Jeff Keller, who runs the technology-focused hedge fund Capelight Partners after operating stints at Salesforce and MongoDB, takes Max Wiethe through the third-quarter split that has software rallying while semiconductors fall. They work through what the market is paying for and what it is refusing to pay for, why Keller reads the broken deal-announcement feedback loop as bullish, how he treats hyperscaler capex as an option purchase rather than a return calculation, and what would have to break before he would short the AI trade.
👤 Guest: Jeff Keller, founder and portfolio manager of Capelight Partners, a technology sector focused hedge fund he has run since 2021, after operating roles at Salesforce and MongoDB
🎙️ Host: Max Wiethe, who interviews fund managers about the business of running other people's money
📰 Published: 31 August 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 1 hr 2 min | ✅ Time saved: 29 min
Key Takeaways
The third-quarter split between software and semiconductors is normal volatility inside a trend that has not broken
Wiethe's own numbers: software and services up 20% since the quarter began, semiconductors and semiconductor equipment down 7.5%
"I don't think that late June was the end of this mega trend. I think this is just kind of normal volatility."
Long-term uncertainty is so high that anything with short-term certainty is trading at a huge premium
Cyber and infrastructure software have the visible numbers, and large names in them are being paid 25-30 times revenue
He says the multiples there are loftier than in AI itself, and that the labs expect to disrupt cyber too
Databases are now chosen by models rather than people, and infrastructure software collects the volume either way
Every application carries a database underneath it, and AI multiplies the number of applications
Usage-based names can spike quickly in the short term; the long run he calls more of an open question
In technology, being in the right theme makes more money than picking inside it
The migration to cloud paid the sector before it paid the individual winners
Dispersion arrives as the open-ended growth story slows, which is where AI is now
The low-multiple AI names should beat the high-multiple ones, because they all trade off the same two drivers
Hyperscaler capex and the ARR of the labs drive the whole complex, at wildly different multiples
Nvidia and Micron are telling investors they are at peak; the behind-the-meter and services names are priced on out-year numbers
Three years is the limit of visibility in technology, and the market is paying for one or two quarters
Five years ago the subject was NFTs, SPACs and the metaverse, and AI was not on the radar
His own frame is a two-to-three-year through-cycle earnings multiple with downside protection
The deal-announcement feedback loop has broken, and he reads that as bullish
Partnership and financing announcements no longer move stocks the way they did six months ago
"You'd rather be in an environment where you have to sit and earn the returns than one where you get stocks moving 20-30% on press releases."
Heavy retail involvement means the clock is ticking on a trade — but shorting an open-ended growth story is deadly
Boring hardware names are now retail favorites, the way software names were in 2020 and 2021
There is no catalyst that ends an open-ended story, and he is not shorting AI here
This is not 2021: the multiples then were less defensible and the damage came from rates
Software carried "50 times revenue" into a pandemic pull-forward in cloud adoption
He does not expect this trade to break until adoption slows
The buildout is getting funded, and the people running it treat a higher cost of debt as a rounding error
On his count the labs, SpaceX and Google will have raised about $500 billion year to date
"There's been maybe a disconnect between the financial analysis of this and the religious analysis that maybe comes from the West Coast."
Hyperscaler capex is an option purchase, not a return calculation, which is why the stock market cannot stop it
Defensive, because not playing could wipe out the business; offensive, because the upside is open-ended
A four or five percent return would not stop the spending — "It's kind of a waste of capital, but it's not incinerating capital."
The megacap trade is over as the easy trade because cloud went from three vendors to nine or 10
Fewer customers, more scaled providers of compute, and market caps large enough that the law of large numbers bites
Amazon compounded roughly tenfold from 2016 or 2017; he thinks 15-20% a year is the realistic profile from here
Meta's offer to release excess compute is a signal to investors, not a plan for the compute
Sub-leasing a slice would earn a couple of billion dollars and not move a company that size
Sebastian Mallaby's book on Demis Hassabis is his evidence for Zuckerberg's habit of going big on every trend
He runs long biased above 50% net, and thinks the old-school long-short model is due a renaissance
Capital wants either cheap long-only or high-fee market neutral, which leaves the middle underserved
Monthly results, and probably quarterly results, are a noise generator he has to manage LPs through
Three things would tell him the AI trade has rolled over: lab ARR, capex, and the price of compute
All three look strong, he has never gone net short, and he does not expect to
"There's going to be a lot of pain, I think, if this ever rolls over. And I just don't think it's yet."
