight inflation by tightening money, speculative units become Ponzi units. That is the precise risk. Not that AI fails, but that the cost of carrying long commitments rises while the assets bought with them age on a short clock.

6. Fire picks the driest wood

If the cycle turns, who gets hurt first? Popular commentary ranks winners and losers by how close they are to the technology. Our research points to a different rule: fire doesn't pick the people closest to it; it picks the driest wood, the piece that borrowed short and has no runway.

Use two questions for any company, fund or household. Does it need new money to keep going? And is short money funding long assets?

By that test, the order looks roughly like this:

1. Weeks to months: the thin layer. Newer AI cloud providers and AI labs. One listed AI cloud provider, for example, spent about four times its operating cash flow on capital expenditure in the first half of 2026; in the second quarter its interest expense was $640 million against adjusted operating income of $128 million. Its 2031 bonds were reported trading at a yield of about 15.8%.

2. Weeks to months: private credit. US private credit is about $1.4 trillion. Semi-liquid vehicles sold to individuals hold about $425 billion, and in the first quarter of 2026 redemptions from perpetual business development companies exceeded inflows for the first time. Most managers cap withdrawals at about 5% of net asset value per quarter. Gates turn a run into a queue. They don't make the pressure disappear. And software is the largest sector in many private-credit portfolios, which makes AI a two-sided risk: if AI succeeds, software borrowers are hurt; if it stalls, data-center borrowers are.

3. One to two quarters: the supply chain. Chip, memory and equipment suppliers, and export economies built around them.

4. Two to four quarters: construction and jobs. This boom is capital-heavy and labor-light. Data-center construction is up 73.2% year on year while all other private non-residential construction is down 5.9%. Information-sector employment is 12.1% below its November 2022 peak.

5. One to two years: wealth and public finances. Through the index concentration described above, and through tax receipts on capital gains.

History offers a useful contrast about where losses land. After 2000, US households lost about $4.76 trillion in directly held stocks and funds, and 87.7% of those losses fell on the wealthiest 10%. Unemployment peaked at 6.3%. After 2006, households lost about $6.19 trillion in real estate, and most of that loss fell on the bottom 90%. Unemployment peaked at 10%. Household debt today is 62.3% of GDP, compared with 93.5% at the end of 2007. On these numbers, today's exposure looks more like 2000 than 2008: concentrated in equity, held mostly by the top. The shock might be absorbed in aggregate. It would not be shared evenly.

7. The strongest case against us

We take the other side seriously, because it is strong.

- The core is still hedge-financed. The four largest spenders generate enormous operating cash flow, carry high credit ratings and have room to cut spending quickly. Microsoft's finance chief has said the short-lived part of spending is precisely the part that would slow first if demand changed.

- Credit markets are calm. The overall high-yield spread was about 3.25% in mid-October. Bank capital is high. Household leverage is low by historical standards.

- Nvidia's guarantee is contingent. Nothing is payable for years, OpenAI must reimburse it, and no AI-related guarantee has paid out so far.

- Old chips keep earning. In mid-October, H100s rented for about $3.13 an hour on one index and even A100s for about $0.95. Power scarcity can extend the earning life of older hardware.

- Strip out the paper gains and earnings still grow. By our rough calculation, stripping these two companies' valuation gains out of full-year 2026 S&P 500 earnings still leaves growth of around 26%. Nvidia's operating profit is real.

- Productive bubbles leave productive things. Even if returns disappoint, data centers, grid upgrades and trained models may support decades of growth.

- AI productivity could raise real growth, which would ease both the debt math and the valuation math.

This is why our position is narrower than "AI is a bubble." We think the fundamentals are still strong and the capital structure is running hot. Those can both be true for a long time.

8. Six signals worth watching

Rather than predict, we prefer to write down in advance what would tell us the cycle has turned. Here is the draft list we track, using public data at the aggregate level only:

1. Capex cuts. The four biggest spenders cut combined next-year capex guidance by 10% or more, or combined quarterly capex falls year on year for two quarters running.

2. Equipment investment. Real US investment in information-processing equipment declines for two consecutive quarters.

3. Cancellations. Three or more publicly disclosed data-center or AI-infrastructure projects of $5 billion or more are cancelled or indefinitely postponed within 90 days, for reasons other than permitting or local opposition.

4. Compute rents. A public GPU rental price index falls 40% or more from its October 2026 reading.

5. Credit. The high-yield spread widens by 250 basis points or more from its October 2026 level, and at least one AI cloud or data-center issuer defaults or restructures.

6. Guarantees. A supplier discloses actual payouts or impairments of AI-related guarantees or capacity commitments totaling $1 billion or more.

Our rule: three of six within any 90-day window means the AI investment cycle has turned. That's a statement about the cycle, not a prediction about prices. As of mid-October 2026, the count is zero of six.

9. What a household can actually see

We don't give investment advice, and nothing here is a recommendation to buy or sell anything. But the system above suggests a few questions worth asking about your own situation.

Are you exposed twice? If AI succeeds, the job that gets repriced could be yours. If it disappoints, your index fund feels it. Our technology piece describes how payroll is being converted into compute budgets. Check whether your income and your savings are tied to the same factor.

How long is your runway? The two questions for companies work for households too. Do you need new money to get through a bad year? Is short money funding long commitments? Borrowed money turns "able to wait" into "unable to wait."

Write the rules while you're calm. Decide in advance what you would do in a drawdown, and don't rewrite the rules during one.

Check less often. Investors who look at their portfolios more frequently see more losses and take less risk than their goals justify. Our companion piece on the nervous system explains why a system that watches itself too closely tends to sound more alarms, and why, in one lab study, five short nights of sleep did more damage to judgment than a single sleepless one.

The accounting question

The technology being real does not make the price right. Railways were real. Fiber was real. Both changed the world, and both ruined many of the people who paid for them, because the financing didn't match the timing.

This boom has something earlier booms didn't: assets that age faster than the debt raised around them, inside an index that most households own whether they realize it or not, at a time when money costs more than it has in almost two decades. None of that tells you when, or whether, the cycle turns. It tells you where to look.

Fire doesn't pick the people closest to it. It picks the driest wood.

See the system. Build leverage. (By leverage, we mean tools, knowledge and collaboration, not debt.)