A New Kind of Business Cycle
In Focus
Financial conditions have been falling fast this year on the measure I’ve used for the last ten years. The GMI Financial Conditions Index, built on the 10-year yield and the dollar, has dropped from the mid-50s in the spring to 44 this week and, on every cycle I’ve watched, a move like that takes the ISM down with it about nine months later.
With the ISM at 54.5, the question isn’t why the economy hasn’t reacted yet based on typical leading and lagging indicators (that timeline makes sense), but rather, why might this economy not feel much of an impact even after that lag runs out?
This article focuses on answering that second question.
Rates no longer run the business cycle. This time, the marginal spender in the US economy doesn’t borrow at the long end, can’t stop spending, and is already bigger than the dotcom build was. That changes what transmits a tightening, it changes which measure of conditions we should be watching, and it changes what eventually ends the cycle. Let me take those in turn.
The Borrower Who Doesn’t Borrow
Normally, rates and the dollar go up, borrowing gets more expensive, companies cut investment, and the cycle rolls over. That works because most companies need cheap money to build.
The hyperscalers don’t. Across the big builders, operating cash flow is running at about $700bn a year against capex of about $586bn over the last twelve months. As I mentioned on Journeyman back in July, speed is everything; whoever secures the most compute wins, and these remarkably profitable companies possess the massive cash reserves needed to fund it.
Take the dumb idea of leverage, just debt as a percentage of market cap; theirs is about 4.5%. The telcos in 2000 were at 25 to 30%, and many were at 100%.
The cushion is being spent, mind. Free cash flow across the group was roughly $1bn in Q2 against $37bn a year earlier, so the operating cash now goes straight out the door as capex, and the gap is being filled with debt – up about $300bn on the year. But look at how they’re borrowing…
AI-related corporates have pushed the average maturity of their new bonds out to about 13.4 years this year, from 11.5 in 2024, at a coupon of about 4.9%. The US Treasury has gone the other way, with new issues averaging about 2.4 years, down from 6.1 in 2024.
So the hyperscalers have locked in 13-year money at around 5%, and a tighter long end doesn’t reprice it. The test for them isn’t the level of rates at all; it’s whether earnings grow faster than the interest bill and, with interest a rounding error against operating income, rates could double, and that test would still pass. The Treasury is now the rate-sensitive borrower, because a short book reprices fast, and that puts more pressure on the Fed to cut. Bill-heavy issuance is also liquidity by the back door, since bills are near-money, so corporates end up supplying the duration the Treasury has withdrawn, which makes them an unlikely pair of bedfellows.
Nor can any of them slow down, because this is the Red Queen. In Alice Through the Looking-Glass, the Red Queen tells Alice that in her country you have to run as fast as you can just to stay in the same place; that’s exactly where Microsoft, Google, Amazon and Meta are. Whoever has the most compute wins the next model, the next model wins the customers, and the customers pay for the next round of compute; the first one to slow down loses, and none of them can afford to be first.
Can a tighter long end stop the build? It’ll slow the rest of the economy, yes, of course, but with that much operating cash flow and 13-year money already locked in, rates don’t fucking stop it.
Bigger than the Dotcom Build
And it’s big enough to matter to the whole economy. US business investment is growing at 9.6% a year, the fastest since 2022, and the part that matters is information-processing equipment plus software, i.e., the AI build. That alone is adding 1.25 percentage points to US GDP growth this year.
The late-1990s internet build averaged 0.93 points a year between 1995 and 1999, and the decade to 2024 averaged 0.34. So, this is already bigger than the dotcom build as a share of the economy, and it’s happening while the long end tightens.…
It isn’t slowing either. Nothing coming out of the hyperscalers suggests a pause; their 2027 commitments are still rising, and their share prices aren’t under any real pressure. The cleanest read is the supply chain that feeds them, and South Korean exports were up 84% year on year in September, with Taiwan’s up 61%, both still accelerating. You don’t get numbers like that from an industry that’s taking its foot off the gas.…
Nobody Has Run Out of Demand
The other objection is that all this compute will find no buyer. The prices say otherwise. The H100 is a three-year-old chip, two generations behind the one Nvidia is shipping now, and it rents for more today than it did a year ago, up around 20% across the big marketplaces. In a normal technology cycle, old kit gets cheaper every quarter, not dearer.
And it’s happening while the price of a token keeps falling, which is the tell: intelligence gets cheaper, so people use far more of it, and the compute bill goes up, not down. Even the open-weight models that are crushing margins at the model layer burn roughly the same compute per token as the frontier ones, so they add to the demand for chips rather than take from it.
Memory tells the same story. Conventional DRAM contract prices nearly doubled in the first quarter of this year and are still rising 10 to 15% a quarter; server memory is expected to stay undersupplied through 2027, and TSMC’s September revenue was up 55% on the year.
