What We Told Hedge Funds This Week

SaaS wasn't dead; the labor market wasn't dead... but Anthropic will still IPO with that backdrop

Let us start with the software wars, because this week settled an argument. You will remember the Citrini “doom-porn” on software from earlier this year (late February to be exact). The SaaSpocalypse thesis: AI eats every seat-based business model, sell everything with a subscription attached. And to be fair, it had glimpses of truth to it: the lazy, undifferentiated seat-sellers ARE in trouble. But the market took the thesis, extrapolated it to everything with a login page, and marked the EXACT bottom in software assets doing so.

What is actually developing is something else entirely: AI is being deployed ON TOP of vertically integrated software, not instead of it. Look at this week’s quarterly results. CrowdStrike (arguably the most vertically integrated software piece on the planet... this is the firm whose sensor update grounded half the world’s airports a couple of years back, THAT is how deep in the stack they sit) beat and raised, with its AI detection ARR up 250% sequentially and Falcon launch partnerships signed with OpenAI, Anthropic AND Google. Salesforce, the poster child of the funeral notices, posted a monster beat with Agentforce ARR at 1.5bn and +240% YoY... the stock jumped 22% in a day. SentinelOne is grinding at its highs, and even UiPath, the market's favorite “AI roadkill”, keeps refusing to die. The pattern is the same everywhere: the software that OWNS a workflow becomes the distribution rail for the intelligence.

And even MORE importantly, when AI runs on top of existing tools, it INCREASES token demand. Every accounting system, every security platform, every CRM that bolts an agent onto its workflow is a new, recurring, enterprise-grade token buyer... the exact opposite of the hobbyist price-shopping that the doom-porn charts were built on. There is EVERY reason to be upbeat on token demand from here. As for Citrini: A remarkably solid EXACT contrarian signal, even if everyone (themselves included) thinks the timing has been great... and do not forget they abandoned the memory trade right when Leopold handed over the keys to Ken Griffin. The worst possible timing in a year full of candidates. But I am of course not “without guilt”. I toured with a similar “software is dead” narrative earlier in Q1 than them, but I have kept the vertically integrated names in high esteem on the above logic and made a killing there (CrowdStrike, SentinelOne, etc.)... So I am only half-embarrassed of that take (I think it has been more right than wrong)...

Now, with Anthropic potentially only 4–6 weeks from IPO’ing, I thought it made sense to look at the case from a MACRO perspective first... because the macro backdrop is quietly making the bull case for them. Ask yourself the question the productivity-deniers cannot answer: how can profit margins be THIS aggressively up, at all-time highs across the economy (not just in tech), when the statistical bureaus tell us there is no productivity? Of course this IS the productivity. Companies are selling more without hiring more.

Sure, the mass exodus of workers promised by the AI executives never materialized, but there is a widening GAP between hiring and profiting, and it can almost only be AI-related. And here is the pattern I keep repeating: when profits rise as aggressively as they do now, you do NOT see layoffs, even though you probably could lay people off without harming the business. You will see them once the profit cycle rolls over. THAT is when the productivity gets harvested for the next cycle.

I heard a beautiful example from a Scandinavian line manager recently: he had instructed his whole team to automate every task with AI, aggressively... use AI, use AI, use AI. Asked why he had not laid anyone off after they automated everything, it took him a while before he indirectly admitted that... well, it is actually a pretty nice thing to be a manager of 20 people. THAT is the labor market of 2026 in one anecdote. The automation is real, the incentives to shrink headcount are not... yet. (And give it a generation: the Gen Z crowd will soon be bragging in the bar about how many AGENTS they manage, not how many people.)

The measurement point matters too, by the way. If you hire a bureaucrat, GDP captures it neatly... the payslip IS the output. How do you measure an agent? The number of live agents in the OpenAI ecosystem has roughly quadrupled in two months; they are doing genuinely economic work, and the national accounts capture approximately none of it. So we have record profits, record capex, an allegedly sluggish economy and “no productivity”... or we have a measurement system built for payslips in a world increasingly run by agents. I know which explanation I am betting on.

