A collaboration between Lewis McLain & AI
Part Two of “Forty Trillion Is the Small Number” — posted one day later, because one day was all it took.
Yesterday I posted a long piece arguing that the United States government had passed forty trillion dollars in debt, and that the forty trillion was the small number.
I stand by every word of it. But I left something out, and I did not have to wait long to find out what. By the evening came around the bond market and three separate wire services had explained it to me.
Here is what I missed.
I wrote that piece as though the federal government were the only large borrower that mattered — as though Washington set the price of money and everybody else stood in line behind it and paid whatever was left. That is how it has worked for most of my career. That is how I was taught it works.
It is not how it worked this summer.
This summer, somebody else got to the window first.
ONE: WHAT HAPPENED WHILE WE WERE WATCHING WASHINGTON
Four companies — Alphabet, Amazon, Meta, and Oracle — have sold about $223 billion in bonds so far this year. Last year, all four of them together sold about $108 billion.
They more than doubled it. In eight months.
And that is only four companies. Add Microsoft and the forecasts get bigger still. JPMorgan raised its estimate for technology borrowing this year to $540 billion. Goldman Sachs expects the big five to borrow about $250 billion this year and $400 billion next year. UBS thinks these companies will spend approaching $770 billion on buildings and chips and power this year alone.
For most of my life, these were the companies that did not borrow much. They generated so much cash they did not know what to do with it. Apple and Microsoft and Google sat on cash piles that embarrassed small countries.
That era ended. The artificial intelligence buildout costs more than even those companies earn, and so they have come to the bond market — the same bond market where the Treasury sells its paper, at the same long maturities, to the same buyers.
And here is the part that ought to get your attention. Bank of America’s economists wrote that this borrowing is potentially crowding out demand for long-term Treasury bonds, and that it has played a major part in pushing yields up.
Think about what that sentence means. For forty years, the complaint ran one direction: the government borrows so much that it squeezes out everybody else. Now the accusation runs backwards. Now there is at least a serious argument that private companies are squeezing out the government.
One analyst went further and raised the possibility that all this corporate supply could eventually force the Treasury to sell fewer long-term bonds to keep from paying too much.
Sit with that a minute, because if you read my last piece you already know why it matters. The single most dangerous thing about America’s debt is not the size of it. That’s a squeaky close second. It is how short it has gotten — a third of it coming due inside a year, which is what makes a rate shock hit the budget in three years instead of ten.
And now the government may be pushed even shorter, not by its own choice, but because Amazon and Oracle got to the long end of the market ahead of it.
TWO: WHAT CROWDING OUT ACTUALLY MEANS
Let me get this out of the jargon, because it is a simple idea wearing a suit.
Picture a small-town bank on a Monday morning. There is one loan officer and there is a certain amount of money to lend that week. That is it. That is what he has got.
If the county comes in at nine o’clock and takes most of it for a new jail, then the farmer who walks in at two o’clock is going to hear a higher rate, or hear no.
That is crowding out. One big borrower ahead of you in line changes the price for everybody behind him or her.
For decades the county was always first through the door. Government borrowing was so much larger than anything a company could do that nobody else really moved the needle.
What changed this year is that a handful of companies got big enough to matter in that line. Two hundred twenty-three billion dollars is not a rounding error. It is real money competing for the same lender’s attention, at the same long maturities, in the same week.
As one strategist put it plainly: whoever is issuing — a government, a tech company, anybody — is now competing with more borrowers. And so yields have to be higher.
The evidence that this is duration and not the Fed
https://yield-curves-lfm.netlify.app
Bookmark the link above. If you are unable to interpret what it means, that is perfectly okay. My audience is wide, so I know only a few would even read this particular blog. Just know that a basis point in 1/100th of one percent. When the rates go up only 1 basis point, sounds tiny, the Federal government will eventually pay $4 billion every year just on that level of current outstanding debt! What if the rates go up 10 basis points? I know for sure that the person sitting in honking traffic isn’t focused on this condition. But hear me out.