AI is making markets less efficient rather than more, by speeding up and crowding narratives
Sell-side research, Substacks and investor chatter are increasingly model-written, and consensus thoughts are not worth much
He disagrees flatly with the argument that rational machines drain fear and greed out of prices
Software Ripped, Semiconductors Fell, and Keller Calls It Normal Volatility
Wiethe opened with his own arithmetic on the quarter. "I actually ran the numbers as of yesterday's close. So since the start of Q3 software and services are up 20%.", he said, adding that "Technology hardware and equipment are basically flat but if you go down to semiconductors and semiconductor equipment we're down 7 and a half% to start the quarter." He called it a massive divergence from the way 2026 began and asked what Keller was seeing.
Keller's answer started with his own background, which is operational rather than analytical — time at Salesforce and at MongoDB — and which he said orients him toward big themes, a longer horizon and extreme dislocations.
He thinks the year opened on one of those dislocations. The war in Iran was distracting everyone from what he called massively inflecting adoption of AI, a long-running story that had its takeoff moment in the first quarter
The second quarter is where the crowd arrived. He would not call it the blowoff top, but there was a 100% gain in SOX and retail got heavily involved, so a choppy trend afterward is not surprising
He also pointed to "situational awareness and deleveraging in Korea" as part of what made the volatility obvious in hindsight
The bull case he repeated is that the base businesses are still compounding. Apple services growing double digits, Google search growing double digits, Meta growing in the 20s — which on his reading leaves room for 10 to 15 years of double-digit growth for the AI labs, lifting all boats
"I think we're still moving upwards on an S-curve with regards to AI adoption."
Software, he said, was thrown out with the bathwater, so some of the bounce makes sense. "I don't think that late June was the end of this mega trend. I think this is just kind of normal volatility."
The Certainty Premium: Cyber and Infrastructure Are Being Paid 25-30 Times Revenue for Visibility
Wiethe asked whether the software rally is the market deciding software is less at risk of AI disruption than it feared, or something more technical — software shorts were paired against AI longs, and degrossing is still running through the hedge fund complex.
Keller said the correlation itself gives the game away. "It's been entertaining to see SOX and IGV basically have a negative one correlation." Bad news for AI, he said, sends ServiceNow and Salesforce up several percent, which is an odd dynamic and clearly technical in part. Then he split the sector in two.
Application software is repricing, not being disrupted. A sector that traded at a premium on a decade of double-digit growth is being marked down to low-double or high-single-digit growth. "And once you get into that five to 15% growth area, it's really hard to trade on revenue multiples and non-GAAP EPS and it's really hard to dream the dream."
On Salesforce and Workday — and the potential acquisition of Workday — he said the initial crash was the sector "coming back down to earth" rather than the market pricing AI disruption of the whole business. "It was just these are low to moderate growers with a little bit left tail risk."
The bounce that matters is in cyber and infrastructure, and it is a premium on certainty. "I think in this market, long-term uncertainty is so high that anything with short-term certainty is trading at a huge premium."
Cyber numbers are known to be good because buyers are rushing to build defenses against new AI threats; infrastructure is known to accelerate because more software adoption flows into existing tools
He is not bearish on them, but he thinks the price is wrong. "You're paying sort of 25-30 times revenue for large companies. I think that the success rate on those is paying up there is low."
"cyber is also not immune from disruption threats" — he said people at the AI labs think cyber is very much something they will disrupt as well
"I think that's, it's a lofty territory that they're in more so than AI, frankly. I'd argue that the real AI winners are CrowdStrike and some like that rather than maybe some of the semis."
The Database Buyer Is Now a Model, Not a Person
Asked for concrete examples of infrastructure, Keller went to the company he used to work for, flagging the bias himself.
