Profit investors require, per year
(the front page’s slider: 10%)10.0%
Ruchir Sharma says the pin is not in Silicon Valley but in Washington. The US government and the AI builders now borrow from the same pool of savings, and the interest rate the government pays sets the floor under what everyone else pays. This page sets his argument beside the site’s model, agrees with more of it than not, and puts his mechanism on a slider.
Every bubble in three hundred years has ended the same way: money got dearer. This time, he says, it is the government’s own borrowing that makes it dearer.
Ruchir Sharma is chair of Rockefeller International and a contributing editor at the Financial Times. On 11 September he made this case on the FT’s Economics Show, interviewed by Chris Giles; the FT publishes a transcript, and his column of the same week, “Why America’s debt binge is starting to matter”, makes the same argument in print. Every figure in this section and the next is his, as he gave it there. None of them feeds a number on this site, and the derivations ledger records them that way.
Four signs of a bubble, three of them ticked. Sharma tests for a bubble with four signs, each beginning with “over”: overvaluation, over-ownership, over-investment and over-leverage. Asked on the show which are present, he ticks the first three. US shares in general, and AI shares in particular, are priced in the top tenth of their own history. Americans now hold more of their wealth in shares than in property, which he says no country has done before, and about 40% of the US market is a direct bet on AI. And AI investment as a share of the economy is approaching 5% of GDP, the level the housing boom reached in 2006 and 2007. The fourth sign he leaves unticked, with a date on it. The builders used to be flush with cash and are now net borrowers, Amazon and Meta among them, but their debt is still small against their profits. At the pace they are borrowing, he says, they may get there in a year or two. §4 says one of them already has.
The debt, and why now. The US government owes about $40T in total, or about $32T once money one part of government owes another is left out. Either way it is roughly the size of the whole US economy, up from about 37% of it at the start of the century. The reason he thinks it matters now, after decades of warnings that came to nothing, is the interest bill. It now runs above 3% of GDP, close to a trillion dollars a year: more than the $800B spent on defence, and on course to be the largest single item in the budget. It keeps rising because the government borrows for five to seven years at a time, so old cheap debt is still being replaced with new dear debt.
The yield, and how it feeds on itself. A bond’s yield is the interest rate the market charges its issuer; the ten-year Treasury yield is what investors demand to lend the US government money for ten years, and it is the benchmark every other dollar loan is priced from. It was about 4% in mid-February and about 4.8% when the episode was recorded. Sharma gives two reasons. Oil prices rose and fed inflation, so lenders asked for more. And the market, after ignoring the debt for years, began to worry about it. That worry is self-feeding: it raises the yield, the higher yield raises the interest bill, and the higher bill deepens the worry.
The government’s borrowing rate is the floor under everyone else’s. When it rises, the builders pay more. And the builders are now borrowers.
This is the part of the episode that turns a story about Washington into a story about data centres. Sharma builds it in five steps, and they are worth setting out one by one, because the slider in §3 rests on them.
1. The government’s rate is everyone’s floor. The US government is the safest borrower in dollars, so nobody lends to anyone else for less than they could get from Washington. When the Treasury yield rises, mortgage rates rise, and so does the rate at which companies can borrow. The whole economy’s cost of money moves up together, and the AI builders are inside that economy like everyone else.
2. The builders have a gap, and the market fills it. Companies worldwide are spending about $1T a year on the infrastructure that AI runs on, data centres above all. By the most generous estimate, they earn about $200B a year from selling AI services. That leaves a gap of at least $800B a year. Part of it is paid from the builders’ other businesses: search advertising, online retail, office software. The rest has to be raised in the market, either by selling bonds or by selling new shares. This, in his words, is why AI “has now increasingly become a capital market story” rather than a technology story. The builders that were once the richest companies on earth are now net borrowers and, he says, the largest issuers of debt in the market.
3. They are competing with the Treasury for the same savings. The US government runs a deficit of about 6% of GDP, which means it borrows about $2T a year. The builders’ $800B gap, and the growing part of it raised from investors, draws on the same pool: households’ savings, companies’ savings, and savings from abroad. If the government pays enough, money goes there first. The builders then pay more for what is left, or raise less of it. This, he says, is the difference from every earlier boom: the borrower competing with the builders is the state.
