What We Told Hedge Funds
A weak USD via a detour? What funding markets need
Two inflation prints, both hotter than our nowcast, and a market that responded to the second one by BUYING bonds and SELLING the dollar. That is not what the textbook says happens on the biggest PPI since 2022; rather, it’s what happens when positioning is stretched enough that the data stops mattering, and it is the first evidence that the extremes we have highlighted for a month, in the Mag-7, in the dollar, in oil, and in bonds, are starting to unwind. Herein: being wrong on macro and what it taught me about positioning, the detour the dollar may be taking, an in-focus on hyperscalers versus neoclouds versus hardware with the IRR study and the balance sheets, and what we need to see for the bullish setup to return.
A short word on being wrong, because I have been…
We nowcasted the September CPI mildly below consensus on core, and it came in at 0.33%, which was hot. We said the PPI would not confirm that, and then it came in at 1.3% on the headline and 0.7% on core, the hottest since March 2022. A reader on X summed up the quarter: I used to be fat and retarded, and now I am fat, retarded, and WRONG.
For the record, I am no longer as fat: one or two gym sessions a day since August has a curious narrowing effect which you can see for yourselves at the next client dinner. The other two charges are ones I can only beat by being right again. But here is the thing about being wrong on macro while being right on positioning: the market has just told you which one it trades on.
The dollar fell on the hottest PPI in four years. Bonds rallied. Small caps held. And by Thursday afternoon the AI trade itself had turned: the chip rout that took the Kospi down 7.6% at the open reversed as yields tumbled, the neoclouds closed higher after Jensen Huang’s comments on demand, and the memory names whipsawed to a flat close after a week in which they lost a trillion dollars of market value between them. The things everyone owns, the Mag-7, the dollar, the crude longs, the bond shorts, were the things that moved against their owners. I was wrong on the print, and the print did not matter by the end of the day. That is the whole conclusion, and I will try to explain why it is good news rather than an excuse.
Macro: A Weaker USD Via Detour?
Let us start with the PPI, which was in many ways worse than the CPI. Final demand rose 1.3% on the month against 0.5% expected, 6.3% on the year, with goods up 2.6%, the largest increase in the series’ history, energy up 10.8%, and core up 0.7% and 5.0% on the year. Two caveats before the conclusion. We ONLY nowcast the CPI, not the PPI, so the PPI call on Thursday was analysis rather than model output; and I am not claiming we were right on either.
The conclusion is the one we have drawn before: a WIDER PPI-CPI spread is not good news for companies with weak CPI pricing power, because their input costs are rising faster than the prices they can charge. It is a margin squeeze for the consumer-facing economy, and it is still a signal to buy whatever is CAUSING the inflation rather than whatever is absorbing it. That has been the right trade since March, and it is the right trade now.
The interesting thing is the market response, because it was a pretty classic position squaring across assets. The ten-year yield FELL as much as 10 basis points on the day to 5.08%, a three-week low, after touching 5.30% on the CPI the day before. The dollar WEAKENED. Small caps were flat to up while the popular large-cap risk-on names sold off. October hike odds went from near 50% on Wednesday to about 30% by Thursday. None of that is what a 1.3% PPI is supposed to produce.
It reminds me a little of October 2022, when the equity market RIPPED higher on a monster hot CPI print, and everyone spent a week explaining why. The explanation then was the same as now: when the data is as bad as it can be and the market goes the other way, the data is not driving the market; positioning is. And positioning was monster short bonds, monster long dollars, 51% long crude, and at an all-time-high net Mag-7 exposure going into the week.
There is something going on when stuff moves the other way from the way it is supposed to, and it is almost always telling you where the crowd is.
Look at the bond positioning first, because it is the largest of the four extremes and the one that moved most. The CFTC data shows the speculative community short more than two million contracts in the 2-year note against a long book roughly a third of the size, the most lopsided reading in the history of the series. That is a market that has bet on the Fed hiking into the inversion, and it is a market that loses money every day the front end rallies. A 1.3% PPI should have been the best day of the year for that position. Instead, the 2-year fell three basis points and the 10-year ten.
