What We Told Hedge Funds
with Andreas Steno
The market is pricing a collapse of the AI build-out… so it is either collapse or a face-ripping rally from here.
In a month where we estimate that more than a hundred macro PMs lost their seats, the oil market priced a new bombing campaign, the semiconductor index posted its longest losing streak since 2008 and then bounced, Meta raised its capex and then fell 7%, and the ten-year dropped below 5% on the same day tech sold off. Underneath it, the AI market has moved to pricing in NO growth from here, while Anthropic cannot get capacity and the price of compute rallies… Herein: the oil market and the knife, Korea and the semis trade, why the productivity is already here, an In-Focus on the data-center delays and whether chips are piling up in warehouses, and a portfolio update on staying alive through a de-grossing.
Theme 1: The physical oil market
Theme 1: The physical oil market is not particularly tight, but the risk is perceived as huge by traders…
Good morning from a desk that is, like most of yours, lighter on P&L than it was a month ago. I will take the themes in order and keep the paragraphs short, because this has been a week for reading quickly.
Start with the spread that tells you what kind of oil market this is. Brent trades at $111-115, WTI at $99-101, and the gap between them, normally $3-5, is $13. In the spring it touched $15. That is not a shortage spread; it is a POLICY spread. Brent prices a Gulf where tankers are hitting mines and flows fell to around 8.5 mb/d last week from more than 20 the week before. WTI prices a United States that has decided to keep its own barrels.
That second part is the thing to understand. The serve-our-own-population thinking that started with Trump’s threat of a diesel export ban in September has become the operating model of the energy war: the Russia diesel deal, under which Putin ships product to American and global markets in exchange for sanctions waivers, is the same logic applied to a different population.
Every government in the chain is now managing its domestic pump price first and the global market second, and the Brent-WTI spread is the price of that fragmentation. More conflict, more regionalization, wider spreads. It does not close on a ceasefire; it closes on the strait.
Chart of the week #1 - The Brent-WTI spread at $13… three times normal, and it is a POLICY spread
Chart 1: Gulf crude flows fell to ~8.5 mb/d last week… from 20+ the week before. The perceived risk is the premium
Now the physical gauges, which still say what they said two weeks ago. The Houston export spread, the cleanest measure of an international scramble for American barrels, is around $1 against $7 in March; nobody is bidding up the marginal US cargo. Chinese state refiners are running at 73% because China is importing LESS, not more; the largest seaborne buyer on earth is not competing for barrels. The crude market is not particularly tight.
What is tight is the perception of risk, and the perception is specific: traders are pricing a new US bombing campaign on Iran, and Brent puts hit a record volume this week on exactly that.
Chart 2: The US export spread at one dollar against seven in March… no international scramble for American barrels
Chart 3: Chinese state refiners at 73%… China is importing LESS crude, which is not what a shortage looks like
Here is where I come out, and it is a view I hold with some sympathy for the man holding the knife. A new bombing campaign is now probably the base case; it has to be, for everyone, after Trump denied the talks on Wednesday and said the war would be over fairly soon, which is not a sentence you say about a negotiation. But look at it from Tehran. The Iranians are getting increasingly panicky.
They have offered the Americans peace deal after peace deal, the seven-day ceasefire, the phased reopening, the indirect track through Pakistan and Qatar that Trump now says does not exist. They are now out of money — on the schedule we laid out in September — and the Revolutionary Guard cannot pay their troops to do any fighting.
A counterparty that is out of money and offering deals is a counterparty you can get a better deal from by pressing, and I think that is exactly what Trump expects: one last campaign to get the knife in, a much weaker Iran at the table, and a settlement on his terms before the midterms are far behind him. I have sympathy for the view.
It is also, for the oil market, the better outcome of the two, because a campaign that ends the war ends the mine-laying, and the mine-laying is the whole premium. The trade is unchanged: short crude from a 51% long into a balance that is fine, long the cracks until the strait and the Russian refineries give them back, and do not confuse the two.
Theme 2: The AI build-out
Theme 2: The market is now pricing a collapse of the AI build-out… So it’s either collapse or a face-ripping rally from here. Nothing in between.