Split this year in half at June 30 and look at what the Treasury’s own yield curve did.
| Short-Term (2-Year) | Long-Term (30-Year) | |
|---|---|---|
| January through June | rose sharply | barely moved |
| July through August | barely moved | rose sharply |
In the first half of the year, the short end did all the moving and the long end sat still. That is a market arguing about the Federal Reserve — about whether rates get cut this fall or next spring. When they quibble over the increments or decrements, the are usually talking about 25 basis points, sometimes 50. Quick quiz now that you are informed. What does 50 basis points do to the Fed’s $40 trillion, not to mention all other governments? Even you.
In the second half, it flipped completely. The short end went quiet and the long end took off.
That is not an argument about the Fed anymore. When the thirty-year moves and the two-year does not, the market has stopped asking what the Fed will do next quarter and started asking a different question entirely: what does a person deserve to be paid for parting with their money until 2056?
And that question — what long money is worth — is precisely the question a flood of thirty-year corporate bonds forces the market to answer.
The thirty-year Treasury closed at five dollars and thirty-one cents per hundred on August 17, and touched about five dollars and thirty-four cents the next morning. That is the highest since 2007. Do you recall what happened back then?
One correction to yesterday’s piece. I wrote that the government sold thirty-year bonds on August 13 at five dollars and twenty-two cents. That was accurate as far as it went — it is the price at that particular auction. But the market kept right on going in the days after, and by the time I hit publish it had moved. Then on the 19th, the Treasury doubled the size of its bond buyback operation to steady things, and yields came down some.
I would not read too much into that last part. As one strategist noted, the buyback operation changes very little about the fundamentals — which remain the need to finance an enormous wave of technology borrowing on top of very large government deficits. You can rearrange the furniture. The house is still the same size.
THREE: THE HONEST OTHER SIDE
Now I have to slow down, because I do not want to sell you something the evidence will not carry.
There is a real argument on the other side, and it deserves a fair hearing.
Corporate bonds and Treasury bonds are not the same thing. They do not draw from a single fixed pot of money. Different investors buy them for different reasons, and a pension fund that buys an Amazon bond was not necessarily going to buy a Treasury bond instead.
These are not shaky borrowers. The hyperscalers are among the strongest credits in the entire investment-grade market. They carry far less debt against their earnings than a typical big company. Treating a company with that balance sheet as a source of stress in the government bond market may confuse a lot of supply with a lot of risk, which are not the same problem.
And there are other suspects. Japanese long-term rates have been climbing for reasons that have nothing whatever to do with data centers, and world bond markets move together. An inflation shock from the war has repriced everything. Foreign central bank buying has been soft. The federal deficit itself is running near two trillion dollars annually. Any one of those, by itself, could have pushed the thirty-year up this summer.
So here is what I actually believe, stated as carefully as I know how:
The AI borrowing is not the cause of higher long-term rates. It is a genuine, measurable, and rapidly growing contributor to a move that has several fathers. Anyone who tells you it explains the whole thing is selling something. Anyone who tells you it explains nothing has not looked at $223 billion.
And the direction of travel is not in dispute. Whatever share you assign it today, that share is getting bigger, because the buildout is not slowing down.
One thing to watch that I find genuinely sobering
Demand for these bonds is cooling even as the supply climbs.
Apollo tracked how many orders came in relative to how many bonds were offered. In February, hyperscaler deals were drawing nearly five dollars of orders for every dollar offered. By July, that had fallen to under two.
And a Reuters review found that of ninety-one hyperscaler bonds issued this year with comparable data, seventy-eight were trading at worse prices than where they were sold. Buyers who took them at issue are underwater. That means they are holding investments that have lost their value.
Supply climbing while appetite falls is how markets get repriced. Not all at once. But steadily, and in one direction.