AI proliferates software, and software proliferates databases. People are producing vibe-coded apps or simply developing more efficiently inside large organizations, and "And you know, every software application has a database line underneath it."
"It's no longer human. It's now various large language models." — the selection of the database has moved to the model, which he said leaves the mix shift unclear even if a rising tide lifts the boat
Anything that meters compute has seen the same pickup, including CDNs, and he said usage-based businesses can spike quickly in the short term
The short-term asymmetry is what he is pointing at. Infrastructure has a higher chance of beating than Salesforce or Wix, which are less likely to blow up numbers in a quarter; the long-term picture he called more open
In Technology, the Theme Pays More Than the Stock Picking
Wiethe asked whether tech is a relative game of picking winners inside a sector or a question of where the pie is growing.
Keller said the work is bottoms-up and company-driven, and then said the money is somewhere else. "But I think in technology a lot more money gets made being in the correct major theme than it does picking sort of within those themes." Knowing ten years ago that cloud software was taking share mattered more than which cloud name you owned.
AI is still in the thematic phase but moving out of it. Dispersion arrives as the open-ended growth story slows, and he said it is already starting to matter who has sustainability rather than who is being lifted by the tide
He put the reason for that in scale terms later in the conversation, when he described the biggest platforms as running at GDP-level market caps and profits
Where He Would Own the AI Trade: the Low Multiples, Not the Hidden Winners
Wiethe noted that technology finished third from the bottom of the S&P's 11 sectors coming out of the bounce — "I mean I also ran the numbers on just the S&P 11 sectors and you know technology came in third. The only things that were worse were industrials and utilities which were really thrown into the AI trade." — and asked where investors who still want AI exposure should look.
Keller started with the risk rather than the pick.
The first risk is extrapolating a very short-term dynamic, and he named the price of compute and the neoclouds as where that is happening
Elon Musk entering the space, he said, "I do think that I think there's a little bit of a risk that this very short-term extremely high price of compute is Elon might move us quicker to the glut."
Everything in the complex runs off the same two drivers at wildly different multiples. "You know, the other thing is it's funny that of all these various, whether they're bottlenecks or semiconductor companies, they trade on the same factor and they trade on basically hyperscaler capex and the ARR of the labs." He credited the point to Gavin Baker
The names telling you they are at peak are the cheap ones. Behind-the-meter and services companies trade at high multiples on out-year numbers, while Nvidia, Micron and the well-known names are not extrapolating. "In fact, they're telling you we're at peak."
The contrarian conclusion: "I think it's likely that the low multiple names across the AI trade in general are going to do better, even if they're at different cycles of earnings growth."
The "hidden AI winners" that investors have piled into are, on his reading, pricing a rosier future than a lot of semis are
Three Years Is the Limit of Visibility, and the Market Is Paying for One Quarter
Asked how he values anything in this sector, Keller began with a joke at his own expense: people, himself included, talk about four- and five-year projections, and five years ago the conversation was NFTs, SPACs and the metaverse, with AI not on the radar.
The valuation risk is cyclicality inside a group that has never been treated as cyclical. These are highly cyclical businesses trading on what he called massively inflated rather than peak earnings, held by a new profile of investor
"Just relying on multiple can be very dangerous because there's high cyclicality here that a lot of tech investors are not used to."
His horizon is set by where forecasting stops working. "But I think, if you have good visibility into three years, beyond that is just hard to forecast in technology."
"I think there's a very high premium on one to two quarters." That is a different game from the one he plays, which he described as a two-to-three-year through-cycle earnings multiple with decent downside protection
How he bridges the gap: the long-short model lets him bet on both sides of the book and try to isolate the factors, and the shorter the market's horizon gets, the further out he is willing to look
"I've never heard sort of more conversations about what's the catalyst path and who's going to beat sellside consensus numbers."
"So, shorting is a dangerous game." — the discipline is knowing how far people can extrapolate and how juicy a near-term imbalance can get, then sitting and waiting
The Deal-Announcement Feedback Loop Broke, and He Reads It as Bullish
Wiethe described the mechanism that made the first half of the year feel different: a semiconductor company would sign with Nvidia and the stock would rip 10-20%, or a company would announce a bond offering to fund capex and the presumed recipients of the money would rally. The announcements are still coming, he said, and the reaction has gone.