So how solid a pillar does the whole buildout rest upon? There is genuinely GOOD news for the 2027 cycle after Bessent’s liquidity gymnastics. The buybacks, potentially funded straight out of the TGA, are the closest thing to QE this cycle has produced (see Monday’s editorial for the full mechanics)... and we get the first glimpse of it at the operation next week.

My out-of-consensus call: I think he SURPRISES the market by buying A LOT MORE (maybe 10bn+), not less. Everything this administration has done for a month points the same way... they are trying to prolong the cycle through the remaining Trump years, with the banking system re-installed as the marginal dollar printer, exactly like the mid-90s template. I do think the PMI cycle peaks in rate-of-change terms around November/December, but that does not make 2027 a negative growth year... we have simply not seen the kind of tightening that kills cycles, and the people in charge are visibly working on the opposite.

And then the Anthropic case itself. Sure, they (and OpenAI) are still catching up to their compute bills... that is the nature of this phase of the buildout, and I have made my peace with it.

But let us just remember what Anthropic has already achieved: NO company in recorded corporate history has reached this level of ARR this fast. Not Google, not Meta, not Microsoft, not even OpenAI... and that holds in inflation-adjusted dollars too, where it is not even close.

The latest investor deck ahead of the IPO has them approaching 70bn in annual recurring revenue... and 15-20 months ago they barely had a client. We charted it from the 1bn mark and forward, against the fastest revenue machines ever built. Look at the red line and remember that Google needed four YEARS to do around a quarter of what Anthropic did in twenty MONTHS.

The only thing that genuinely worries me in the case is the compute capacity. Look at the one sub-category in Anthropic’s accounts that has looked weak from a rate-of-change perspective: Claude Code revenue, flatlining since June... and it happened exactly WHEN they became constrained. That is not a demand story, and I do not buy for a second that Code sidelined itself due to the model or the price. It is the exact reason they lost momentum... and it is why they paid SpaceX 2–3x what everybody else pays for compute in that panic deal ahead of the summer.

They were simply too laid back on their compute needs while Dario was out boasting about how SENSIBLE they were being about compute... and then suddenly they were sitting on the hottest product the world has ever seen and could not supply it. A 2.8x premium for emergency megawatts is what that realization costs.

Which brings me to the slightly heretical conclusion for a house that loves the Anthropic story: I think there is a case to be made that OpenAI currently sits on the more interesting PIPELINE, and that Anthropic is rushing for the IPO precisely because it has lost a bit of momentum that it would rather not have become too obvious before listing.

You can feel the compute gap as a user: OpenAI and Gemini have visibly more inference capacity, which is why Claude is slower in responding... You can live with that since Claude is the better model, but if it was NOT the better model, it would be a very big issue.

The IPO at a fresh (2-trillion-ish) valuation is exactly the tool that fixes the constraint: equity plus cheap debt to buy compute at scale. So do not read me wrong... the demand for intelligence is the largest demand story for any single product ever produced, and BOTH of these companies will be dramatically larger in 2030. But if you are asking me which IPO I am more tempted to size up at the margin right now, the honest answer is the one with the super compute pipeline... and that is currently not the one listing first.

Warsh at Jackson Hole: the Fed just made itself SMALLER...

So, the speech. At first glance, a disappointment for those of us hoping he would dust off the full productivity gospel... he touched it, but he did not preach it. Read closer, though, and the touch was heavier than the headlines suggest. He opened by declaring a “hinge point in history”, described a “hyper-Moore’s law” playing out, and, remarkably, quoted the token sales of the two leading AI labs at more than 100bn annualized, up 500-plus percent in a year... the Fed chair reciting ARR figures from the Jackson Hole podium (try to imagine Volcker doing that). He then listed, almost verbatim, OUR bottleneck framework as the Fed’s open questions: will AI lift productivity and WHEN, will tokens complement or compete with labor, and how much of the surplus lands with the owners of scarce assets... “AI labs, chipmakers, energy producers and cloud providers”, in his words. The framework is fully installed in his head. He just refuses to trade on it yet.

And THAT refusal is the actual news, because he told us exactly why: the task force on productivity and jobs (one of FIVE task forces, including the ones rebuilding the inflation measurement) will report later, and “their recommendations will come later and have no bearing on decisions we make in the current policy conjuncture”. Translated from Fed-speak: nothing moves until the new measurement apparatus is in place. Which is precisely my base case, fully intact after Wyoming: Warsh stands pat until a NEW inflation target/measurement framework lands later this year or early next... and then the game changes. “I stand here today committed to a discipline, not to a decision” was the money quote of the whole speech.

