What We Told Hedge Funds This Week
The ECB cannot print molecules either...
A quarter or two ago, the line was “the Fed cannot print molecules”. The past month or two, the molecules moved continents: the genuinely tight energy market is European Nat Gas, where storage sits 25 points below seasonal norms and EU law forces price-insensitive buying into the hole. Including: the USD update (the trade now runs on the GROWTH channel, not the inflation spread), what Waller and Williams just told us about September, TWO In Focus sections... the gas squeeze (with the Romanian trades) and the OpenAI comeback via GPT-6 Astra... plus a salty portfolio note on the “peak AI” pundits who have re-emerged exactly on schedule...
The ECB cannot print molecules either...
Let us start with the scoreboard, because that is how we run this letter. The short USD trade has faced some troubles over the past week. Energy prices refused to play ball, Warsh handed the spec USD longs a lifeboat in Wyoming a week ago, and the dollar has traded firm against almost everything since, at least until Waller came to the rescue of the trade again...
But the more important update is the REFRAMING: the trade’s engine has changed. The original case was an inflation-spread story... US disinflation versus European reflation. That leg got muddier this week, with our US rising-inflation probability ripping from 7% to 41% in a month (energy-driven) while the Eurozone sits at a screaming 84.5%. Both sides rising = a less clean divergence. The trade from here is a GROWTH story instead: the US is the ONLY major economy on our dashboard with a falling growth probability (28%, from 41% in June), and our new lead-lag work says the economic surprise index follows the nowcast impulse with a ~6-week lag... positive surprises today, softer ones into the autumn.
Softer US data relative to the world is the classic dollar-negative impulse, and it needs no Fed cut to work... only US data losing its exceptionalism at the margin. Same trade, new engine, humble sizing... and the ammunition reloads at Bessent’s buyback operation next week, where I keep the 10bn-per-operation call on the table.
And speaking of the nowcasts earning their keep: the biggest analytical finding of the week is that the US GROWTH nowcast has quietly been one of the best leading indicators in the world. Push it two months forward, and it traces the momentum equity index through this entire cycle... the March peak in the nowcast preceded the June 2 top in momentum, and the June-July roll preceded the correction we have lived through since (momentum is -16% and bad on ALL scores since early July).
To be clear, this is NOT a fresh “sell momentum” call... the correction already happened, and it rhymed with our data almost to the week. The read from here: the washout is largely behind us in level terms, but the old goldilocks leadership does not return until the growth nowcast BASES. It has not based yet.
Now contextualize the central bank picture with what Waller and Williams handed us this week, because it defanged the post-Warsh hike panic considerably. Both restated their summer reaction functions with a DATE attached: Waller holds if the August CPI (September 11) shows inflation cooling and hikes “if it reverses”... Williams is encouraged and willing to sit steady. Our August core nowcast at 13-15bps clears BOTH bars comfortably.
So the swing votes stand pat in September if our data on inflation is right (also given today’s NFP number, which we were spot on in calling way above consensus), whatever Warsh’s “more work to do” rhetoric suggested in Wyoming... and the RBNZ just previewed the template for what a move would look like from the big central banks: a dovish hike with a magnifying-glass track. So we think September will be soft (relative to expectations), but with a US inflation probability at 41% and climbing, the risk is clearly RISING that the doves cave in later this year if the energy impulse leaks into the core prints. September is safe on our numbers. Q4 is a live fight... and it is exactly why the products deal out of Hormuz (or a Russian refining recovery) remains the single most important macro catalyst on the board.
On oil, we keep the contrarian lean: the front month rolled INTO the war, the DATED physical price sits galaxies below the March/April stress, and escalation has been a SELL signal all year in this sine-wave conflict. Short the barrel, short the cracks... It doesn’t sound like Trump is willing to massively escalate this into the mid-terms.
In Focus I: The European Nat Gas squeeze... and how we trade it in equities
Now to the market that actually deserves the fear. The TTF day-ahead price is accelerating almost daily, and the mechanics underneath are brutally simple: Europe does not have the storage where it needs it, and the LAW forces the buying regardless of price. German storage sits around 52% full versus a ~77% seasonal norm for the date... a 25-point gap with roughly two months of injection season left.