Let me state what the market is pricing, because it is a stronger statement than most people realize. The semiconductor index fell eight sessions in a row, its longest streak since 2008, lost 15.7% over ten days, and six memory makers shed a trillion dollars of market value. The memory names trade at 7-10x run-rate earnings against a market on 22x. On any reasonable model, that multiple on that earnings base prices NO growth from here. Not slower growth. None. The AI market has moved to pricing that the build-out is over.
Now put that next to the data from the same fortnight. Korea reported exports for October 1-20 up 105% year-on-year to $61.6bn, with semiconductors up 222% to $27.5bn, now 44.7% of everything Korea sells abroad, the highest daily export average ever recorded for any 20-day period. And that is in a won that is 14% stronger than in May, which understates the dollar growth rather than flattering it.
Korean exports have led US forward earnings by six months for 25 years, and pushed forward, they say US EPS grows 40-50% into the second quarter of next year.
Chart of the week #2 - Korea, October 1-20… exports +105%, chips +222%, a record on every measure, with a STRONGER won
Chart 4: Korea is a solid indicator for semis… and on the correlation, they are now more than 150% too cheap
And then the two facts that make the no-growth assumption VERY odd. First, Anthropic is capacity constrained. Its prospectus says so in its own words: it sought more capacity than the hyperscalers would supply and is building around them with third parties, which is what the $750bn is for.
Its growth slowed from June because it could not serve the demand it had, and OpenAI, which had the capacity, took the growth; the combined run-rate kept adding $11-12bn a month. Second, the price of compute is rallying like crazy. The Blackwell rental index is up 35% this year at $5.29 an hour, and the H100, the previous generation that every bear model has written down to scrap, is RISING with it.
There is life in the old dog yet.
Chart 5: There is still life in the old dog… H100 rents are rising with the B200, which no bear model allows for
I understand the funding stress. It is real, and it has to be dealt with: the ten-year above 5%, the neocloud bonds at 15%, Meta’s free cash flow negative and a new share class to fund the capex, credit spreads at their widest since May 2025. A hiking cycle reprices everything built on debt, and the AI build-out is built on debt.
But the analogy the market is reaching for, the dotcom bust, is SO far off that it needs saying plainly. In 2002 roughly 3% of the installed fiber was lit and bandwidth prices fell 90%. Nobody used the capacity, and the price of the output collapsed. In 2026, contracted GPU capacity runs near full, the labs are rationing their customers, and the price of the output is UP 35%. That is not under-utilization; it is OVER-utilization of compute.
The dotcom build-out failed on the two numbers that matter, and this one passes both.
Chart 6: The dotcom analogy fails on the two numbers that matter… 3% of fiber lit vs 90% of GPUs, bandwidth -90% vs rents +35%
Which is why there is nothing in between from here. Either the build-out actually collapses, in which case the obligations become liabilities and the current prices are right, or it does not, in which case the earnings arrive on the Korean schedule and a market that has priced zero growth has to re-price to something, and the something is a face-ripping rally.
Korea exported fewer chips at much higher prices this year; what happens if volumes rise in 2028-30 and prices only moderate? Probably still a good business, and certainly not a 7x one. The hyperscalers report on the 28th and 29th with $1.2tn of obligations and a template from Meta, which raised the capex floor. I know which outcome I expect.
I also know the market is positioned for the other, and that is Theme 3 and the portfolio section.
Chart 7: Korea has exported FEWER chips at MUCH higher prices… what if volumes rise in 2028-30 and prices only moderate?
Theme 3: The margins
Theme 3: If this is not productivity, what is? The margins are already telling you
The bear case has a second leg beyond funding, and it is the one I hear most from thoughtful people: fine, the capacity is used, but where is the productivity? Where does the AI show up in the economy rather than in the P&L of the people selling it? Here is one answer, and it is on a chart that has nothing to do with Nvidia.
US corporate profit margins are at 15%, a record, and they have gone UP through a hiking cycle, an energy shock, and a tariff war. Margins are supposed to compress when input costs rise, and the cost of money goes up. They have expanded. If this is not productivity, I do not know what is.
Chart 8: If this is not productivity, what is? Profit margins keep going up… through a hiking cycle and an energy shock
The historical comparison helps here, and I will give you three, because none of them fits on its own. The dotcom cycle, as above, fails on utilization and price.
