The Upsidedown: Financial Chicanery and the Reality of the AI Bubble

Financial Chicanery and the Reality of the AI Bubble

Let us offer an alternative world model for you to contemplate as you see fit. We don’t mean to kill the vibe, but we’re here to call balls and strikes as we see them. We could be wrong, but we contend that the gulf between perception and reality – in just about all macro domains – is reaching an imminent breaking point. The popular market and macro narratives supporting this euphoria are about as divorced from reality as these things get. At least historically, reality has won these battles. We’ve commented on many such narratives in bits and pieces scattered throughout our work, but always emphasized that the narratives are ex post, while the flows drive the action. This is a manifestation of the way human beings operate. Decisions are unconscious, actions are taken, and afterwords the mind makes up the explanation for the action, to preserve the illusion of control. The more untethered those ideas become from reality, the greater the cognitive tax, the worse the misallocation of energy, and in this case, also of capital.

Since we view narratives as non-causal in the main, we haven’t focused on debunking them. Instead, we’ve tried to convey the archetypes of macrophenomena at play. How bubbles work. Why technological adoption curves necessarily lag what gets priced in. How technological revolutions are borne out of financial bubbles, not the other way around, despite becoming the namesake of those bubbles. Do we really think tulips were the future? Or did the Dutch Empire happen upon a load of precious metals in faraway lands, exploding money and credit that needed an outlet? Every bubble worked that way. Railroads. What happened there? Greenbacks, and then an enormous gold and silver influx from South and Central America in a post-war bimetallic currency standard is what happened there. What about radio and the 1920s? You guessed it, suspension of gold coin convertibility by the British, the adoption of a gold-exchange pound standard across a decimated, overvalued Europe, and inappropriately easy Fed policy designed to maintain that standard and juice credit in service of industry. At the end of the day, technological revolutions and government-sanctioned money printing are two sides of the same (debased) coin. Money printed will be money spent. There will always be a reason. Narratives may not be causal, but they do serve a function – they reinforce the actions being taken, whether they are productive or not. What gets built eventually gets used, but not before tight capacity stokes inflation, erodes real income, contracts liquidity, and the bubble collapses into recession.

LIQUIDITY DRAINS ARE THE BEGINNING OF THE END

We’re in the early innings of one of the nastiest contractions in liquidity in recent times – the sort that topples financial pyramids and the assets and economies tethered to them. While bubble HQ just delivered the wildest, most levered blow-off top this side of the Civil War, dollar liquidity and cross-border lending are tightening acutely this year. That dichotomy is an early warning sign for global growth and the most levered markets in modern American history.

We’ve just seen one of the biggest contractions in the pace of inflows to EM equity markets on record (above right, blue line). EM is very much the ginger stepchild of global liquidity flows, so large moves like this tend to coincide with the inflection in the global cycle. Sharp deteriorations in global growth (red line) have always followed such large sudden stops in cross-border flows. This all-time record blow-off in US momentum wasn’t driven by real money inflows, so much as from historic amounts of retail leverage. This is the aspect of levered retail blow-offs that makes them inherently self-reversing – a classic end to a bubble already facing an unfolding liquidity contraction. This now marks an unprecedented fifth (!) levered retail blow-off top to the post-Covid cycle (Feb 2021, Nov 2021, Dec 2024, Oct 2025, and now, Apr/May 2026). If this were a normal cycle, each of these would’ve bookended it. But so much money was printed and spent after Covid, that after 2022’s inflation, it sat parked on the sidelines in the Fed’s RRP until it exhausted itself late last year. This blow-off, to our minds, represents the last available flow into the bubble, since real income is falling and the liquidity contraction has arrived. Steep underperformance of the market’s former leadership (inc. Mag 7, PE and dollar-sensitive indices like H-shares), and near-record equity funding costs in what is typically the loosest quarter of the year speaks to the dichotomy between liquidity on the one hand, and leverage on the other. The squeeze is even impacting short-dated bills in a way we haven’t seen since the 2022 bear market, and before that Covid. This contraction is working its way out the risk curve from money markets to FX, to sovereign bonds to low grade credit, up the credit stack, to global equities, to all non-momentum domestic stocks, and now, seemingly, to momentum.

Late-cycle Fed easing has preceded (but failed to forestall) all modern recessions, pushing 2s well below Fed Funds. That liquidity release stokes a final burst of spending and levered asset buying. These drive up asset and consumer prices, create a short-lived end-of-cycle commodity blow-off, and contract liquidity (pulling 2s back up to a premium). Today that rate premium sits at about + 50bps – the level it reached right before bearish market catharsis in each of 2H01, 2H08, and 2H21. This pattern and degree of liquidity tightening after a late-cycle easing has, at least over the last twenty-five years, reliably preceded busts and the onset of recession. We’ve grown accustomed to that being followed by emergency easing – not a return to hikes, though conditions today are more structurally stagflationary. Stocks have experienced some of their worst historic drawdowns in the six months after this exact sequence of events.