US business spending on computers and peripherals, the GDP line closest to AI hardware, has more than doubled since late 2023 to a record $420bn a year and has risen six quarters in a row. That’s not an industry building ahead of demand; it’s a demand the industry can barely keep pace with.
My key point has always been that the demand for intelligence is infinite and that it compounds… more people use it, each of them uses more of it, and every new model uses more compute per answer than the last as it reasons before it replies; additionally, around 90% of frontier-lab compute now goes to inference and post-training rather than to training the model in the first place.
Then the agents arrive, working around the clock and spawning other agents, and inference volume goes up by orders of magnitude again. That’s exponential on exponential, probably a double exponential, and it’s why the capex doesn’t stop when the first wave of chips is paid for.
So Which Conditions Matter?
If the biggest spender in the economy is indifferent to the long end, then a measure of conditions built on the long end is measuring the wrong thing. The 3-month version is the one I originally used. We moved to the 10-year about a decade ago because statistically it should be the better measure, and it was, and we’ve been running the two side by side for a while now. Now we’re now de-emphasising the 10-year version although keeping it on the dashboard for comparison; the thing it measures no longer transmits to the part of the economy that’s doing the growing.
The old GMI build uses the 3-month rate and the dollar instead of the 10-year, and it tells a very different story: conditions were the easiest they’d been in years this January, and they’ve since gone back to neutral — not tight, just no longer easy. The front end and the dollar are what the real economy borrows and trades at, and they’re where the Treasury’s short book lives too.…
Logically, a rising long end should be a big problem, and for the rate-sensitive bits of the economy, it is. But for the sector that’s actually driving growth, the long end barely registers, making the 10-year build a measure of how hard the rest of the economy is being squeezed rather than a forecast of the ISM. Neutral front-end conditions don’t give you a recession; they give you a slowdown in the parts of the economy that need cheap money.…
What Actually Ends It
If rates can’t stop the build, something else has to, and we’ve already seen what it is…
In 2022, rates rose, shareholders punished the spenders, Meta was forced to cut, and the AI contribution to GDP fell from 0.57 points to 0.34. The only brake that works on companies that don’t need to borrow is their own equity holders bailing on them.
The thing that brings that revolt more than anything else is inflation, and we don’t seem to be under that kind of pressure. Headline CPI is 3.7%, higher than anyone wants, but it’s oil, with energy prices up 20% on the year, and core is 2.6%. It’s not accelerating, and the bond market agrees, with 10-year breakevens sitting at 2.35%.
Then there’s the politics. The race with China means it’s very hard for Bessent or for Warsh at the Fed to let conditions tighten to the point where the build stalls, because if it does, their entire economic plan goes out of the window… and they’ve already put it in jeopardy once with the war in Iran. They can’t afford to do it twice.
Who Pays In the Meantime
The economy is running hot, with nominal GDP growing 6.3% a year against a 10-year yield of 5.08%, so money is still cheaper than the economy by over a point. But Q2’s 8.5% annualized nominal growth was 6.1 points of prices and only 2.2 of real growth, and when the two biggest spenders don’t care about rates, the tightening lands on everyone else.
It lands first on the marginal borrower inside the AI trade itself. The neoclouds, the independent data-centre developers and the private-credit vehicles behind them don’t have the profit machine underneath them that the hyperscalers do, and the weaker tier looks a lot more like 1999 than the big four do… Oracle’s free cash flow has been negative for six quarters. We fully expect some of them to fall by the wayside, especially if liquidity isn’t flowing, and the likely end state is the hyperscalers absorbing their assets at a discount. That’s how the last build ended too.
And it lands on Main Street, the rate-sensitive private economy of business owners and house owners and car owners, housing most of all. Sales are back at 1995 levels, and the mortgage burden is back where it was in 1989 and 2006; mortgage rates are too damn high, and house prices are also too damn high. That’s the other arm of the K-shaped economy, and it’s why the headline numbers look fine while the economy feels so strange to so many people: the top is building like mad while the bottom is stuck, and you can’t read that through the average.
Where the Productivity Shows Up
There’s a second-order effect that matters even more than the first. Capex of this size doesn’t just hold the ISM up; it turns into productivity, and the dashboard tells us roughly when: labor productivity follows the ISM by about a quarter. The ISM has gone from 48 at the turn of the year to 54.5, so productivity, which printed 2.2% in Q2 after 2.9% in Q1, should be picking up again through the winter.…
That’s the number to watch for the other half of the K. Productivity minus inflation is running at about minus 1.8 points right now, which is why the “hot” is mostly the price spread turning positive, which lets rates fall and feeds excess liquidity back into the system. The last build did the same thing, with the late-1990s capex showing up as the productivity boom of the early 2000s. And if this build is bigger, so will be the payoff.