Keller called it sobriety, and he prefers it.
The comparison he reached for was 1990s, then 2021. He was not investing in the 1990s, but he said "You know, the euphoria was multiples higher in 2021 than it is now, which is interesting because I think we have a much more bankable long-term technology trend."
The muted reaction is partly about unreadable economics. With data centers, one company leases from another and "We don't know who's on the hook for extra costs. We don't know what happens if there's a delay."
The association trade has stopped working. A year ago any association with OpenAI was gold; in the first half of this year any association with Anthropic was; now there is more scrutiny about sustainable long-term profits
He pointed to the crypto miners converting to data centers, where the euphoria has come off entirely
Wiethe noted that Bitcoin is ripping and the miners are ripping with it, so they get the best of both worlds. Keller: "They might have to shut down the data centers and go back to Bitcoin mining."
The bullish reading: "You'd rather be in an environment where you have to sit and earn the returns than one where you get stocks moving 20-30% on press releases." A longer, slower cycle is easier to stay invested in
Heavy Retail Is the Tell, and Shorting an Open-Ended Story Is Deadly
Wiethe raised the holder base: type a hot ticker into X after one of these announcements and you find retail cheering the deal and asking why the stock is not popping the way it did three to six months ago.
Keller's rule of thumb is blunt. "Whenever you get heavy retail involvement in any trade, that usually in my experience has meant that the clock is ticking for the end of that trade."
He listed the precedents: gold and silver at the start of this year, crypto, and software in 2020 and 2021, when people were launching paid newsletters about MongoDB, Elastic and — Wiethe supplied the name — Fastly
"I think for sure that you have these you know boring hardware companies for 10 years that are now all the rage among retail investors. I think that's certainly dangerous."
The other side of the trade is why he will not act on it. "You know, one thing I've learned is shorting an open-ended growth story can be deadly because there's no catalyst." A negative data point helps, but nothing puts a nail in the coffin
He thinks the current stretch is an information vacuum. People are latching on to every ARR leak from Anthropic or OpenAI, and he expects that to clear when the labs are public
"I like short selling. I'm not piling into shorting AI here." Cleaning out the leverage, he said, is healthy, even if he does not expect a rip to all-time highs
Why This Is Not 2021, and Why the Damage That Year Came From Rates
Wiethe put the parallel carefully. The 2021 deleveraging around Archegos and the Cathie Wood ARK names marked the top for those names even as the broader market powered on until inflation took everything but energy down in 2022. He said comparing those businesses to today's AI companies would be disingenuous, then asked whether there is more pain to come for what he called the tourist semiconductor investor.
Keller quibbled with parts of it.
The real businesses held up longer than the speculation. Tesla and software peaked in late 2021, well after the SPACs and the GameStops cratered around the first quarter
Inflation and higher rates did the damage in 2022, and the starting multiples were worse: "50 times revenue on software" after a pandemic pull-forward in cloud adoption
Today's multiples he calls more reasonable, with more legs under the trend than that pull-forward had
"I don't think that this trade will break until you start to see adoption slowing." Right now, he said, adoption is explosive, and that is the difference from 2021
Funding the Buildout: About $500 Billion Raised, and a Religious Analysis From the West Coast
Wiethe laid out the financing problem — the Fed possibly hiking by the end of 2026, fiscal concerns, and an AI buildout funded through credit markets to the point that it is crowding out the Treasury market — and asked whether long-end yields threaten the companies' ability to keep raising.
Keller declined to forecast rates, saying far greater minds have spent a lot of time trying to figure out which way the bond market goes, and that it is entirely possible the 10-year is 30 or 40 basis points lower in two months. Then he answered the question underneath it.