**

THE AI BUBBLE IS ACTUALLY **SMOTHERING AMERICAN GROWTH

The ubiquitous belief is that AI is boosting growth. From our vantage point, that’s a total inversion of the truth. It hinges on first-order, nominal spending only, without following that spending through the economy or financial pipes. It doesn’t take account of where it goes (i.e. imported goods and commodities), whether it’s real or nominal, or its inflationary and capital-consumptive effects. It doesn’t factor the consequent crowding out of everything else, especially everything else debt-funded and interest-rate sensitive (such as the entire rest of the American economy). AI is currently a net negative to growth, and a bonfire of cash (at least up front) at a time when liquidity and real incomes are already contracting. That mix is – at least right now – sucking the rest of the economy dry. For the top branch of the K-shaped economy to exist, the bottom branch must also exist. They aren’t separate phenomena.

It’s not mechanically possible for AI to scale returns as priced in, over the priced-in timeline. We’re not asserting this as mere opinion – it’s inescapable because at the end of the cycle, available (physical) resources are depleted, no matter how many dollars are printed or borrowed. Further spending simply creates inflation, keeps rates up, and squeezes real income. This explains the memory and broader commodity price moves. Every dollar of AI spending in this zero-sum situation crowds out a dollar of potential demand one-for-one. So AI will accelerate the recession before it can generate revenue, because AI-related activity is smothering its own potential customers. Normally, capex aims to increase the supply of things already in strong demand, thereby alleviating economic tightness down the road. But emerged out of a weak economy…and isn’t responding to any demand signal at all (which is why none of it is financed out of revenue). It isn’t building any physical supply of goods or commodities to increase economic runway – just compute. And most of the spending flows abroad, via imports. So it crowds out domestic income and ships it to other economies. In other words, it’s currently a net drain and a productivity drag, which is why it’s also inflationary, not deflationary. So it’s a no from us on the “deflationary productivity boom” narrative. There’s no capex on the horizon that will bring down the cost of real tangible things, so it’s hard to see how non-inflationary demand returns.

► Just to be clear – technological change has been enormously, incredibly beneficial over the long run. None of this is to say whatever gets built won’t be used, or helpful, or productive in the future. But that’s always going to be true no matter what – whether it gets used doesn’t speak to whether it is positive RoI, or whether that capital was put to its best possible use relative to others. This seems more uncertain than most in that respect, because most technology comes in response to a given identifiable problem, there’s at least some indication of a clear use case upfront, and some demand and cashflow funding it. It was clear to everyone how railroads would benefit them. AI is, if anything, scaring people, rightly or wrongly. And they’re also aware that it’s already costing them income.

► Real US personal income is now declining nearly on track with the average recession, and faster than several. This is exacerbated by spiking imports. When economies overheat, and yet capex spending continues, imports must rise as a share of spending, because capacity is tight. Accelerating imports (of goods and commodities) to supply capex booms at the ends of cycles are income drains that leak abroad and accelerate the cycle’s end.

► Here’s another way to visualise the crowding out of income, non-import spending, and domestic liquidity:

The left-hand charts show the 2007-2008 cycle, and the charts on the right show this cycle, leading into and since Covid. The blue line is the same short-end yield curve move we show to proxy liquidity tightening. As imports rise as a share of the economy, this depletes local / onshore liquidity and steepens the curve. This effect is visible throughout 2008, 2022, and this year. It also occurred in 2025’s tariff frontloading, but the still-positive RRP balance at that time flowed out in response to this tightness, preventing net steepening until now. Finally, in all three cases this tightening corresponded to local peaks in the labour market, as already slowing non-import spending got crimped by a simultaneous decline in real domestic income and liquidity.

► In other words, incomes leave the US economy and benefit the economies of, in this case, North Asia. It’s not just a one-off drag to growth in a given quarter when imports surge (as you see in the GDP report, almost cancelling out the GFCF contribution). Since one man’s spending is another man’s income, any spending abroad creates income abroad, in lieu of creating income at home. It compounds as lost income that’s no longer circulating in the domestic economy for all future quarters. This is true whenever imports aren’t immediately income-accretive (via profitable production).

► The memory price blow-off – itself the tail wagging the dog – accelerated that drain. It operates the same way any commodity price move does (like oil prices) – transferring income from losers to winners. This isn’t theoretical. Chip workers in Korea are getting nearly half a million-dollar bonuses while the prime-age US labour force just registered its biggest monthly decline outside of Covid. The labour market is clearly weak.

► The cumulative YTD percent decline in household survey employment is bigger than the recession years of 2001 and 2008. The labour differential points to a sharp move higher in the unemployment rate, which has only remained contained because folks are dropping out of the labour force in size.

► This isn’t a retirement issue. Prime-age workers are giving up on the job search faster than ever (ex-Covid). Businesses are strained. Rapid late-cycle drawdowns in C&I lines reliably point to unanticipated cashflow gaps. Businesses then respond to cashflow pressure by cutting spending, including on labour.

► Weak income growth, an inflationary shock, job losses, depleted savings, maxed out borrowing and crisis-level delinquencies, are driving consumers to the wall. 1Q26 PCE was revised down to just ~50bps annualised. And showed weakness across both goods and services, giving back tactical frontloading and war stockpiling.