FOUR: AND NOW LET US TALK ABOUT TEXAS
Everything above is a national story. Any newspaper can write it.
Here is the part that belongs to us, and I have not seen anybody put it together, and it is the reason I sat down with Claude to write this at all.
Texas is paying for this twice.
Once at the bond window, in higher interest on every road and school and water line we finance.
And once at the tax office, because we are subsidizing the very buildout that is raising our cost of money.
The numbers
Texas exempts qualifying data centers from sales tax on their equipment. Two sections of the Tax Code do the work — 151.359 for regular data centers, 151.3595 for the very large projects. To qualify, a facility generally needs at least a hundred thousand square feet, a commitment of two hundred million dollars over five years, and twenty qualifying jobs.
When the Comptroller’s office scored that exemption back in 2020, it estimated the cost to the state at about $29 million for fiscal 2025.
The actual figure for fiscal 2025 came in at over $1 billion.
Read those two numbers again. Twenty-nine million estimated. A billion delivered. That is not a forecast that missed. That is a forecast that was off by a factor of more than thirty.
And it is still climbing. The Comptroller now projects the annual value approaching $1.8 billion by 2030, and the state expects to forgo about $3.2 billion over the next two years.
Texas has more than three hundred operating data centers, with a hundred forty-two more under construction — the most of any state in the country. Fifty-nine new certifications were approved this year alone, and the Comptroller’s office expects roughly ten new ones every month going forward.
Now here is the part that lands directly on city hall
Section 151.359 exempts the state sales tax. But section 151.3595 — the one for large projects — exempts the state and local sales tax both.
That is your city’s money. That is your two percent. That is the revenue line your finance director builds a General Fund budget on.
Layer on top of it the property tax side — Chapter 312 abatements, the JETI program — and a city can find itself hosting an enormous capital investment while collecting a fraction of what a facility that size would ordinarily generate.
And nobody is checking
This is the part that would end a municipal finance career if it happened at a city.
Of the one hundred thirty-eight data centers certified for the exemption, the Comptroller’s office has audited twenty.
Of those twenty, six were found in breach of the terms. Four failed to create the jobs they promised. One did not build to the required square footage. One reported that its required power agreement had fallen through.
Six out of twenty. Nearly a third of everything they looked at.
And a hundred eighteen certified facilities have not been looked at yet, while ten more get added every month.
I want to be careful here, because a thirty percent failure rate among the first twenty audited does not necessarily mean thirty percent of all hundred thirty-eight are out of compliance — auditors often start where the questions are. But that is exactly the point. We do not know. The Comptroller’s own testimony was that the office has little information about whether these facilities are meeting their promises. Facilities found in breach do have to pay the money back, which is the system working. But it only works on the ones somebody looks at.
Accounting for results with such a few number of data centers involved (few for those of you who manage hundreds of big projects), is simply not that hard. Were project tracing and results reporting not part of the agreement in the first place?
The Legislature noticed. The Senate Finance Committee held an interim hearing on this in July, and both the Governor and the Lieutenant Governor have raised questions about the exemption’s cost. One senator put it about as plainly as it can be put: no tax exemption should operate on autopilot.
The fair case for the other side
I am not going to pretend there is no argument for these incentives, because there is one and it is not foolish.
A data center brings enormous capital investment into a county. Even abated, the improvements eventually hit the tax roll. Construction employment is real while it lasts. Utility load can improve a municipal system’s cost recovery. And Texas won this industry over Virginia and Arizona partly because of these incentives — if we repeal them, some of that development goes elsewhere, and we get the higher interest rates anyway with none of the buildings.
That last point is the strongest one against me, and I want it stated in its best form: you cannot stop the national borrowing wave by taxing Texas data centers. The bond market does not care what Collin County does. Repeal the exemption tomorrow and the thirty-year Treasury does not move a basis point.