4. The line is 5%, held rather than touched. He gives two reasons for that number. The first is a market habit: for twenty years the ten-year yield has been capped at about 5%, and every time it has neared the cap it has fallen back. A decisive break above it would read as a new regime. The second is arithmetic. The US economy grows at about 5% a year before inflation is taken out. Once the government pays more than that on its debt, the debt grows faster than the economy that services it, which is the textbook definition of a debt that is getting out of hand. Above the line, he expects three things at once: capital is pulled towards Treasuries, the builders borrow at higher rates, and investors’ appetite for risk shrinks.
5. This is how bubbles have always ended. Looking back three centuries, he finds every bubble ended as a monetary event: money got dearer, whether through a central bank raising rates or, before central banks, through something like the gold standard. This time, he argues, a smaller rise is enough to do it, because the stock of debt is so much larger, and most of that debt is the government’s.
Where the yield stands. On 10 September, the day the episode came out, the ten-year Treasury yield closed at 4.95%, its highest since October 2023, when it peaked at 4.98%. It has not closed above 5% since July 2007. Sharma expects the Treasury to fight the line, by talking yields down and buying back bonds, and names two things that could hold it: inflation easing, so that lenders stop asking for more, and a productivity boom arriving fast enough that the economy grows out of its debt. §4 takes up the second.
What this site already records on the builders’ borrowing costs. The pattern he describes is visible in figures already on these pages. SpaceX’s June bonds, the largest sum the build-out has yet raised from public markets, pay an average of 5.855%. A bond backing a data centre leased to Microsoft sold at 5.7% in April; the same issuer paid over 7.2% in August, the rate a company with a junk credit rating pays. Oracle, the heaviest spender of the BIS’s five, was downgraded by S&P to one notch above junk and then sold $20B of new shares to pay for its building. The tracking page’s third clock already counts the debt raised in 2025 and 2026 as coming due in 2028 to 2030, to be borrowed again at whatever rate lenders then demand. Sharma’s argument is about what that rate will be.
The front page holds fixed the one number Sharma’s argument is about. Here it moves.
The front page’s revenue floor is the revenue the AI spending must earn every year to pay for itself: the chips wearing out, the running costs, and the profit investors require for putting up the money. That last item is a slider there, set at 10% a year. But 10% is really two numbers added together. The first is what the safest borrower pays, the ten-year Treasury yield. The second is the extra that investors demand for the risk of a data centre over the risk of lending to the US government, which this page calls the premium. Sharma’s argument is about the first number. The panel below splits the front page’s 10% into the two, puts each on its own slider, and redraws the floor. At rest they add up to the front page’s 10%, so the floor starts where the front page leaves it: $1.30T a year.
How this was built: the floor is the front page’s own arithmetic at its default settings, $3.5T of spending spread over five years, $250B of running costs, and a required return on the $3.5T. The only change is that the return is written as yield plus premium instead of one number. The premium’s starting value is set so that the two sliders add up to the front page’s 10% at the yield of 10 September, so the page adds no assumption of its own about how much investors want; it only shows which part of that 10% the bond market controls. Each point on either slider adds $35B a year to the floor, because it is one per cent of $3.5T. The Treasury yield is the US Treasury’s daily reading for 10 September 2026, confirmed against two independent reports. The premium is held fixed as the yield moves, which is the gentle case: in a scare, lenders widen the premium at the same time as the yield rises. That is what the data-centre bond in §2 did between April and August, 5.7% to over 7.2%, which on this panel is a point and a half on the second slider, or about $52B a year. Move both sliders to see that case. The BIS’s own test, unlike this floor, subtracts the interest actually paid on the debt (its Graph 11.B measures revenue minus capital spending minus debt service); this page’s floor charges a return on all the capital instead, which is stricter and, at any given yield, larger. Full arithmetic: the derivations ledger.
The front page’s own slider prices his case without the split. Its default is 10%; Sharma’s world, with the yield above 5% and lenders nervous, is the upper part of its range.
| Profit investors require, per year | Floor, per year | Multiple of Microsoft’s office business |
|---|---|---|
| 10% (the front page’s default) | $1.30T | 9.6× |
| 12% | $1.37T | 10.1× |
| 14% (the top of the slider) | $1.44T | 10.6× |
Microsoft’s office business, meaning Microsoft 365, Office, LinkedIn and Dynamics together, turned over $135.3B in the twelve months to March 2026; the front page uses it as a yardstick for what a business of the floor’s size looks like.
About $35B a year per point, and about 11% across the whole slider. That is real, and it is second-order. Rates move the floor; they do not transform it. A floor of $1.44T is not a different bet from a floor of $1.30T. What rates change is not the size of the bill but who can still pay it, and that is the subject of the next section.