The EU aggregate is in the low 60s versus an ~82% norm, the LOWEST level for the date since records began in 2011. Brussels has already blinked once, relaxing the November 1 target from 90% to 80% (plus 4 points of flexibility)... and even the relaxed target requires injections of roughly 4,400 GWh per DAY from here, while the Qatari LNG that normally balances the European autumn is stuck behind the Hormuz mine-clearance timeline until early Q4.
A legally mandated, price-insensitive buyer forced into a tight market on a deadline... that is not a market with a lot of participants on the short side of the trade. The consolation: ~76% is likely sufficient in a NORMAL winter per the INES modelling, so what is being priced is a cold-winter tail risk. But tail risks with legal deadlines attached typically get bought/speculated.
How to express it in equities if the squeeze continues?
Two ideas, one admittedly off the beaten path. First, the ROMANIAN angle, which I am genuinely tempted to add (but haven’t found a venue that allows me to): OMV Petrom and Romgaz, the 50/50 partners behind Neptun Deep in the Black Sea... the only major NEW gas development anywhere in the EU, first gas targeted for 2027, roughly doubling Romania’s production and making the country the largest gas producer in the European Union. The EU just legislated itself into being a forced buyer of Nat Gas every autumn; the Russian pipes are gone for good, Qatari flows are hostage to a strait... and here is a domestic EU producer bringing the only meaningful new supply online, trading at Bucharest multiples rather than Brent multiples, and paying you to wait. The risks are the usual Romanian ones (windfall-tax creativity, a government that treats energy as a fiscal piggy bank)... size accordingly.
Second, the “boring” blue-chip expression: Equinor, the largest single supplier of gas to Europe post-Russia, with cash flow overwhelmingly linked to exactly the TTF and NBP prices that are accelerating. (Honourable mention if Hormuz clears in Q4: the LNG shippers, where ton-miles explode the moment the strait reopens... different trigger, same complex.)
In Focus II: OpenAI is BACK... and that is bullish for the compute cycle
The second In Focus this week, because the AI duopoly just flipped again.
OpenAI launched GPT-6 Astra, and the independent benchmarks say it is not marketing fluff: on Perplexity’s WANDR agentic benchmark, Astra scores 0.682 at $11.98 per task... clearing Anthropic’s Fable 5.1 (0.601 at $12.76) on BOTH axes, better AND cheaper, and sitting alone at the top of the Pareto frontier. OpenAI has reclaimed the marketing momentum, the developer chatter, and the front pages... and the interesting question is not “is the model good” (it is), but WHY OpenAI dares to launch aggressively and invite a usage explosion right now.
The answer is the one we have been hammering for a month or two: they are NOT capacity constrained... and Anthropic is. Remember the receipts: Claude Code revenue flatlined around the 15bn mark from June, exactly when Anthropic hit the compute wall... they paid SpaceX 2-3x market rate (50m per MW-year) for emergency capacity in the summer panic deal... their secured power pipeline runs meaningfully below OpenAI’s through 2027-28... and they are rushing an IPO precisely to buy their way out of the constraint.
When you cannot serve the demand you already have, you do not launch demand-generating products. When you CAN, you launch Astra. The capacity gap has become a product-velocity gap, exactly as we argued it would.
And before anyone dismisses the benchmark as cherry-picking: the independent capability trackers agree. On Epoch’s Capabilities Index, Astra sets a new RECORD at 169 (90% CI 165-174)... not just above Claude Fable 5’s 160-161, but slightly ABOVE the reasoning-era trendline of +15 ECI points per year.
Read that carefully, because it kills two bear cases at once: the scaling line is NOT bending (so the “diminishing returns” crowd waits another year), and the lab that everyone said was falling behind on compute just shipped the most capable model ever measured. Capability progress intact + capacity to serve it = the usage flywheel spins on.
And here is the equity read, because it is genuinely tradeable. The market has spent the last 6-9 months selling the “OpenAI bets”... Nvidia, SoftBank, Microsoft... relative to the “Google bets” (Alphabet, Broadcom) on the bailout-tweet noise and the TPU success story. Look at the chart: the OpenAI complex has lagged the Google complex by a wide margin since the spring... and Astra is exactly the kind of catalyst that closes that gap from below.