And the Tide…
Conditions are one thing, and liquidity is another, so it’s worth checking the tide itself. Global liquidity is still growing, but the pace has collapsed… on a three-month annualized basis, it was running at about 17% at the February peak, and it’s 2.9% now. The six-month rate is 2.9% too, down from over 11% in May, and the level is below its September high.
The US isn’t the problem. Our broad US measure is at a new high and up almost 5% on the year, and even the old narrow measure has turned up, from −0.7% to +1.8% over the last month. It’s a liquidity slowdown, not a growth slowdown, and it’s coming from outside the US, with China drifting and the rest of the world not adding much.
What matters for assets is excess liquidity, the money left over once the real economy has been fed, and that’s still scarce.
The six-month flow is running about two points below nominal GDP on the broad measure and 4.5 points below on the narrow, and it’s been in the scarce regime since August. A 6.3% nominal economy eats liquidity, and right now it’s eating all of it. It’s also why breadth in the US stock market is so narrow, with the advance-decline line rolling over since the middle of September while the index holds up. Same issue: There isn’t enough liquidity to drive all assets, so it concentrates into the few.
So the tide is coming in more slowly than it was, and the economy is taking the lot. With conditions back to neutral on top, that’s a headwind for assets, not yet a strong one, but enough to give equities a wobble. I don’t think it ends the cycle, because the tide hasn’t turned, so this stays a hold… we don’t add here, and we don’t have the full green light to buy the dip yet either. Crypto is still cheap to trend and can catch up when the flow comes back.
What This Means for Markets
So where does that leave the trade? Start with the thing that’s been scaring everyone, the semis.
The semis are back to fair value. SMH is sitting right on its long-term log trend, within a fraction of a standard deviation of it, and from here it could get to one standard deviation cheap just by moving sideways for a few months. Whatever excess froth there was in the summer has gone.…
The Nasdaq is already there. It’s about 4% below its trend, roughly one standard deviation cheap on this channel, and most of the sideways grind since the summer has been fear that this cycle is about to go bust, which – if everything above is right – is going to be wrong. What we should be on the lookout for is a pickup in excess liquidity or liquidity overall, because that’s what gives it the next leg higher. We’re not there yet. Hence the hold. But the setup is cheap to trend, not expensive.…
The weakest of the big players is Oracle. Pull the chart back to 2002 and what you see is a stock that compounds at about 12% a year, overshot its channel by two standard deviations in the summer and is now correcting back towards trend. It’s about 11% below the mean, not quite one standard deviation cheap. It isn’t clear anything has really changed in the business either, and the chances of Oracle going bust are slim… I’d never bet against Larry Ellison.…
Where Am I Wrong?
Three ways, and I’d rather list them myself than wait for you to do so.
First, the long end matters after all. If the 10-year build is the right one and the lag has simply stretched, you’d see it as an inverted yield curve and an ISM falling back below 50 over the next year; this whole argument was a timing illusion. Neither is happening yet, and those are the two things that would tell me it is.
Second, the numbers flatter the build. The 1.25 points of GDP are gross of imports, and a lot of the hardware is imported, so the domestic boost is smaller than the headline, maybe meaningfully so.
Third, inflation reaccelerates. That’s the one thing that turns the shareholders because it takes the Fed out of the picture; a 2022-style revolt would cut capex far faster than any rate rise.
What We Watch
The next read comes quickly. The hyperscalers report Q3 in the last week of October, and their 2027 capex guidance is the number that matters most. If they hold or raise it, the argument holds; if they trim it, the long end gets its lag back.
Then the supply chain. Korean and Taiwanese exports are the fastest read on whether the orders are still flowing. They won’t keep growing at 80% a year obviously, and a slowdown from there is fine; what would worry me is the growth rate heading towards zero, and that looks highly unlikely from here.
Then the curve. The 2s10s steepened to 0.47% on 8 October, right at the top of its channel, and has since fallen back to 0.35%. A break above the channel would be how liquidity comes back through the banks, and that’s what eventually lifts the other half of the K.
A Different Cycle
So let me sum it up. This is a different kind of cycle because the Treasury is issuing at the short end, which supplies liquidity, while the hyperscalers are issuing at the long end, and they have enough margin to absorb any change in rates. They’ve locked in their funding for ten years and more, so they can keep going for as long as they generate the cash flows… and the size of the capex is so big it pretty much dwarfs everything else in the economy.
That’s why the long end can tighten this hard without the ISM following, why we watch the front end and the dollar instead, and why the thing to fear isn’t a rate rise but a shareholder revolt.
What it doesn’t do is fix the K-shaped economy, which won’t resolve until financial conditions improve significantly, as does liquidity. And we’re not there yet. But it’s the one reason I think this might end up being a supercycle. Liquidity still sets the tide for assets, and that tide is slowing. For the real economy (at least for now), the AI build is the larger force.…
And that’s a sentence I wouldn’t have written two years ago.