When the most crowded short on the board cannot make money on the best possible news for it, the short has begun to cover, and covering is what drives the next leg down in yields regardless of what the next print says.
The dollar is the same story a week later. Net speculative USD longs peaked in late September at the most crowded level since 2022, and they have begun to trim; the dollar index made its high for the year on the CPI, on Wednesday, and could not hold it on the PPI, on Thursday.
The chart below is the one I use for the relationship between the cross and the positioning, and the honest read is that EURUSD is still sitting below where the positioning says it should, which means there is room for the dollar to fall even without a change in the fundamentals. That is the detour: the dollar does not need a soft inflation print to weaken; it needs the longs to stop being added to — and they have.
Which is where the detour comes in. The fundamental case for a weaker dollar has not changed: the US is the softest economy on the board on growth, the ECB and the BoE have firmer paths, and the inflation that the Fed is hiking on is energy and two lagged CPI lines.
But the direct route to a weaker dollar, soft US data leading to a Fed pause, has been closed by two hot prints. The detour runs through positioning: a market this short bonds and this long dollars squares up on its own weight, bonds rally, real yields fall, and the dollar follows the bonds down whatever the inflation print says. That is probably what paves the way for a slightly better funding environment, which is an interesting schism in itself: the worse the inflation data, the more the bond shorts cover, the lower the yields, and the better the financing backdrop for the one sector that needs it. Maybe bonds become more constructive here. If they do, it is the thing AI needs most for the 2027 continuation, and it arrives via the ugliest possible route.
The two charts that follow are on why the inflation impulse is closer to its peak than the prints suggest. The necessity basket, input costs for food, energy, and chemicals, are up around 25% on the year, which sounds bad until you put 2022 next to it, when the same basket was up 100%, and the ISM fell to the low 40s. The impulse leads the ISM by fourteen months, and has already rolled — and this one is a third of 2022’s.
And financial conditions lead the inflation impulse by two quarters: they tightened through spring and summer, and the lagged inflation impulse has followed them up into October, which is exactly what we are living through. But conditions stopped tightening months ago, which puts the peak in the impulse right about now. The live core readings, Truflation and the rest, still sit well below the BLS. Two hot prints do not change the direction. They change the timing, and I have been wrong on the timing twice.
The second chart is the one I would ask you to hold on to through the next two inflation prints, because it tells you when the heat ends rather than how hot it is. Financial conditions — the composite of rates, spreads, the dollar and equities — lead the inflation IMPULSE, the second derivative of the CPI, by about two quarters, and they have done so cleanly through 2008, 2021-22 and the 2023 unwind. Conditions tightened from March through August on the hikes, the curve, and the dollar, and the lagged inflation impulse has followed them up into exactly the prints we have just had. On a two-quarter lead that puts the peak of the impulse in October or November and the turn in the first quarter.
And the live data has not changed its mind. Truflation’s daily core series, which has led the official BLS core lower for the whole of this year, sits around 1.7% against a BLS core at 2.6%, and the other live aggregators we track sit in the same place.
Two hot official prints have not moved the live readings, because the live readings are measuring rents and used cars as they are transacted today while the BLS is measuring them as they were transacted months ago. The spread between the two is the largest it has been since the series began, and it closes in one of two ways: the live data catches up to the BLS, which has not happened once in three years, or the BLS catches up to the live data, which is what has happened every time.
In Focus: Hyperscalers, Neoclouds and Hardware Providers: Who Holds the Bag, and Who Should?
Markets have, at least until the past few days, bought the Mag-7 and sold their suppliers, both the neoclouds that rent them capacity and the hardware makers that fill it. Goldman’s prime book shows net hedge-fund exposure to the seven at an all-time high, 22% of total US exposure, while the picks and shovels have been shorted. Does that make sense? Let me take the three layers one at a time, because the answer is different for each.