The railway mania of the 1840s is the closest on SCALE: Britain spent around 7% of GDP a year at the peak, against about 2% of US GDP for the AI build-out today, so this is twice dotcom and a third of the railways; the railways also produced a crash, and then a century of productivity.
And the electrification of American industry in the 1920s is the closest on SEQUENCE: the power was built in a decade, and the factory floors took thirty years to reorganize around it, because the gains came from redesigning the work, not from the electricity itself.
That is the stage we are in. The infrastructure phase is nearly done; the bottleneck phase, memory and power, is where we are, and the deployment phase, agents and inference at scale, is where the margins chart says the gains are already starting to land. The market reads late- cycle because the price action is late-cycle. The capacity data says 1997 and the margin data says the electrification of the 1920s. Neither says collapse.
Chart 9: How big is this build-out? Twice dotcom, a third of the railways… and the only one where the output price ROSE
Chart 10: Where are we in the AI cycle? The bottleneck phase, not the commoditization phase
In focus: The chip glut question
In focus: Is there a GLUT of chips lying around due to delays? Not really… The scarcity is real…
The best question I had this month came from a reader: if grid queues run four years and transformers take three, the hyperscalers physically cannot finish what they have announced. So does the hardware sit stranded in warehouses and force a pause? I spent two days on it, and here is the condensed answer.
The delays are real and worse than reported. Of about 16 GW of US data-center capacity announced for 2026, around 5 GW was under active construction by the summer; a third to a half of this year’s promised capacity will slip into 2027 and beyond. A data center takes 12-24 months to build; a grid connection in a major US market takes 36-84 months to build; transformers run 160 weeks; heavy-duty gas turbines are sold out into 2029.
And the queue is fiction: 700 GW of data-center grid requests for a sector that draws 49 GW, most of it duplicate and speculative filings, which is why Texas froze new connections for an audit.
Chart 11: The constipation factor… of 16 GW announced for 2026, around 5 GW is actually being built
The response is already on the ground. The industry stopped waiting for the grid eighteen months ago: more than 130 GW of behind-the-meter generation is proposed to serve US data centers, 80% of it being gas, with 17 GW of dedicated gas already under construction and fuel cells deploying in ninety days where turbines take years. Mobile generation is bridging sites like Abilene while the transmission catches up.
A parallel power system is being built at a 15-25% premium to grid power, because a gigawatt of GPUs sitting dark is far more expensive than a gigawatt of on-site gas.
Chart 12: The answer so far: build your own grid… 130 GW of on-site power announced, 80% of it gas
And nothing is stranded, for a structural reason: GPUs are ordered against POWER, not against buildings. The accelerator supply chain runs 12-18 months from wafer to rack, the deployment cycle is under 18 months, and Nvidia allocates against confirmed power rather than press releases.
When a site slips, the hardware goes to a site that did not: densified into existing campuses, where liquid cooling has lifted rack densities from 20 kW to 100-130 kW; re-routed to Texas, the Gulf or the Nordics; or energized on the behind-the-meter bridge. Delayed energization is, from the compute market’s point of view, a SUPPLY shortfall, and it is a large part of why rents are up 35%. There is no glut of chips lying around.
The scarcity is real, and the delays make it more real, not less.
One more thing the map says: the buildings are American and the chips are Asian. The US hosts 45% of the global AI data-center pipeline and makes about 12% of the AI chip stack by value; Taiwan and Korea make 70% of the chips between them. The constraints are different, the beneficiaries are different, and the geopolitical risk sits on the half the market is not pricing while it prices grid queues. Own the power, own the chips, be careful with the shells.
Chart 13: The buildings are American, the chips are Asian… 10,000 km apart, with different constraints
Portfolio update
Portfolio update: It hurts, but remember to stay alive through these things…
The book is down 2.6% on the year, from +11% a month ago, after +31% in 2025 and +73% in 2024. I will not dress it up, and I will not pretend the hurt is only ours. There is a LOT of blood on the street.