The bigger structural message was the Fed deliberately making itself LESS important for markets. The entire middle section was a demolition of forward guidance (“you can call it an outline... just don’t call it forward guidance”), a warning about the hall-of-mirrors problem, and the flat statement that the Fed “should not indulge a regime in which market participants are looking primarily to the Fed for their next trade”. Fewer promises, quieter communication, more meeting-by-meeting discretion... that means MORE volatility around the prints and the meetings, and LESS Fed-centricity in between. For macro traders, honestly, that is a gift. And do not miss the monetarist nod either: “money matters”, including money created by the banking and financial systems... that is a green light, from the Fed chair himself, for the whole Bessent private-printer architecture we covered on Monday. The two of them are more aligned than the real-yield spat suggested, I guess.

Now the part that matters most for the inflation trade: notice what he did NOT do. He gave essentially zero credit to the CPI softness. His words: “while this summer’s PCE and CPI readings were better than expected, they do not tell me that underlying trends have meaningfully improved”... and every number he leaned on was a PCE number: 3.7% on 12 months, 4.1% annualized on six, 54% of the basket running above 3%. The PCE, the LESS soft of the two indices, remains THE variable until the task forces are done. He even re-anchored it explicitly: the 2% target “as measured by the PCE price index, is a firm, fixed target”.

Which makes our analysis from Wednesday twice as important now, so let me repeat the most important takeaways. The record-wide core spread between the PCE and the CPI is index plumbing: a portfolio-fee echo (financial services running +7.55% YoY in the PCE, largely IMPUTED from the level of the stock market... the S&P at highs literally shows up as “inflation” in the Fed’s preferred gauge), PPI-fed health care priced off what providers receive rather than what patients pay, and a shelter weighting only half the CPI’s, so the melting shelter glacier helps the CPI mechanically more.

And the spread closes DOWNWARDS: the CPI does the steering historically, the predicted PCE from matching CPI categories sits around 2.9% versus the observed 3.7%, the portfolio-fee leg fades the moment equities merely go sideways, and the PPI pipeline is already turning with the cracks rolling over. Savor the irony: Warsh’s hawkish anchor is partly INFLATED by the very stock market he watches for “market internals”... and when the PCE folds toward the CPI over the autumn, the last data argument for standing hawkish folds with it. But a hike between now and the end of the year is not ruled out... but probably just one... And I didn’t get my productivity gospel, admittedly...

In focus: Nvidia is BACK... but they are handing over margins to the memory companies

The quarterly result itself was a monster, and the market treated it accordingly: revenue of 96.2bn (+106%), data center at 89bn (+117%), and the first 100bn+ quarterly guide in corporate history at 108bn... with ZERO China compute assumed, so any summit outcome is free optionality.

The stock added roughly 435bn of market cap in a single session, Vera Rubin is in full production with purchase orders from every hyperscaler and already ~20% of next quarter’s data center revenue, and AWS signed up for an additional 2 MILLION GPUs through early 2029. Nvidia is clearly BACK after this report... no arguments from me.

But the still undertold story of the week sits one layer below the headline, and it is the one we have been hammering all month: Nvidia is handing its margins to the MEMORY companies. Follow the receipts.

Saturday: Bloomberg reports 15%+ price increases to customers, varying by memory configuration.

Wednesday: supply commitments explode from 119bn to 279bn in a single quarter, and management says out loud that it is “mostly memory” for the Vera Rubin ramp... while guiding Q4 gross margins DOWN to 71–72% on rising memory costs before “stabilizing” in FY28.

Thursday: the Korea August NAND export price prints 99,573 dollars per kilo.

The largest chip company on earth spent an entire week confirming, in print, that the scarce input in this cycle is memory... and that it is willing to surrender 3–4 points of the best gross margin in the industry to secure it.

Every one of those surrendered margin points lands in the pockets of the memory vendors, and our SanDisk/Micron sleeve remains the cleanest way to stand at that toll booth. Still the undertold story... even the sell-side spent the week celebrating the guide instead of reading the commitments footnote.