Start with the hyperscalers, because they are the ones who actually hold the bag. Between them, Oracle, Microsoft, Amazon, and Alphabet carry about $1.2 trillion of remaining performance obligations, revenue their customers have contracted for and they have committed to deliver: Oracle around $455bn, Microsoft around $370bn, AWS around $195bn, Google Cloud around $155bn.
That is the whole point about who is exposed if things really turn south. The hyperscalers have sold the compute forward. They need the capacity to exist, on time, from somebody, or the obligations become liabilities. The market has decided that makes them safe, because they have the cash flow to build it themselves if the neoclouds cannot. I think it makes them committed, which is a different thing. A company with $370bn of obligations and a bond trading at 66 cents does not cut capex; it finds a cheaper way to fund capex, and renting from somebody else’s balance sheet is the cheapest way there is.
But what I am actually more interested in, going into the reporting season on the 28th, is how the executives see the INVESTMENT CASE for data centers, because I think they will have concluded that despite all the funding stress, the case has improved rather than deteriorated. Here’s why…
The rental price of compute has not fallen. The Blackwell index, the market price for renting a B200 or B300 GPU in the US, sits at $5.29 an hour, up from around $4 at the start of the year, with a high of $5.30 in May. Every study, every sell-side model, and every bear thesis assumes that cloud prices COME DOWN sharply from here, because that is what happened in every previous compute cycle. It may happen. It is not what we are seeing, and it has not been what we have seen for a year.
Demand for inference has absorbed every GPU that has come online at flat to rising prices. An executive looking at that curve and at his own backlog is not looking at a worse investment case than a year ago. He is looking at a better one with a higher cost of capital, which is the IRR question, and I did the arithmetic.
The study. Take one Blackwell GPU rented at the index price, 85% utilization, which is conservative for contracted capacity, and an all-in capex of about $75,000 per GPU: the chip, the server and networking around it, and the data-center shell and power to house it. That gives about $39,000 of revenue per GPU-year against roughly $8,000 of operating cost, over a five-year life with a small residual.
On FLAT rental prices, the unlevered project IRR is about 32%. If prices fall 10% a year, 23%. If they fall 20% a year, 12%. If they fall 30% a year, which is what the bears assume and what the previous cycles delivered, the IRR is about 1%, which is to say the project destroys value at any cost of capital. And then the scenario nobody models because it is not supposed to happen: rental prices RISING 10% a year, which is LESS than what the Blackwell index has done over the past twelve months, gives an IRR above 40%.
I include it not as a forecast but because it is the only scenario consistent with the price data we actually have, and because there are two revenue sources on top of it that the entire sell side is ignoring. First, the LLM providers have not even begun to monetize advertising. The consumer AI assistants have hundreds of millions of users and no ad load; the day that changes, the revenue per GPU-hour of inference goes up without the rental price moving, and the pressure on the buyers of compute is to buy MORE of it, not less. Second, the one visible instance of an AI lab’s revenue growth slowing this year, Anthropic’s, was a CAPACITY problem, not a sales or demand problem: the company could not get the inference capacity to serve the demand it had, which is the whole reason the neoclouds exist and the whole reason the Blackwell index does not fall. A business whose growth is constrained by its supplier’s capacity is not a business whose supplier should be priced for a demand collapse.
Now put the funding costs on the same chart: investment-grade hyperscaler debt around 6%, GPU-backed term loans around 9%, and CoreWeave’s unsecured bonds at 15% this week. The picture is clear. At flat or gently falling prices, which is the world we are in, a data center earns 20-30% on capital against single-digit funding costs, and the spread is enormous; it is the best infrastructure project on earth. At a 30% annual price decline, it earns nothing, and no funding cost is low enough. The entire debate between the bulls and the bears on neoclouds is a debate about the slope of that line, and the market has priced the bear slope into the suppliers while pricing the bull slope into the hyperscalers who own the obligations. One of those is wrong.
Which brings me to the chart that answers the “is AI demand slowing” question with numbers rather than vibes, and I want to use it carefully.