The OpenAI supply chain is broader than the big three, by the way: Oracle and the Stargate consortium, CoreWeave as the flexible-capacity provider, AMD with the multi-gigawatt accelerator deal, Broadcom on the custom-silicon co-design (yes, it sits in both camps), Arm via the SoftBank nexus, and the Korean memory names via the Stargate MOUs. A re-accelerating OpenAI is a purchase order for ALL of it.
And do not miss the macro importance of that purchase order: the lab with the BEST compute pipeline just retook the product lead, which is precisely what fuels the next leg of compute buying into next year. Look at the construction data if you doubt how load-bearing this is: data-centre construction is up 50bn versus late 2023, while ALL OTHER private construction is down ~125bn... the AI capex line is single-handedly holding up the entire US construction cycle.
An OpenAI that presses the accelerator is not just a tech story. It is THE cycle story.
Portfolio update: The “peak AI” pundits are back... right on schedule
Before the numbers, allow me a moment of saltiness, because the “we are past peak AI” chorus has re-emerged this week with remarkable confidence.
Exhibit A: a viral post from Citrini Research declaring “the bottleneck trade is over... I don’t think there are any new bottlenecks left to discover”. With respect: announcing that all bottlenecks have been discovered is the analytical equivalent of the 1899 patent-office chief declaring everything inventible already invented... and even if it WERE true, “a rotation of capital between the bottlenecks we’ve already found” is not the end of the trade; it IS the trade. Compute, then memory, then power... we have been trading that rotation for the past years and getting paid for it.
Exhibit B is the better one though: “what if peak AI expectations are now behind us?”... illustrated with a chart of ANTHROPIC’s implied valuation falling from 1.4 trillion to ~850bn. This is clearly written by someone not in the know about the shift of momentum WITHIN the duopoly (soon to become more than a Duopoly btw...). The Anthropic drawdown is not “peak AI”... it is a capacity-constrained lab losing product velocity to a rival that just launched the best model on the market. Reading it as the end of AI is like calling the death of the smartphone because Nokia lost share. The demand for intelligence did not peak... it MIGRATED to the lab(s) that can actually serve it.
And one more pattern for the sceptics: AI sentiment shows genuine SEASONALITY, and we are in it. Almost exactly a year ago, the same peak-AI chorus sang the same song into September... followed by the usual pick-up into year-end and early Q1, when enterprise budgets reset, the December compute crunch arrives, and the labs ship their headline releases. Astra just fired the starting gun on that seasonal pattern two weeks early. We fade the chorus, again.
The hardware vs software reversal... we think the anti-hardware sentiment is about to PEAK
And the pundit chorus above connects directly to the most interesting positioning signal on our screens right now: the options market has completely stopped paying up for hardware.
The ratio of semiconductor to software implied volatility (SMH vs IGV, 1-month 25-delta calls) peaked at 2.14 during the hardware euphoria of late 2024... and now sits at 1.03. PARITY. The market is pricing semis and software as equally boring/entertaining, after two years where every incremental AI dollar was assumed to land in silicon.
That is what peak anti-hardware sentiment looks like in cold, tradable numbers... and sentiment extremes in this ratio have marked the hardware-vs-software turning points throughout the cycle. Notice the irony versus this week’s tape: the SOFTWARE names just had their vindication week (CrowdStrike, Salesforce), the hardware names carry the “peak AI” discount... and the trigger for the reversal is sitting in plain sight: OpenAI’s success. The market spent months doubting OpenAI’s compute pipeline (the bailout-tweet episode, the Google TPU scare)... and Astra just demonstrated that the pipeline DELIVERS. A re-accelerating OpenAI with the largest secured compute base in the industry is a purchase order for the entire hardware complex, into a market that has priced hardware optionality at zero premium. We like that setup enormously.
Two dark horses to watch on this reversal.