By our count, more than a hundred macro PMs have lost their seats at the multi-manager platforms this month alone, and we can see it in real time: team after team being kicked off the Bloomberg chats as their drawdown limits trip. Goldman’s prime desk called last Thursday one of the broadest hedge-fund drawdowns since Covid, with short leverage at the 99th percentile over five years.
That is a de-grossing, and a de-grossing has one property worth remembering: it sells what it owns and covers what it is short, regardless of what either is worth.
Chart 14: There is a LOT of blood on the street… more than 100 macro PMs out this month, by our count
Which is why the positioning story is now fitting together across assets, with one exception. In bonds, the two-year short, two million contracts at the peak, has started to cover, and the ten- year fell 14 basis points below 5% on Wednesday, the day tech sold off, which is the detour we described a week ago arriving.
In the dollar, the longs peaked in late September and have trimmed; the dollar failed at its high on the PPI and has not recovered it. In equities, the levered market is net SHORT Nasdaq futures, the most since the spring, and the Goldman data says the shorts are at a five- year extreme.
The exception is oil, where speculators are still 51% net long into a balance that the physical gauges say is fine, and that is why crude is the one short we keep.
Chart 15: Markets are already monster short bonds… and the shorts have started to cover
Chart 16: USD longs have become a little less crowded… and the dollar failed at its high
Chart 17: The levered market is net SHORT Nasdaq futures… the setup for a squeeze, as in April and May
Chart 18: The oil market is very long again… the one place positioning has NOT squared, and the one short we keep
Read the equity positioning with the April-May episode in mind. The last time the levered market was this short Nasdaq futures was the week before the first Iran ceasefire headline, and the index rallied 25% in six weeks on nothing more than the absence of bad news.
A short squeeze is clearly a possibility from here, and the bar for triggering it is low: a hyperscaler guiding 2027 capex up, a Fed sentence acknowledging tighter conditions, a Hormuz headline, any of them does it. We do not need good news. We need a little benign news into a market that is short and de-grossed.
And then our own trigger, which I will be specific about. We are on growth watch in the nowcast data. The US rising-growth probability leads momentum in US risk assets by one to two months; the March low in that series preceded the April rally by six weeks, and the July peak preceded the autumn fade. It fell to 19% in September.
It is at 23% now, and the rate of change, the ten-day change in the probability, has printed its first green bars this week after two months of red. That is the early stage of a turn, not a confirmed one. But the rule has held through this entire cycle: as soon as the rate of change in US growth ticks up, the market rallies again. We are watching it daily, and the book is positioned for it.
Chart 19: Growth watch… the rate of change in the US growth nowcast is turning, early stages
Chart 20: Regime Model USA… growth at 23%, inflation above neutral, liquidity ample. Lame Duck, for now
The book, position by position in brief. Memory and semis held in full; the gap to the Korean data is the widest it has been, and we would rather own it than explain why we sold it. The neocloud held, not added, with the add trigger, the ten-year below 5% and capex confirmed, half met.
The power leg held; the power semiconductor name is down 37% on entry and the state has just underwritten its capacity, which is why conviction stays at 10. Uranium, down 49%, held as the hedge for the regime the rest of the book does not like, and the first thing we trim if the dollar sleeve has to shrink.
The end-consumer scarcity names, security software, and the biotech position are what is holding the book up, and they are doing it well. No changes this week. The job in a de-grossing is to stay alive through it, keep the positions that the data supports at a size that survives the positioning, and be there when the rate of change turns. We are.
Chart 21: Our portfolio composition… updated shortly after the open on October 23
Trades and themes across asset classes
In FX: short USD as a lean, small, unchanged. Positioning has turned, the fundamentals have not changed, the trigger is the growth nowcast confirming. Cut on a delivered second hike.
In Rates: the position squaring in bonds arrived on Wednesday via the detour. Ten-year below 5% is the number; constructive on funding before duration.
In Commodities: short crude into the one positioning extreme that has not squared, long the cracks until the strait reopens. A bombing campaign that ends the war is bearish the barrel, not bullish.
In Equities: the market prices a collapse of the AI build-out; the data prices a face-ripping rally. Nothing in between. Held in full into the hyperscalers on the 28th and 29th, with the levered market short and the growth nowcast turning.
Have a great weekend, stay alive out there, and good luck!