► The wealth effect in the top-branch of the K adds to the inflationary spending from the capex itself, and both have prevented rate cuts that would offer relief to households on the lower branch of the K. In aggregate, consumers are experiencing some of the highest loan delinquency rates ever, across all loan types ex-mortgage. In short, there are always trade-offs and opportunity costs that surface-level growth narratives ignore.

MONEY PRINTED WILL BE MONEY BORROWED WILL BE MONEY SPENT

But why and how are we’re in a K-shaped economy in the first place? No bubble before has ever occurred in such a weak economy, let alone during three full years of economic decline. And how did a capex boom emerge from this economic backdrop, that’s completely divorced from any demand signal? Since strong end demand normally brings about capex, it’s normally funded by profits. But that’s not the case today. How was it financed, since rising economic savings are normally required to take up bond supply (as with demand for government debt during recessions)?

The answer is that it was all, effectively, monetised. The assertion we made in the intro (and will make until our dying day) is that the government creates the money bubbles that birth non-productive borrowing, which takes whatever form it can at the time. Whether the debt-funded spending it enables it’s ultimately useful or not, and to what degree, is completely secondary. The extremity and length of this unprecedented “K” is the exception that proves the rule that money, once printed, will be borrowed, and then will be spent, regardless of whether or not there’s any underlying demand for that spending. The post-Covid cycle was a government-sponsored money and credit expansion that got channelled first to households, and then to folks with access, who basically spent it on more of what they were already doing (building compute). The injections were vast enough to outstrip global capacity (driving inflation), fund the 2021 all-asset bubble, and still leave nearly $3tr leftover to be parked in the RRP. In 2021, the money went into housing and stocks as always, but it also went into compute expansion and use cases for compute – crypto, NFTs, web3, metaverse. Remember all that? None of it was productive. It also went by way of PE and VCs into droves of loss-making software start-ups (providing a wave of demand for…compute), many of which may be viable, but many of which have died or been floated along as zombies hooked on private credit. The compute will be built and the reason will follow. This time may be more useful, but our point is, there’s no a priori evidence of that. The RRP was dry powder for a record surge in repo borrowing that directly and indirectly financed the AI bubble – the capex and the speculation.

► In other words, this bubble was financed by an unprecedented three-year binge of monetised financial leverage, even as growth slowed sharply in the rest of the economy. To the extent the Fed’s rate hikes, QT and import spending all tightened liquidity, there was dry powder in this facility to offset it. Money must go through financial intermediaries first. Banks borrowed in the repo market and lent it onwards to non-bank financial players and investors. These loans now total over $3tr. The dry powder ran out in October 2025. Before October, as the middle chart shows, banks could simply borrow repo and make these loans one-for-one. They moved together. After October, repo flows have moved inversely to these loans. In other words, repo is now a zero-sum game. Banks borrow repo for many normal reasons – but late in the cycle incomes decline and bond supply rises, so demands for cash outstrip available wholesale funding. Interbank repo falls, and loans to financial players get cut. You can see this entire source of funding for the bubble, for the three-year “K”, is hitting a wall. In 2Q26, there was a brief window when government bond and bill supply went negative, easing repo markets and providing balance sheet for the fastest and biggest spike in equity margin borrowing on record (front page). That drove the stock blowoff.

► The challenge post-2022, in a stagflationary economic slowdown with high nominal rates and broad-based credit distress, is where to find borrowers with both the willingness and ability to borrow. So where did all these “non-bank” and “unclassified” loans go? Lenders require collateral and/or cashflow, while borrowers require a productive use to spend the funds on. Or do they? A third of these loans are untraceable, but of the $2tr or so we can break down, the biggest chunk of lending (c. $1.2tr) went to private credit, private equity and ABS pools (consumer and mortgage). They used that borrowing to fund payouts to investors in lieu of exits, replace lost income as borrowers went delinquent, fund new AI start-ups, and extend new PC loans, many to roll pre-existing loss-making start-ups, and many to fund... securitised AI-related capex. More on those structures shortly. The residual – some $600bn in level terms – are direct equity margin loans.

**

SOME ALTERNATIVE AI NARRATIVES TO CONTEMPLATE**

We understand that the casual narratives appear compelling – there’s an AI revolution, the spending will go up since that’s what companies say, semis are no longer cyclical, etc. On the surface this appears to be what’s going on. But a technological revolution (and its adoption curve) and a financial bubble are two completely different phenomena with different physics. They transpire over very different timeframes. History shows that the financial bubbles precede and birth the technology bubbles, profitable adoption takes much longer, and many players fail in the process. The three-year divorce between stocks and the economy is first and foremost a result of latent funding in need of a destination. This whole thing is occurring in an otherwise depressed economy, so these challenges are all the more real. So the narratives are one thing, but how does the spending (both on the capex and on the assets themselves) continue to get funded in this environment? Having now provided our view that it basically can’t because of inflationary and liquidity constraints, we will turn to the narratives themselves. Because today, across many dimensions, the narratives are already breaking apart. Datacentre construction growth is declining, we’re at hard constraints to resources (energy and related equipment, grid connectivity, chips, etc), the LLMs are loss-making with negative unit economics, bond indigestion is showing up everywhere (whether in hyperscaler spreads, low bid/covers, hung debt at banks, and increasingly, the turn to equity raising), LLM returns aren’t there (and can’t be, for the cyclical reasons above), attempts to switch to (still unprofitable but less subsidized) token pricing have resulted in immediate pushback and rationing of corporate token spend, pilot programs are being cancelled for lack of productivity (again a mechanical outcome of a declining economy), American players are railing against IP sovereignty issues around closed-source LLMs, and the Chinese are dominating adoption share with cheaper open source options.