He accepts the framing and is not worried. What used to be buybacks — hundreds of billions, soon near trillions — is going away and requiring debt issuance instead, which he called a more dangerous way to finance a buildout
His evidence that the market is absorbing it: "You know, especially after the Anthropic IPO, these major the labs and SpaceX and Google will have raised about $500 billion year to date.", alongside a lot of debt, and markets are broadly okay
Top lines are still growing double digits, in the 20s for some, and he said there is probably more cost base to take out if it came to that, as these companies did in 2022
The people spending the money are not making a financial decision. "This is essentially the greatest race of their career."
"You see Sergey and Larry coming back to Google to work on this. You see Elon basically pivoting all of his attention to building data centers."
"And if they need to borrow at 150 bips higher than they thought they did, and it ends up in aggregate being whatever that math is, big dollars, 10-20 billion dollars more a year, I don't think that's going to slow them down."
"I think there's been maybe a disconnect between the financial analysis of this and the religious analysis that maybe comes from the West Coast." He grew up on the West Coast, he said, so he likes to think he can do a little of both
The limit he does put on it: credit markets might try to slow the spending, and unless yields really blow out he thinks there is enough money to fund it for a couple of years. If the return never materializes, "we'll have a problem"
The Megacap Trade Is Over as the Easy Trade Because Cloud Went From Three Vendors to 10
Wiethe pointed to Google's first negative free cash flow quarter and asked whether the rest of the hyperscalers are heading the same way, and whether buying big megacap tech has stopped being the easy route to outperformance.
Keller agreed, while saying he owns some of them and thinks the stocks are reasonably cheap.
The historical return profile is the part that is gone. "You know, Amazon, you can check me on this, but is probably up something like 10x since 2016 or 2017." Making 15 to 20% a year from here is possible; making that again is not the base case
The cloud business that drove the group has changed shape. "You know that market is dramatically changed from basically three vendors with you know thousands of customers to now there's probably nine or 10 scaled providers of compute."
He counted Oracle, SpaceX, the labs procuring directly, the neoclouds, CoreWeave and Nebius among them, and noted that Apple and Meta were never part of the cloud engine anyway
Scale advantages remain, but there are fewer customers, more competition and more uncertainty — which is why he understands the multiples
The law of large numbers is doing the rest. At GDP-level market caps and profits, he said, two, five or ten times returns are behind the group: "I think that was a golden period over the last 5 to seven years but I think you're right that that's over but doesn't mean they can't still be good risk-rewards."
Capex Is an Option Purchase, Which Is Why the Stock Market Cannot Stop It
Wiethe pulled Google's adjusted free cash flow estimates up on his screen mid-conversation and asked whether the consensus path presupposes too much certainty about the return on this capex — whether the firms are really betting on five or ten years out, over exactly the horizon Keller had just called unforecastable.
Keller reframed the question away from return on invested capital.
"it's almost like buying an option" — and he says it is two options at once
Defensive: it is existential, and not playing the game could wipe out the business
Offensive: the cloud numbers show new products and open-ended upside if you end up the leading provider of compute
The structure of the spend is itself optionality. He cited Microsoft's latest call describing powered shells bought without chips in them: "You're essentially buying a couple years of optionality where if you need to, you can ramp up a lot of capacity really quickly."
His analogy inverts the usual one. People think of it as concrete — put $1,000 in the ground, what comes out? "You got to put it in because it might be this open-ended upside." and because the alternative is too risky
A low but positive return still clears the bar for the company, if not for the shareholder. "Number two, it might not be great for shareholders because if the ROIC is only like four or 5%, that's not great."
"It's kind of a waste of capital, but it's not incinerating capital."
The conclusion is that the spending continues regardless. The risk of stopping, he said, is much greater than the optionality bought by staying in the game
Wiethe pressed the point: if 2027 arrives without visibility into 2028, does that uncertainty keep punishing the hyperscalers? Keller said there is a decent amount of queasiness, and gave the cleanest illustration of how far expectations have moved.