First: how the market treats the ANTHROPIC IPO now that they are visibly losing momentum into it. A softly priced or poorly received listing will be waved around as “peak AI” evidence by the same pundits above... even though a capacity-constrained lab repricing lower while its rival accelerates is a SHARE story, not a demand story. Expect noise; position through it.
Second: how the market treats ALPHABET, with Gemini currently not in the race at the frontier (Astra and Fable 5.1 have left it behind on the agentic benchmarks). The “Google bets” have outperformed the OpenAI complex for months on the TPU story... if the model race narrative flips back to OpenAI, that relative trade unwinds from BOTH ends, and the rotation lands exactly where our trading book sits.
Now the book.
Very good performance overall since last Friday, which we are satisfied with given that the biggest macro lean (short USD) misfired and the market obsessed over energy.
The resilience came from the right places: the memory names have done perfectly OK despite everything... Micron keeps compounding (+246% on entry), SanDisk stabilized after the poorly-timed June entry, and the licence-cliff optionality into the summit costs nothing to hold. Memory demand is set by the AI capex cycle and Nvidia’s 279bn of supply commitments, not by the TTF curve... and the Astra launch only strengthens that: a re-accelerating OpenAI with the best compute pipeline is incremental demand for every toll booth we own. It is also why I keep arguing that US TECH is quietly one of the best trades available if markets keep obsessing over energy prices: the US runs on cheap domestic gas while Europe pays the squeeze, and the AI complex is the energy WINNER via our electricity sleeve rather than an energy victim.
The sleeves that WITHSTAND higher energy prices remain most of the book by design: Electricity (Bloom +132%, NextPower, Enphase, Stem, Hammond) is a direct beneficiary of expensive molecules, Energy Fuels carries the uranium-plus-rare-earths angle, Barrick and the Decoupling sleeve hold the debasement hedge now that China is potentially gearing up it’s threat ahead of the September summit, Steel Dynamics (+43%) rides reshoring capex, and the End-Consumer Scarcity trio (Dell +107%, HP +47%, Apple) sells dollar-priced hardware regardless of the gas curve. Definium remains the biggest single position (13.9%), with Panorama this month as the binary (we have high expectations)... and it could not care less about gas prices, which is its own kind of hedge. Year-to-date scoreboard for transparency: +9% after +31.3% in 2025 and +72.6% in 2024... a pedestrian year so far, but still very good through this cycle...
Trades and themes across asset classes
In FX: short USD as a GROWTH-channel lean... the US is the only major economy with a falling growth probability, the surprise index rolls into autumn on our lead-lag work, and Bessent reloads the ammunition at next week’s buyback operation (we still say 10bn per op). EURUSD towards 1.2050, AUD as the secondary leg. Humble sizing until the energy impulse fades.
In Rates: received in the US front end. The swing votes told us the trigger is the CPI on September 11, and our core nowcast at 13-15bps clears both the Waller and Williams bars.... Honest caveat: the Q4 cave-in risk is rising with the 41% US inflation probability, so this is a September trade first, a Q4 trade only if the core prints keep cooperating.
In Commodities: short flat oil from this zone and short the crack spreads (escalation remains a SELL signal in the sine-wave war, and either a Hormuz products deal or a Russian refining recovery releases the disinflation wave)... long the debasement sleeve (gold, Bitcoin, uranium) with the TGA-funded buyback flow as the running tailwind. In Nat Gas, we do not chase TTF directly... we express it in equities below.
In Equities: the memory complex on top (SanDisk, then Micron, tilted US over Korea into the licence cliff and the summit)... the electricity sleeve as the energy-squeeze WINNER... the AI-on-top software rail (CrowdStrike, SentinelOne) keeps its earned sleeve... Definium into Panorama as the binary of the month... the Romanian gas pair (OMV Petrom + Romgaz) on the watchlist as the Nat Gas trade... and a fresh eye on the lagging OpenAI complex (Oracle, CoreWeave, AMD, SoftBank nexus) plus the hardware-vs-software reversal, where the implied-vol parity says the entry is HERE. Wolfspeed stays as the 2027-30 power-wave option. We accordingly UP our conviction on several of the hardware names in the book, here, while urging a little caution on software (in relative terms).
Have a great weekend, and good luck out there!