“COMPUTE SHORTAGE”: Is It Really A Shortage When It’s Given Away?

► It’s very easy to manufacture “a shortage” by giving things away for free, to recipients of free money. There was a shortage of Uber drivers back when Uber was loss-leading at half the price of a regular cab. LLMs are banking on a VC scaling model that worked before, but without the same conditions. VC-funded companies and technologies have routinely taken scale by operating at a loss to displace incumbents in existing industries that already had revenues they could capture (again, think of the Uber example). This is dumping, and it’s only possible if money is available to fund their upfront losses. Another reason (aside from the macro) it’s not possible for AI to seamlessly transition into a profitable exponential profit curve that validates the pricing, is that there’s no incumbent revenue to capture this time.

► The LLMs are selling each $1 of compute for c.30 cents, tried to raise pricing and immediately faced pushback, are talking about a price war even from these levels, and it’s very hard for negative free cashflow companies (LLMs)

to scale commoditized products whose unit economics are deeply loss-making, during a liquidity crunch, when they’re running out of cash, competing against foreign state-subsidized open-source competitors. Re: compute supply, each expansion of tech capex since 2015 has conveniently involved expanding compute at the same pace as before, right at the point where growth started to level off. The current wave of growth pre-dates the release of ChatGPT and is tiny vs. GDP – another reality that flies in the face of the growth narrative. Most of the money is going on imported chips and equipment. Growth has already stalled out, as projects get halted. So hang on, where are all the chips going?

► It might be cynical to see this as a free money search for more uses of compute, to justify building more compute. But at the same time, it might not be. When the rate of capex growth slows, the lag in D&A that flatters profits in these buildouts catches up, and profits fall. There is a profit motive, even if it ends up being an uneconomic, non-cash one.

► What’s indisputable though is that post Covid, compute demand growth came from loss-making companies, who funded it with venture capital. Flows into the PE/VC/PC industry were in turn, a product of monetary and fiscal expansion. Before Covid, the key driver of compute demand growth was corporate cloud migration. This was big and came with recurring annual demand. But as a growth driver, it was a level change. Once the majority of large companies made the switch, growth inevitably slowed. That growth slowdown was the status ante, coming into Covid (as you can see in the charts above), which then saved the day with another wave of free money and more users of compute: SaaS start-ups, a record VC boom, crypto, NFTs, Web3, Metaverse, and now AI. How’s all that doing? The free money firehose was already fading by early 2022, but not before catalysing a new wave of datacentre expansion that has continued since then. Where would compute demand be without this free money? We can’t know, but in stagflation, we may find out.

► Compute is being built today in anticipation of a shortage tomorrow, predicated on widescale adoption of AI. While we think the technology is more akin to a highly-effective knowledge management software than proto-consciousness (and that LLMs are a dead-end in that respect), we agree that knowledge management is a huge economic deal. But for instance, for AI to generate returns through higher revenue, companies would need to borrow more to spend on new things, which creates that revenue, in anticipation of end demand also keeping pace via higher spending (itself requiring higher borrowing). Yet balance sheets are maxed, bankruptcies are high, funding conditions are tight, inflationary constraints are palpable and rates are rising, not falling. Alternatively, for AI to generate returns through cost savings, workers would have to get fired, which decreases their spending and therefore economy-wide income, negating the cost savings for businesses in aggregate via revenue reduction.

► Meantime, the “compute shortage” narrative is starting to come apart, at least in places. While everyone involved claims there’s a shortage right now, even if that’s true, it’s because of training demand from LLMs and the other post-Covid demand drivers listed above, a lot of which won’t survive this liquidity contraction, let alone this cycle, in our view. Compute capacity tripled in the decade to Covid and was overbuilt before the post-Covid explosion of free-money. In June, the Information published a report stating that certain of Meta and xAI’s facilities were running at low utilisation rates. While this was dismissed, subsequently both Meta and xAI announced their intent to resell / onlease compute they had built and/or committed to buy ostensibly for their own use. None of that is to say that other players aren’t constrained, just that these two – who made the same assumptions about the inevitable explosion in demand – suddenly find that they aren’t. None of this proves anything, but at the very least, it’s circumstantial grounds to be sceptical.

VENDOR FINANCING: SPVs & Unprecedented Financial Engineering

Since the AI build-out is debt-funded, and many players are themselves loss-making start-ups with limited firepower, creative financing has been necessary. Even Nvidia – the biggest beneficiary of the picks and shovels spending – reported some $50bn of free cash last quarter but failed to grow cash and is now raising $20bn of debt. It’s true some of that went towards buybacks, but the main cash drag was an increase in “short term investments” related to vendor financing Nvidia is offering to finance many of the purchases of its own chips.