His funding gap and the revenue floor are the same fact counted two ways. The difference is what each thinks pulls the trigger.
Where they agree, and it is the core. Sharma’s gap is spending minus revenue: about $1T a year against about $200B, so $800B a year to be found from somewhere. The front page’s floor asks a stricter question of the same numbers: not what the spending costs this year, but what it must earn every year to be worth having made. That comes to $1.3T. Set against his $200B of revenue, the floor is six and a half times what AI earns today. The floor is the harder test of the two, because it charges a return on the capital as well as replacing it. His $1T is worldwide; the site’s $840B is the five BIS companies’ plans for 2026 alone. Same order of magnitude, same picture.
Where he adds something the front page holds fixed. The front page treats the profit investors require as a dial and leaves it at 10%. His argument is precisely an argument about that dial: if the ten-year yield breaks 5% and stays there, everything reprices from it, including what investors demand for funding a data centre. §3 prices that. The answer is that the dial moves the floor by about $35B a year per point, real and second-order. The bond market can raise the bill by a tenth. It cannot make the bill small.
Where this site’s evidence is ahead of him. He leaves the fourth of his four signs, over-leverage, unticked, because the builders’ debt is still small against their profits, and says they might get there in a year or two. The strain chart and the Oracle print of 10 September say one of the five is already there. Oracle spent $28.5B on capital in a single quarter, more than half of what it spent in the whole of the previous year. Its free cash flow, the cash left after that spending, was minus $5.4B. Its capital spending over the last twelve months came to 161% of the cash its business generated, and about 244% once the $11.4B its customers paid in advance for capacity that does not exist yet is taken out. And it completed a $20B sale of new shares, the first of the five to fund the build with equity rather than out of the business or by borrowing. That is what his fourth sign looks like from inside the accounts. It is ticking now, in the company the BIS’s five has most exposed to the rate he is watching.
Where the two mechanisms differ, and it matters. His trigger is outside the industry: Washington borrows $2T a year, the builders need $800B, they compete for the same savings, and the government wins. The front page’s §5 circuit is a trigger inside it. Amazon booked a $53.4B gain inside its profit, which was the rise in the value of its stake in the AI lab Anthropic, not cash. Anthropic and Alphabet together pay SpaceX about $2.2B a month to rent computing. SpaceX spent $15.8B on chips in one quarter. The loop closes through a valuation rather than a payment. Both endings have financing withdrawing. His needs a bond market to move first. The circuit can unwind with rates exactly where they are, because it needs only the mark-up to reverse. The Oracle prepayment is the cleanest instance of it yet on the record: the customer financing the builder, and the money counted as the builder’s operating cash.
Where his optimism sits on this site. Sharma’s way out is a productivity boom big enough that America grows out of its debt before the yield settles above 5%. The front page does not rule that out; it is one of its named branches. §5’s fork carries the side where AI delivers as well as the bust, and §6’s dial is the growth story in full, with the condition attached: in the BIS scenarios where AI does not replace workers, whether business as usual or a one-off lift to productivity, workers stay essential and their share of income does not move at any growth rate. In the world Sharma hopes for, he is fine and so are workers.
The difference is the clock. §6 runs fifteen years and prices the claim on income after the bet has paid off; the front page’s own note keeps that apart from the floor, which tests the down payment. Sharma needs the boom to show up fast enough to hold the ten-year below 5%, which is a two- or three-year question about the bond market. Both arguments want the same productivity boom. Neither has evidence that it has started. He concedes as much on the show, and the readings on the tracking page do not contradict him: the companies’ results beat expectations while jobs and wages stay flat.
Four readings would settle who is right, and all four are dated.
The ten-year yield, held above 5%. The US Treasury publishes the reading every trading day. Sharma’s test is not a touch but a stay: a close above 5% that holds for weeks, not an afternoon. As of 10 September it is five hundredths of a point short.
The builders’ own borrowing costs. The next large bond sale by one of the five, or by SpaceX, prices the premium in §3 in public. A repeat of the data-centre bond’s move from 5.7% to over 7.2% would be the premium widening while the yield rises, the case the two sliders show together.
Oracle, around mid-December. Whether the customer prepayments repeat, whether free cash flow stays negative, and whether the $20B of new shares was the last equity it needs or the first. That is his fourth sign, read quarter by quarter.
Productivity, in the statistics rather than the results. His escape and the front page’s §6 dial both need growth to arrive. The reading page carries the measured series so far: a labour share at its lowest since 1947, and no displacement in the payroll data on the scale the bet needs.