Put the two largest labs together. Anthropic’s annualized revenue has gone from $9bn in December to $74bn in September; OpenAI’s from $21bn to $50bn on its own conservative basis. Combined, from $30bn to $124bn in nine months. And look at the DOLLAR growth per month in the third quarter: $11bn in June, $11bn in July, $11bn in August, $12bn in September. The combined growth of the two, in dollar terms, is still SOLID; it has been running at a steady eleven to twelve billion of new annualized revenue every month for four months, with no sign of rolling.
Now look at the split, because this is where the bears found their story. Anthropic’s monthly growth rate slowed from 17% in June to 6.5% in September, and that slowdown was reported as demand fading. It was not. From June onwards, Anthropic was CAPACITY constrained: it could not get the inference compute to serve the demand it had, which anyone who used the product over the summer experienced directly. OpenAI, which had the capacity, accelerated over the same months, from 9% monthly growth in July to 19% in August and 16% in September, and the combined dollar number did not budge.
That is what a supply-constrained market looks like: the growth moves to whoever has the capacity, and the total keeps climbing. It is the single best argument for why the Blackwell index does not fall and why the neoclouds exist, and it is on this page because the market has spent a week pricing the opposite.
Now the neoclouds themselves, which is where the market has placed the whole cost of the funding stress. The structure is the one we have described: they sold (parts) of their revenue forward on fixed-price, multi-year contracts to the hyperscalers, but they have not locked in the cost of building the capacity to deliver it. The build-out is funded in the debt and equity markets, and both just got more expensive: a hot core puts a second Fed hike back on the table, the ten-year is above 5%, the hyperscalers’ own long bonds trade in the 60s, and the private-credit data-center deals price off the same curve. Fixed revenue, floating cost, and the cost just moved. That is a potential dilution story, and the market is pricing it: if you cannot borrow for the next gigawatt, you issue shares for it, and the existing shareholders pay. CoreWeave proved the point on Wednesday by announcing a 35-million-share at-the-market programme and a $3bn convertible on the day its bonds hit 15%.
But they are not all the same, and the table is the point. CoreWeave is the leased giant: $104bn of backlog delivered from 49 data centers it mostly does not own, $35.6bn of debt, net debt at 2.9 times annualized revenue, interest expense five times its adjusted operating income, and the bonds at 15%. It has locked in the most revenue and the least cost, and it funds the gap in the debt market every quarter. Nebius is the opposite: $46-49bn of backlog anchored by Microsoft and Meta, more than 75% of its capacity in data centers it owns, about $8bn of cash against debt that is almost entirely convertible, net debt at 0.2 times revenue, and, crucially, customers who PREPAY 50-60% of the build cost on its landmark deals, which means the funding comes from the counterparty rather than the bond market.
Cipher is a landlord rather than an operator: $8.5bn of contracts with AWS and Fluidstack, where Google backstops the Fluidstack lease so the landlord can borrow against it; it owns the shells and leases them out, which is the lowest-risk structure of the four and also the lowest return. And IREN, the one we own, sits in between on funding but has done two things the others have not. It has sold the LEAST forward, $9.7bn to Microsoft and a smaller contract with Nvidia against 4.5 GW of secured grid power, which means it can lock in new deals at today’s higher rental prices rather than last year’s; and it owns its power, which is the asset nobody can buy quickly. Its debt is around $7.8bn, its cash around $2.8bn, and its FY27 capex of $25-30bn is the funding question. I think management has been decently prudent in not selling too much forward, and I think the market is punishing that prudence as if it were CoreWeave’s leverage.
Which leads to the trade. The neoclouds will be THE buys if the funding stress alleviates even a little, and the route to that is the one in section one: a position squaring in bonds that brings the ten-year back below 5% and the GPU-loan curve with it. The IRR study tells you why the leverage works both ways: at a 9% funding cost and a flat price path, the equity in these businesses is worth multiples of today’s price; at 15% and a falling path, it is worth nothing. The market is priced for the second; the data is consistent with the first; and the trigger is the bond market, not the hyperscalers. We hold IREN; we do not add yet, and we add the day the ten-year closes below 5% with the hyperscalers having confirmed 2027 capex. Both could happen in Q4.