► The use off off-balance sheet vehicles has been widespread both for datacentres and the chips that go in them: There are really two different models that are key to how this is being funded and where the debt is being hidden. Firstly, the datacentre projects themselves are ringfenced in SPVs and held off the balance sheets of hyperscalers and neoclouds. In this set-up, the datacentre and its lease commitments provide the collateral for project finance from the banks (again, either directly or indirectly…ultimately, all of this is funded by the liquidity created by that RRP release, counteracting the impact of QT as previously discussed). But the players involved – both lenders and borrowers – have every incentive to structure it this way.The lease commitments don’t actually get recognised by the hyperscalers / neoclouds until the facilities are operational. So the debt-funded spending happens, but not only is there no recognition in the P&L via costs, there’s no recognition in the balance sheet. Meantime, the vendors recognise equipment sales upfront. This is why index earnings look strong. It’s the same profit illusion and structure as the Dot Com. We’re not saying it’s illegal or fraudulent. We’re just saying it’s a non-cash accounting mirage that deserves a PE of zero.

► These lease commitments now total some $850bn or so across the space: This number changes rapidly with every announcement of projects, while many never seem to break ground, but as it stands, we’re approaching $1tr of future lease commitments to rent out compute. Of that, only about $200bn is currently recognised on the balance sheets of hyperscalers and neoclouds. So when you see the existing deterioration in their financials, bear in mind that there’s lots more where that came from. This way “debt” is transformed into a “lease obligation” that’s not even recognised. While lenders require legally-binding commitments, the ringfencing via SPVs and the existence of force majeures imply to us that these structures are probably designed with a way out. That’s speculative, but if not, it’s even more imprudent.

► These second type of funding is lending against the equipment itself, especially chips: At the end of the day this is a collateralised lending cycle. Take out debt, use it to buy chips, value of chips goes up, take out more debt, use it to buy more chips, and so on. Housing – pre-GFC – is a textbook example of a collateralised lending cycle that led to overinvestment. The problem is that when debt funds the purchases, and more purchases are predicated on the values of the underlying collateral, whenever the debt stops flowing, there’s a glut in the collateral and the price falls. If there has been, in fact, hoarding of chips in a perceived shortage (thereby creating a shortage), and now only a fraction of datacentres that those chips are meant to fill are actually breaking ground, what happens to all the chips? Satya Nadella let it slip a couple quarters ago that the chips are not the bottleneck. The “warm bodies” to plug them into, are.

► So why do we have a chip bubble and are these chips really no longer cyclical? In our view the answer again comes back to the funding. Zuckerberg’s attitude to paraphrase is ‘it’s ok if we don’t use all our compute – we’ll just resell it’. That sort of attitude speaks to the lack of demand signal here. If this were a normal capex response to demand, there would be at least some visibility in the forward range of demand outcomes. This is a race between egos to build capacity defensively. And at the same time, in the post-GFC world, bank lending has to be collateralised.

► So in our admittedly sceptical, alternate world view, the demand for chips and projects rose until it exhausted the supply of free money, and it was financed in a secured, collateralised form that lenders could get on board with, and was flattering to the borrowers:

1. Lenders require guarantees to lend. The compute and chip purchase agreements had to be legally binding. At the same time, those making commitments to buy compute and chips want protection against the possibility that they can’t actually afford those purchases. So somehow, for the lenders the agreements had to be legally binding, while for the folks borrowing, they had to have a legal way out / not be. Bankruptcy-remote SPVs solve that problem for both.

1. The lenders also require security / collateral against the chips themselves, which wouldn’t be possible if they were capitalized into the customer holdcos as PP&E. The lenders - PC and banks - wanted favourable regulatory treatment on these loans. They wanted to call them loans to non-bank financials (NBFIs) rather than corporate loans because the capital treatment is >10x better. They were made to reclassify a lot of NBFI loans into corporate in two big chunks last year, for a total so far, of about $450bn.

1. The equipment / chip vendors were a party to many of the deals, because their customers are lossmaking start-ups with limited firepower. SPVs under minority interest accounting don’t need to be consolidated. A revenue transaction between a consolidated subsidiary cancels out on consolidation. Thus, chip vendors wouldn’t have the nice revenue boost without these structures. They also wanted to provide non-equity vendor financing, also with legal protection. It’s typical for these agreements to have 3 parties, and often one is the vendor or designer. The beauty of that is that no one party has more than 49%.

1. At the same time the chip vendors wanted the downside protection that comes from a ringfenced minority equity stake in an off-BS entity. The main company can’t be bankrupted by any related liabilities - legal, financial or otherwise. So voila chip vendors can finance their customers in what is effectively off-balance sheet triangular finance (this is the term used by Chinese bank analysts to describe that system especially under Hu Jintao, but pyramids are triangular so it’s extra appropriate here).

1. The hyperscalers and neoclouds want to avoid recognizing depreciation on the chips, equipment and datacenters. Housing the whole thing off balance sheet in non-consolidated JVs achieves that. Classifying it all as lease obligations that don’t start for years is also helpful. The D&A of an SPV never hits the ultimate owner’s profits.