Portfolio Update: Short USD sleeve working. Next catalyst is probably the China summit and the buybacks pace...

It's been a solid week for the software sleeve. CrowdStrike performing well on the back of earnings, SentinelOne still trending up, and UiPath is also rebounding as we speak. Even Salesforce (not part of the portfolio) is back! This is basically the earnings season wiping the floor with the "software is dead" narrative. I am still of the view that stuff that is already vertically integrated will be AMPLIFIED by AI (CrowdStrike a perfect example), while it will be very difficult to launch new SaaS at high price points due to AI. Alongside this, we are also seeing the Crypto Utility case rebounding, and we are thinking about whether to load up ahead of the 4-year cycle pattern (which will certainly attract a crowd for that very reason in Q4 and early next year).

The short USD trade has overall done well for us since putting it on during the early innings of July, and our data keeps pointing in that direction, albeit not as clearly as 1-1.5 months ago. We have seen a rebound in nuclear, metals, Crypto Utility and everything "soft USD proxy"... and we continue to hold a high conviction there.

We are also of the view that the “hardware scepticism” will be put to bed by the Nvidia results, and even if memory is still licking its wounds slightly from the Leopold-gate in July, we think the numbers MORE than add up... and we see the US/China summit as a potentially strong catalyst for the semiconductor space, especially the US-listed names.

Remember the setup: the Korean licenses for the China fabs expire December 31, the rare-earth control suspension expires in November, and BOTH sit on the summit table... which hands the US-listed memory names a rare two-way bet. A renewal keeps the boom untouched, a non-renewal removes an eighth of world memory supply from the competition, and either outcome gets sold as a win in the US. Meanwhile, the Korea export data says the semiconductor complex trades 35-40% too cheap versus its own fundamentals... just make sure your exposure is on the right side of the Pacific, because the Korean-listed names (and the ETFs stuffed with them) carry the tail risk that SanDisk and Micron collect on.

We have updated our convictions accordingly, and see strong cases in the above themes. We have approximately 23% of the portfolio in the "short USD trades" currently, but may be tempted to add into Q4, especially if Bessent really ramps up his buybacks from next week, which is increasingly likely. Warsh did not truly support those bets, so it hinges on Bessent now.

In terms of macro calls: we remain short oil (this zone is a good one to short from), and short crack spreads as well... with long airlines (or long miners) as the equity expression of that trade. We like memory and rare earths into the China summit, and we like everything short-USD still. the AUD is probably a good expression of the USD trade as well, given its beta to both the metals and the Chinese impulse.

Highest convictions by asset class

In FX: short USD, unchanged and reinforced. the transatlantic inflation gap keeps widening, the positioning washout is incomplete, and Warsh just told markets to stop trading the Fed. Maybe we should trade Bessent's buybacks instead? EURUSD towards 1.2050, with the AUD as the secondary leg.

In Rates: received in the US front end, but it is not a clear trade currently. The market prices hikes into a stand-pat chair awaiting his own new inflation measurement... but I don't think we can rule out a hike between now and that new measure, so we size it as a lean rather than a conviction.

In Commodities: long the debasement sleeve (gold, Bitcoin, uranium) with the TGA-funded buyback flow as the running tailwind, short oil from this zone, and short crack spreads... still both the trade AND the key that eventually unlocks the front end.

In Equities: the memory complex on top (SanDisk, then Micron, tilted US over Korea into the licence cliff), the AI-on-top software rail (CrowdStrike, SentinelOne, UiPath) has earned its sleeve back, keep Wolfspeed as the power-wave option, and Definium into Panorama. And whichever AI lab you prefer at IPO... own the toll booths they BOTH must pay: memory, power, and the racks.

Have a great weekend, and good luck out there!

横向滑动查看完整图表

Andreas Steno

Andreas is the core macro consultant in Steno Research. Building on years of experience as Global Chief Strategist at Nordea Bank, he is one of the most quoted and sought-after macro analysts out there. Andreas’ expertise is the FX/rates, energy, real estate and equity spaces, but doesn't shy away from hot takes on other topics if the underlying analysis is strong enough. Andreas anchors the weekly editorial ‘Steno Signals’.