Then the hardware providers, and specifically memory, where I want to make two arguments the bears are not making. First, the VOLUME case. The Korea chart below is the one that matters: Korea has exported fewer chips in volume terms at MUCH higher prices, which is what a shortage looks like. The bear case for 2027-28 is that volume comes back as the new fabs land and prices fall. Fine; but what if volumes are up and prices only MODERATE?
SK Hynix plans to lift HBM wafer starts 30-40% in 2027, Samsung 40-50%, Micron 60-80%, and the binding constraint on all of it is ASML, which is expanding EUV capacity 30% a year. If 2028 volume is 50% above 2026 and prices are 20% lower (which there are NO signs of yet), revenue is 20% HIGHER, and the contracts have floors. That is not the scenario priced at 4-5x earnings. That is a scenario the market would pay 15x for. The question is not whether prices moderate; it is whether volume growth outruns the moderation, and the capacity plans say it does. The same volume logic runs through the rest of the chip stack, and the clearest example in our book is Wolfspeed: the only scaled US producer of silicon carbide, the power semiconductor that sits between the grid and the rack, now funded by a conditional $1.5bn thirty-year commitment from the Department of War to build out domestic capacity.
How much volume, concretely? The three memory makers have published their plans.
SK Hynix runs around 160,000 HBM wafer starts a month and plans to add 30-40% in 2027 at Cheongju and the new M15X fab. Samsung runs around 120,000 and plans 40-50%, most of it at Pyeongtaek, contingent on its 12-high yields finally clearing. Micron runs around 55,000 and plans 60-80%, the fastest growth from the smallest base, in Boise and Hiroshima. Add it up and the industry is targeting roughly 40% more HBM wafer starts in 2027 than in 2026, on top of conventional DRAM capacity that is growing in the teens.
The binding constraint on every one of those plans is the same: EUV lithography tools from ASML, which has raised its guidance, is expanding capacity 30% a year, and still cannot ship machines as fast as the fabs want them. Which is why ASML is the chokepoint, not the fabs, and why the volume arrives on ASML’s schedule rather than on the memory makers’ ambitions.
Meanwhile, while the market debates 2028, the spot market is doing something in 2026 that the bear case says it should not. DDR5 spot prices, the conventional DRAM that goes into servers and PCs rather than the HBM that goes into accelerators, have re-accelerated over the past two months after a summer pause, and are at their highs of the cycle. That matters for two reasons. It is the memory that CXMT competes in, so if Chinese supply were already weighing on prices, you would see it here first, and you do not. And it is the memory that the inference remix needs most, server DRAM for the CPUs that orchestrate agents, which means the price signal is coming from the demand side the bears have not modelled rather than the training side they have.
Second, the China case, which deserves a precise answer because it comes in two very different versions. The version about CXMT in HBM is utter nonsense, and the reason stems from yields. HBM is a stack of twelve DRAM dies bonded on top of each other, and the yield of the stack is the per-die yield compounded twelve times: a 95% per-die yield gives you a 54% stack, a 99% per-die yield gives you 89%. SK Hynix runs HBM3E 12-high at around 80% and says HBM4 is approaching the same maturity. Micron is in the 70s and has the cleanest ramp of the cohort. Samsung has struggled to get above the 60s and spent eighteen months failing Nvidia’s qualification on exactly this. CXMT is shipping HBM2 samples, targets HBM3 in 2026 and HBM3E in 2027, two full generations behind, with stack yields that the industry puts around 30%.
Nobody buys 30%-yield HBM for a frontier accelerator; the economics do not work at any price. That is not a 2027 story or a 2028 story. The version that IS real is conventional DRAM: CXMT is at 300,000 wafers a month with DDR5 yields at 80% and targeting 90%, and that is a genuine addition to commodity DRAM supply in 2027-28. It pressures the low end of the DRAM market, the PC and phone bit, and it is why we own the names with the highest HBM and server-DRAM mix rather than the commodity exposure. Added Chinese RAM is a reason to be selective within memory. But it is not a reason to be out of it.