USA EARNINGS EXCEPTIONALISM: An Illusion from Debt-Funded Capex

One of the most pernicious things about capex booms is that they create the illusion of profitability, at the time. This is always true in a capex boom, but heavy use of the structures detailed above makes it all the more potent. It’s hard to know exactly how much this dynamic (delayed or ringfenced depreciation) has inflated index earnings because of the opacity of these funding structures, but our estimate is something in the range of 15% by now. All capex cycles are optically very positive for both index earnings and margins, because they increase index sales (by capex providers), while those doing the spending amortise the spending over the useful life of the investment so only a fraction shows up in index costs. With the AI spend, what this means is that people like Nvidia, the utilities and other suppliers are getting AI-related revenue, even though the costs mostly aren’t showing up in the bottom lines of the hyperscalers, because these costs are treated as capex (or held out of sight in SPVs). The second key point about so-called US earnings exceptionalism, is that fiscal deficits are almost always good for margins because they give households income to spend (generating corporate revenues), but without that income needing to come from employment (which would be a cost to companies). So once again we arrive at the conclusion that this part of the narrative is also a product of monetised fiscal spending. At least in the early part of the cycle it was real cash earnings (the fiscal deficits were real), but the capex portion is a non-cash mirage.

Make no bones about it – the trailing earnings multiple is still historically high even with this inflation in profits. But look at the price-to-sales and the price-to-cashflow! Earnings quality is deteriorating to new lows this cycle. Free cash as a share of profits is just 75%, the growth is all non-cash, and despite “strong earnings” there’s been an inexorable rise in net debt per share. If so profitable, why so much debt growth? Call us old-fashioned, but in a “productivity boom”, by definition companies produce more with less, so income growth must naturally outpace debt.

MEMORY PRICES: Actually Just Cyclical, Like the Dot Com & 2022?

► Even if funding were ample, and this may come as a shock to the terminally online, the real physical world is out of stuff for these companies to buy in their pursuit of ever-expanding capex. The data centres are lagging because there’s not enough energy, grid connectivity, chips, turbines, specialty equipment, or funding to build what has been announced. The grid backlog is over three years long. There’s a second reason, and that’s, to paraphrase a recent conversation TOTEM MACRO had with a believable party, that “datacentres are polling lower than cancer in the United States”. The pushback, the class action lawsuits…all of that is gathering pace. As and when this capex growth slows down – and to be clear it doesn’t need to shrink, the pace just needs to slow – the dreaded D&A wall catches up. That part is just math, and was a big factor in the profit decline in 2022 and the Dot Com. And yet, that’s exactly what the datacentre construction chart a few pages back already shows is going on. Just as the margin-funded equity blowoff indicates the funding constraint, the memory price blow-off reflects the physical supply constraint. Taiwanese chip volumes are up a mere 2% year-over-year, memory volumes hit capacity, and the price move reflects that crunch. The broader commodity blowoff reflects overheating writ large.

► Yet something perturbs us. Why is Korean chip production falling? If demand is so insatiable, and if the margins are upwards of 75-80%, and if we were commodity memory chip producers, we would certainly be maxing out our production to capture those margins. There’s an argument which we can’t disprove but are highly sceptical of, that suggests that the pivot of DRAM capacity to much more wafer-intensive HBM capacity mechanically causes real IP to decline. This would be inconsistent with how any other data agency globally calculates real IP and doesn’t really pass the sniff test. But even if true, it certainly wouldn’t explain the sequential drop (to well below expectations) that happened month-over-month. Hynix recently announced they were actually pivoting capacity back to commodity DRAM since the margins there are higher than the more AI-centric HBM line – which says something unflattering about the relative strength of AI demand. So this rationale makes no sense from a sequential standpoint at the very least. The second point is that most customers order contracted amounts that generate a volume peak in the first quarter. But again, if demand is so insatiable, Korean vendors would simply keep production as tight as possible and sell into the spot market. Incidentally, the three main memory providers are being sued by class-action DRAM consumers on the allegation that they intentionally squeezed supply. We have no view on the veracity of that suit of course, but the near-15% fall in output doesn’t look great at first blush.

We arrive at a different plausible story. What if the entire memory price squeeze is a function of seasonal consumer tech demand hitting an AI-depleted market in the usual pre-holiday ramp and handset cycle that begins in September of every year? What if the memory shortage is no more structural than the chip shortage we faced in the first euphoric phase of the post-Covid bubble (2020-21) when “chips and ships” constrained global activity and drove up inflation? What if it’s no more structural than the memory squeeze that took place from late-1999 to the third quarter of 2000, similarly mooted to be a newly-secular growth industry, only to unwind sharply just a few months later?

As late as February 2000, capex and earnings were being revised up at an accelerating rate, there was a leverage-funded blow-off top in semis in Feb/March, and literally by late March/April, earnings were being sharply cut on write-downs, backlog shrinkage, and the inevitable D&A wall, because liquidity and spending were already tightening at the same time analyst and market euphoria was peaking. Semi sales then peaked in September of that year and declined sharply. This memory squeeze – bear in mind – was in commodity consumer DRAM, not HBM. And it also hit NAND – also not particularly AI-intensive. So if it is indeed true that the memory price squeeze was a seasonal artifact of holiday demand hitting a depleted market, and now we’ve frontloaded consumer tech demand during Covid, during tariff frontloading, and then during war-related stockpiling, and finally to front-run memory related price hikes…and if we’re now going into recession and the consumer is hitting a wall, the “memory shortage” and related pricing will unwind just as quickly and just as unexpectedly as it did in the Dot Com. A lot is riding on the bet that this time is different.