"I remember two years ago everyone was praying that AWS would grow like 18%." It is now growing in the 40s, he said, and the stock is only incrementally higher
The second worry is customer concentration, with Meta, on his phrasing, "kind of on the ropes with it"
He thinks executives are less stock-price sensitive than investors assume. It took a long time before Meta started making cuts in 2022, and a 10-20% drawdown would read as a bear market without changing the plan
"I struggle to see the stock market forcing them to stop capex because in my view the ROIC is good enough that the stocks will never get so cheap that they change course." He credited Andy Jassy with laying the hyperscaler return case out well, and said people are on board with it for now
Meta's Offer to Release Compute Is a Signal to Investors, Not a Plan for the Compute
Wiethe opened the Meta bear case: the company was punished for saying it was building compute purely for itself, then rewarded when it indicated it would sell out any excess. If Mark Zuckerberg cannot "figure out what to do with the compute", Wiethe said, that seemed to him "a bear case for AI technology in general".
Keller thinks renting the capacity out would be shortsighted, and that renting it out was never the point. A three-month contract could earn a couple of billion dollars in profits, helpful but not needle-moving for a company that size
"I think it's more almost telling investors, look, we get it and we will stop if need be. We will release our capacity if need be."
The Meta problems he does see are internal, not evidence against AI: a new team hired, an original open-source focus abandoned, a lot of bouncing around
"I think as long as the aggregate is making progress, that's more relevant to the broader AI trade."
His reservation about Zuckerberg comes out of a book. He cited a passage in Sebastian Mallaby's book on Demis Hassabis in which Zuckerberg is asked about AI alongside other technology trends and is equally enthusiastic about AI, crypto and NFTs
"Zuckerberg far richer than I am, far more successful, but he has a tendency to go big on tech trends that it's not really obvious that he needs to do that for the core business."
As a shareholder that makes it a binary question, and one he says he cannot underwrite — while insisting Meta's troubles are not evidence of broader AI troubles
Wiethe's own objection was to the people holding both views. There is no plan at the corporate level in most companies, he said, and it is genuinely hard to walk into senior management proposing a technology whose bull case is that everyone gets fired
Keller agreed and put it back in option terms. Meta is the biggest spender, gets the most questions, and needs its investors to give it rope: "I need to keep playing the game and I need my investors to give me rope."
Positioning: Contrarian on the Data Center Backlash, Long the Left-for-Dead Growth Stories
Asked what the book looks like now, Keller described a contrarian streak pointed at the crowd's political worry.
He thinks the data center backlash is overstated as an investment risk. Stocks are implying that data center starts will be challenged — by labor, or by the sites applying for power — and he calls that unlikely and short term
Past the midterms, and whether or not particular states impose moratoriums, "I think they're going to get built."
"We've seen various technology moral panics over time and usually the market is strong enough."
So the long is the deceleration that he does not think is coming. He is looking at names priced for a slowdown in data center starts and sees legs there
The other bucket is idiosyncratic growth left for dead — e-commerce names, recent IPOs, businesses that stand on their own legs and used to be bid up to high multiples, and that nobody is talking about while the whole conversation is about AI and software
He is wary of the reverse everywhere. Anywhere near-term certainty is high, he looks contrarian, because those mini-cycles come down the back slope
Long Biased, and a Bet That Old-School Long-Short Is Due a Renaissance
"I refer to the fund as long biased which I count as having sort of more than 50% net exposure", and he said it can run higher than that
The argument against market neutral is that the beta is free and worth having. "I like being long biased because markets go up over time." Capturing it for investors is not something to get paid for, but it is something to capture
The crowding he sees is on the other side. A great deal of money is market neutral, relative-value focused, factor balanced, or forced to delever at the wrong time, and being unconstrained in that way is the opportunity
"I think that the kind of old school long-short model is probably in for a renaissance."
The reason it is hard to run is commercial rather than intellectual: "You know, people want either a cheap long only or a high fee market neutral shop."
What Would Have to Break: Lab ARR, Capex, and the Price of Compute
Wiethe asked whether Keller ever goes fully net short, and what would have to happen to get him that bearish — noting that for every Cliff Sosin who made it out of 2022 and back, several managers did not.