Portfolio Watch: What do we need to see for markets to return to a bullish setup?
Two things, and both are visible in the past 48 hours without yet being confirmed. A weaker dollar and slightly calmer funding markets. We are watching the trends since the inflation reports closely, and the honest status is that they have started but not finished. The ten-year is 20 basis points off its high, the dollar is off its best level since May 2025, and the two-year bond shorts have started to cover. That is the detour in section one beginning to work. It is not yet a trend.
The thing I watch most is the GROWTH parameter in our nowcasts, because it tends to lead momentum in the risk-on names in the US, semiconductors included, by one to two months. It has done so through this entire cycle: the March low in the growth probability preceded the April ripping rally by six weeks, and the summer peak preceded the autumn fade by about the same. The US rising-growth probability fell from 27% to 19% through September and has shown the first signs of stabilization and a rebound in the rate of change this week. Until that rebound is confirmed, my March/April analogy does not hold fully.
The POSITIONING reminds me of March: retail at the lows, crude at a record long, the dollar crowded, breadth the narrowest since 2000. But in March the growth probability was already turning when the positioning was at its extreme, and today it is only beginning to. We need better momentum for the analogy to materialize, and the next two weeks of nowcast data decide whether it does.
On the dollar, it is a choppy call between a USD-bullish and a USD-bearish regime, and I will not pretend otherwise. The regime engine has flipped three times in five weeks and sits in Lame Duck QE today, which is dollar-positive, after the hot core. The fundamental case says the dollar falls: softest growth on the board, Fed one hike from done on a headline that financial conditions have already dealt with, firmer central banks abroad.
The energy and inflation channel says the dollar holds. What tips it, slightly, towards weakness returning, is positioning: the longs are crowded, they have started to trim, and the dollar fell on the day it should have rallied. The triggers, in order: the PCE on the 29th, which inherits the airfare and health-care heat from the PPI and will be hot, so the trigger is the market’s reaction rather than the print; the Fed on the 28th, where one hike with no-rush language is dollar-negative, and two is not; a Hormuz products deal, which collapses the energy premium that has carried the dollar since July; and the hyperscaler capex guidance the same week, which decides whether the funding stress eases or deepens.
The trading book, as it stands: Short USD as a lean, small, unchanged; we would add on a ten-year close below 5% with the growth nowcast confirmed higher, and we cut it on a second Fed hike being delivered rather than priced. Long semiconductors and memory in full; we do not add after the week we have had, and we do not trim either, because the volume case and the yield case above are the reasons we own them. IREN held, not added, with the trigger stated above. Uranium held as the hedge that works in the regime the rest of the book does not, and first in line for a trim if the dollar sleeve has to shrink. Short crude from a 51% long, long the cracks. And we are looking at agriculture as a 2027 hedge, not yet a position.
Trades and themes across asset classes
In FX: short USD as a lean, small. A choppy regime call tipped towards dollar weakness by positioning. Triggers: the Fed on the 28th, the PCE reaction on the 29th, Hormuz. Add below 5.00% on the ten-year; cut on a delivered second hike.
In Rates: the position squaring in bonds is the detour to a weaker dollar and a better funding market. Constructive on funding before duration; a ten-year below 5% is the number that matters for everything in the AI sleeve.
In Commodities: short crude into a record long, long the cracks. Gold holds in both regimes. Agriculture under the microscope for 2027.
In Equities: the hyperscalers hold $1.2tn of obligations and will confirm capex; the neoclouds are THE buy when funding eases, IREN first; memory on the volume case, selective within it on the China DRAM case; hardware over software until the growth nowcast turns.
Have a great weekend, and good luck out there! And if you see a slightly less fat Dane in the gym twice a day, that is me working on one of the three.