GLOBAL GROWTH IS LIKELY NOW TIPPING INTO RECESSION

It wasn’t just the AI bubble that generated spending without underlying cyclical demand. Consumers also pulled forward their demand, while producers have built inventories to front-run tariffs and war-related disruptions, which combined to generate two large blow-offs in global trade (last year and this year). There’s a variety of evidence that suggests this is stalling out, alongside end demand. Consumers brought forward goods demand, while allowing services spending to take the brunt, to get ahead of war, tech, and tariff-related inflation. But they put these balances on their credit cards in April and May, while recent lending data suggests they’re now back in deleveraging mode, where they were before the Trade and Iran Wars. Meantime, the war and the Supreme Court’s invalidation of tariffs – both of which happened in the final week of February, caused another spike in import activity on the part of both consumers and vendors. This meant that, similar to 2025, the year’s trade got pulled forward to the first half, because new tariffs are very likely to come into effect later in July. You can see this surge in the chart below, right. Secondly, as we’ve discussed before, trade-related borrowing is a key driver of deposit creation in the US. When trade finally turns down, deposits contract. This squeeze is what mechanically generates market deleveraging. Production data across Asia – at the very beginning of supply chains (especially in this cycle) – has already peaked and is generally declining.

Of course, trade contractions hit exporter economies before that weakness ripples across to importer economies. This is why Chinese data and growth surprises tend to lead the US, an especially pronounced feature of the post-Covid cycle, given the centrality of trade. The sort of liquidity contraction we’re seeing in markets also tends to correspond to the peak in global and local growth surprises, which generally look the strongest right at the onset of recession. On the front page we show how liquidity contractions – as proxied by flows to EM – lead major global growth deteriorations.

Here we’ll add a few more thoughts. China’s weak oil purchases are getting a lot of press for saving the day – but it looks to us like they don’t actually need the oil right now. Chinese data surprises look an awful like the gyrations in oil around the war, because as we pointed out above, the war and tariff-relief induced trade-related frontrunning that mechanically boosted the demand for oil and materials, and hence oil and commodity prices. Physical supply gaps were filled by SPR releases in the US, China and Japan (in all cases providing local income and liquidity relief by replacing oil imports with domestic supply). And while it is true that speculators short-covered a lot of oil positions at the start of the war, and have rebuilt them now, and those moves had some impact on the price, in our view the main impact was a temporary surge in goods trade and production, and demand weakness now.

The fact that the commodity weakness is so broad-based, and not confined to just oil, tells us that it’s a demand problem. It’s not really about Hormuz, directly, and probably never was. The oil weakness does, however, make a deflationary return to easing more likely again. But ironically, it will do so without offering much relief to US households or stalling consumer demand, because the blowout in crack spreads is keeping gas and distillate prices high even as oil prices fall. As you can see above right, there’s been a meaningful unwind in industrial metals, which also tends to lead US surprises, which makes sense. Here’s how growth-sensitive commodities looked in prior inflections:

Originally published for Totem Macro clients: July 10th, 2026

**

IMPORTANT DISCLOSURES AND DISCLAIMERS**

DISTRIBUTION. THIS REPORT AND THE INFORMATION IN IT IS PROVIDED FOR YOUR INTERNAL USE ONLY. THIS REPORT IS PROPRIETARY AND CONFIDENTIAL, AND MAY NOT BE REPRODUCED, DISTRIBUTED, OR DISCLOSED, IN WHOLE OR IN PART, IN ANY MANNER, WITHOUT THE PRIOR WRITTEN CONSENT OF TOTEM. THE INFORMATION IS PROTECTED UNDER THE APPLICABLE COPYRIGHT LAWS IN THE US AND OTHER COUNTRIES. UNAUTHORIZED RETRANSMISSION, REDISTRIBUTION OR OTHER REPRODUCTION OR MODIFICATION OF INFORMATION CONTAINED IN THIS REPORT IS PROHIBITED AND MAY BE A VIOLATION OF LAWS, INCLUDING TRADEMARK OR COPYRIGHT LAWS AND COULD SUBJECT THE USER TO LEGAL ACTION.

Data Sources. Totem Macro uses data and information from public, private, and internal sources. Sources include Haver Analytics and Bloomberg Finance L.P., as well as various central banks and government statistics agencies. While we consider information from external sources to be reliable, we do not assume responsibility for its accuracy.

Analyst Certification. The author who prepared this report hereby certifies that all of the views expressed in this report accurately reflect her/his personal views about the financial instruments and economic factors discussed herein at the time of publication. No part of her/his compensation was, is, or will be directly or indirectly related to the inclusion of specific recommendations or views in this report.