Fine. Granted. But that is an argument for getting the exemption right, not for leaving it on autopilot at thirty times its estimated cost with eighty-five percent of the participants unexamined.
FIVE: WHAT A CITY COUNCIL SHOULD ACTUALLY DO ABOUT IT
Enough diagnosis. Here is the practical part.
Reprice your capital plan, and do it now. If your five-year plan assumes anything close to what money cost in 2021, it is fiction. Long rates are at twenty-five-year highs and there is a very large private borrower standing in line ahead of you who was not there before. Run your debt service at today’s rates plus a cushion, and find out what falls off the list before the bond market finds out for you.
Read your economic development agreements again — specifically the sales tax section. If you have a large data center project under 151.3595, find out exactly what local sales tax you are forgoing and for how long. Put that number in your long-range forecast as a line item, not a footnote. If you have a Chapter 380 or 381 agreement layered on top, add it up.
Ask for the compliance file. If a facility in your city is certified for the exemption, somebody should be able to tell you whether it has been audited and whether it met its job and investment commitments. If the answer is that nobody knows, that is your answer, and it belongs in front of your council.
Watch the Legislature. The 2027 session is going to take up this exemption. Cities with certified projects should know what a repeal, a cap, or a retroactive modification would do to their revenue — in both directions. This could be a revenue gain for a host city or a lawsuit, depending on how it is written.
And watch your own investment portfolio, gratefully. The same rates that are punishing your bond issues are paying your operating funds better than they have since 2007. Book the income. Do not build a payroll on it.
CLOSING
Yesterday’s arithmetic has not changed, and I would not take back a line of it. Forty trillion borrowed. Something near eighty-eight trillion promised and never written down. The trust funds empty, with dates already on them — 2032 and 2033.
What one day added is that Washington is no longer the only one standing at the window.
For my entire career, the government has been the largest borrower in the world by so wide a margin that everyone else was a rounding error. That is ending. Not because the government got smaller — it has never been bigger — but because a handful of companies decided to build something enormous, and discovered they could not pay for it out of pocket either.
So now there are two great borrowers in that line. Both of them want long money. Both of them want it this morning. And neither one of them is going to stop.
The lender is the same lender he always was. There is only so much in the drawer on Monday morning.
And the man who walks in at two o’clock — the school district refinancing a bond, the city building a fire station, the young couple buying a first house in Princeton or Anna — that man does not get told about any of this. He just gets told the rate.
Here in Texas, we have done something that would be funny if it were not our money. We have written a tax exemption we scored at twenty-nine million dollars, delivered a billion, and put it on autopilot heading toward one point eight billion a year — to subsidize the construction of the very thing that is raising the interest rate on our own bonds.
I do not say that to argue the exemption should be repealed. Reasonable people are lined up on both sides of that, and it will get settled in Austin in 2027, not by me.
I say it because a city that does not know it is paying twice will only budget for it once.
SOURCES
Hyperscaler issuance volumes: LSEG data as reported by CNBC and Reuters, August 2026. Forecasts: JPMorgan, Goldman Sachs, UBS, and Barclays analyst notes, 2026. Crowding-out analysis: Bank of America economics notes and Bloomberg reporting, August 17, 2026; Federal Reserve Bank of Dallas, “How AI debt financing impacts duration supply and interest rates,” February 10, 2026. Order-book cover ratios: Apollo Global Management. Secondary market performance: Reuters analysis of LSEG data, July 28, 2026. Yield curve movements and auction data: U.S. Treasury daily yield curve and TreasuryDirect. Treasury buyback operation: August 19, 2026.
Texas data center exemption figures: Texas Comptroller of Public Accounts testimony before the Senate Committee on Finance interim hearing, July 27, 2026, as reported by The Texas Tribune and others; Texas Tax Code sections 151.359 and 151.3595; 34 Texas Administrative Code section 3.335. Data center counts: Comptroller testimony and industry analysis.
The interpretation, the Texas argument, and any errors in either are my own.