"It's unlikely I would ever get net short." "I think late 2021 was a once in 10 or 15 year environment", with government support, retail involvement and clear cyclicality in what had been secular industries all at once
He concedes the AI trade will eventually have those ingredients: a huge change in underlying business activity, heavy investor enthusiasm — people talking about curing cancer and money becoming obsolete — and cyclicality inside an investor base used to secular growth
The three things he actually watches:
Lab ARR and the adoption S-curve. "I don't think there's any crack in there yet."
Capex. It does not need to keep growing at these rates; leveling off at a high level would still be fine for the stocks
The price of compute. That is where an overbuild would show up first, and neither pricing nor build times are signaling a glut, even allowing for Musk
"I think the ARR is strong. I think the capex is strong." — and if and when it does crack, he said, a short-selling skill will be worth having, because a lot of highly cyclical businesses are trading peak on peak, "and even probably some stocks that I own"
"There's going to be a lot of pain, I think, if this ever rolls over. And I just don't think it's yet."
AI Makes Markets Less Efficient, Not More
Wiethe's own worry is the echo chamber. Investors are doing genuine fundamental work on where companies sit in the supply chain, building theses that feel bulletproof, and then sense-checking them against a chatbot that tells them the thesis is smart — a private version of the 1990s chat rooms, and one the market has no obligation to agree with.
Keller took the point further.
"I mean, the amount of AI generated content that I've seen, whether it's in sellside research or substacks or even just investor chatter is surprising to me because the whole job is you got to come up with your own thoughts." Standard consensus thoughts, he said, are easy to get and not worth much
He rates the tools highly for the first pass. AI is great for ramping up and great for facts and figures; the echo chamber is the cost
The bigger opportunity he sees is behavioral. Phone addiction, the speed at which narratives spin up in politics or stocks, and shrinking attention spans are all working the same direction
"I think you're going to get faster narrative cycles. I think you're going to get more crowded narrative cycles."
He named the argument he disagrees with, and disagreed with it flatly. He had heard someone on another podcast say advanced AI would strip fear and greed out of markets because machines are perfectly rational
"So I think that there's a an emerging behavioral inefficiency. I think markets probably getting less efficient over time and I don't think AI really helps with that."
He allowed one exception for quantitative models, then misspoke on the last clause. Wiethe stopped him to clarify which way round he meant it, and Keller took the correction: "Yeah, that yes, thanks for catching that."
Factor Awareness Is an LP-Management Problem, and Extremes Are the Opportunity
Wiethe's framing was that the speed at which a stock stops being idiosyncratic alpha and becomes a bucket, a theme or a trend has gone up sharply, and asked how a bottom-up investor sense-checks himself against that.
Keller says his own attention to it is recent. He has spent much more time on factor exposures in the last year or two and was not paying attention four years ago, and he is not sure whether that is his awareness changing or the market
He treats it as a time-horizon and an LP problem rather than a portfolio-construction one. "Quarters are probably insignificant and so it's kind of just a you know a noise generator" given how much factors move
His worked example: a software investor fielding questions from LPs about what was wrong with the process, when the answer was that AI funds were levering up and shorting software, driving negative momentum into the stock — a two-month effect in names that round-tripped over four months
The rule he draws is conscious factor bets, not neutral ones. Be aware of them, be deliberate about them, and manage the people who are watching monthly
The payoff for that awareness is at the extremes. "One of the things I like to go to is extremes."
The signals he names are the ones that sound like capitulation: momentum having one of its worst five-day stretches in 20 years, a sector called uninvestable, or the largest hedge fund shorting day in seven years
The July Playbook: Buy the Country That Is Getting Margin-Called
Wiethe said nobody listening got through July without an AI position falling every day on no company news, and asked what a professional actually does in that stretch.
Keller described a sequence.
"I mean, there's kind of a life cycle of at first you think, oh, this is great. My stock's down. I can buy more." Then a little worry. Then the recognition that bigger factors are at play
What resolved it was locating the seller. The pain was concentrated in Asia, especially Korea: "You saw some notable percentage of the Korean population getting margin calls. Who knows if that's kind of rumor, but you saw a lot of margin calls, a lot of leverage in Korea, huge deleveraging."
"You have huge hedge fund deleveraging, huge retail deleveraging in a secular trend that I actually think is still going pretty well for the basically two large companies in that country."