Conflicts of Interest. The authors of this report may have a position, material interest, or effect transactions in investment opportunities mentioned in this report. You should assume that as of the publication date of this report Totem’s principal, analysts, and employees, as well other clients of Totem’s, may have positions in investments related to or covered in this report, and therefore may stand to realize significant gains in the event that the price of the position rises or declines. These individuals may continue transacting in or otherwise change the form or substance of their position regardless of the views set out herein, and without notice to you. The author(s) of this report receive compensation based on a number of factors including the quality of research, client feedback, firm profitability and normal competitive factors.

Timeliness. Opinions in this report constitute the current views and opinion of the author as of the date of this report. All other information contained herein is believed to be current as of the date of this report unless indicated otherwise. Any opinions or views stated by Totem and all information in this report are subject to change without notice. Totem is not obligated to update or supplement any reports or any of the information, analysis and opinion contained in them. Totem is not obligated to update or supplement this report or to otherwise notify you in the event that any information, analysis or opinion set forth herein, changes or subsequently becomes inaccurate.

Disclaimer of Warranties. The information in this report is provided “as is”. We do not warrant the accuracy, adequacy, or completeness of this information and we expressly disclaim all endorsements, guarantees and warranties, express or implied, of any kind, including without limitation all warranties of merchantability, fitness for a particular purpose. Although the information provided in this report is obtained or compiled from sources believed reliable, Totem cannot guarantee the accuracy, quality, reliability, timeliness, completeness, or continued availability of any information or data made available. Moreover, Totem has relied upon and assumed without independent verification the accuracy and completeness of all information available from public sources.

Limitation on Liability. To the fullest extent permissible under applicable law, Totem will not be liable for any direct, indirect, consequential, incidental, special or punitive damages that may arise out of the use of this information by you or anyone else (including, but not limited to, lost profits, loss of opportunities, trading losses, and damages that may result from any inaccuracy or incompleteness of this information). To the fullest extent permitted by law, Totem will not be liable to you or anyone else for any loss under any tort, contract, negligence, strict liability, products liability, or other theory with respect to the information in this report.

No Offer or Solicitation. The information herein is provided for informational purposes only and should not be construed as an offer or solicitation to buy or sell any security, financial product, or other instrument. This information is not, and under no circumstances should it be construed as, an advertisement or offering of any of the securities described herein.

Not a Fiduciary or Registered Financial Company. Totem is not a fiduciary. Totem is an independent investment research provider and is not a member of FINRA or the SIPC. Totem is not a registered broker dealer. Totem relies upon the "publishers' exclusion" from the definition of investment adviser under Section 202(a)(11) of the Investment Advisers Act of 1940 and corresponding state securities laws, and is not a registered investment adviser either with U.S. Securities and Exchange Commission or with any state securities regulatory authority.

Not Investment Advice. This report is for informational and illustrative purposes only and should not be regarded as investment advice or as a recommendation regarding a course of action. Totem is not acting in a fiduciary or advisory capacity to you. This information is provided with the understanding that you will make your own independent decision with respect to any course of action in connection with the research herein and that you are capable of understanding and assessing the merits of a course of action and evaluating investment risks independently, or with the aid of other advisors.

Risk of Loss. Investing, including in the investments or opportunities discussed in this report, is speculative and involves substantial degrees of risk including the loss of the entire principal amount of any investment. You should not assume that any investment or opportunity discussed in this report will be profitable. Any discussion of risks in this report is not complete.

Past Performance is Not Indicative of Future Results. There can be no assurance that the future performance of any specific investment, investment strategy, or product referenced in this report will be profitable.

Not Suitable for All Investors. All information is general in nature and provided without regard to particular circumstances. These reports do not take into account the investment objectives, financial situation, or particular needs of any particular person. These reports do not provide all information material to an investor’s decision about whether or not to make any investment. As to whether any such course of action is appropriate or proper based is based on your own judgment and your specific circumstances and objectives; and may not be suitable or desirable for all investors. Investors should make their own investment decisions based upon their own financial objectives and financial resources and should seek advice from professional investment, legal, and/or tax advisors.

Not Legal, Accounting or Regulatory Advice. This report is not intended to represent the rendering of accounting, tax, legal or regulatory advice. A change in the facts or circumstances of any transaction could materially affect the accounting, tax, legal or regulatory treatment for that transaction. The ultimate responsibility for the decision on the appropriate application of accounting, tax, legal and regulatory treatment rests with the investor and his or her accountants, tax and regulatory counsel.

Forward Looking Statements. The information in this presentation contains forward-looking statements or projections regarding future events, targets or expectations. Forward-looking statements are based on assumptions and estimates of amounts not yet determinable, and therefore involve known and unknown risks, uncertainties and other factors which could cause actual results to differ materially from those in the forward-looking statements. Even slight differences or variations from the assumptions used can produce vastly different results. There is no guarantee that projections will be realized or achieved.

Not Replicated In Actual Portfolio. Data presented herein is unaudited and hypothetical and does not account for potential fees, transactions cost, or other outlays that may affect the ability of investors to execute the opportunities discussed in an actual portfolio. The information in this report also does not disclose or contemplate the risks, liquidity, hedging, or exit strategies that may need to be considered to properly execute the opportunities. No representation is made as to whether any of the positions herein can be successfully replicated in a portfolio.

© Totem Macro, LLC. All rights reserved