The conclusion, once the pieces are in place: "Look, I want to be a buyer when people are getting margin called and liquidated."
Absent that clarity he does nothing. Separating fundamentals from technicals is hard, and it takes a lot of signals lining up before he treats a drawdown as an opportunity rather than information
Optionality Beats Being Fully Invested, and Leverage Is What Destroys It
Asked whether he holds cash for those moments, Keller broadened it.
"I think optionality is generally one of the most important things when managing a portfolio" — cash, or lower gross exposure, or anything that keeps him from being the one deleveraging when the opportunity shows up
The frequency of the events is the argument. Three- and four-standard-deviation moves, or what he called once-in-a-20-year events, now seem to arrive every several months
"I'm trying to wait until something hits me in the face." There is a graveyard of managers trying to trade short-term moves, and he wants moves large enough to be unmistakable
The reason it matters more than being fully invested: "But you make your money on one or two good ideas a year, a couple big ideas every couple years."
Leverage is the direct enemy of that. He said he runs with very low leverage, comparatively, and had not realized how much others were running — "leverage is a great way to reduce optionality"
Wiethe brought up realized volatility in Korean equities and in semiconductors against a calm headline market, and Keller said that gap is the defining feature of this year.
"The headline indices are very tranquil and you're actually seeing these incredible vol spikes", with the dispersion sitting inside sectors and inside individual stocks
He compared the market to a duck: placid on the surface, kicking hard underneath, with a year of rotations and chances to be nimble at the extremes
"So, my guess is that that will increase and sort of the behavioral advantages will increase." Shorter horizons, more uncertainty and more money in less fundamental strategies all push the same way
Into Year End: the Anthropic S1, and Whether a Listing Sucks Capital Out of the Trade
Wiethe closed by noting that with the exception of Nvidia, which had not yet reported when they recorded, earnings season is largely done, and asked what Keller is watching into the end of 2026.
"I think we'll get the Anthropic S1 probably, it sounds like next week or the week after." He does not expect much new information — his guess is the numbers are broadly leaked — but he expects precision, and metrics instead of third-hand guessing
The same three-item list runs into year end: lab ARR and open-source adoption, capex — which he called locked to be strong for 2027 with attention moving to 2028 — and the forward price of compute, which he called very tight and likely to stay tight
"I think if you're going to see an overbuild in data centers, you'll see it there."
On the stocks themselves he is neutral to cautious. "It would not surprise me if we chop around for a while." After the retail and investor enthusiasm of June, he said, things usually do not rip straight back to new highs
The exception would be a technical breakthrough: "I don't know what it is, but there's recursive self-improvement is kind of floating out there."
Wiethe's last question was whether the Anthropic listing creates a capital vacuum, the way investors sold the space also-rans the moment they could buy SpaceX directly — and he noted there is no pure-play lab exposure in public markets today.
Keller thinks the analogy is imperfect because the businesses are real. "I think that equity issuance is a huge signal for kind of euphoria and blowoff tops.", but issuance into real businesses at realistic prices is a different thing
"So, I think that equity issuance into realistic businesses at realistic prices is less capital sucking than you know, we need to put $10 billion into Nikola or whatever we were doing back in 2021."
He would quibble with the SpaceX valuation, and noted the space stocks doubled and were then cut in half, leaving them back where they started
He expects appetite rather than a vacuum. The euphoria and positioning have cleaned out of the semiconductor trade, and "I think the market's ready to digest Anthropic, and I think it'll want lab exposure. Frankly, I think it's going to trade in at a crazy price."
He was careful to add that he is not saying he would buy it
Signing off, Keller pointed listeners to his account on X — "I'm on Twitter, Jeff Keller1." — and to an email address on a domain he said he had picked because he never paid for the .com. Wiethe's reply: "Nobody's paying for the com anymore."
Keller's bottom line is that the AI trade is still an open-ended growth story with adoption, capex and compute pricing all pointing the same way, so the thing to fade is not the trend but the price the market is currently paying for short-term certainty.
Products, Companies & Tools Mentioned
Capelight Partners (Keller's technology sector focused hedge fund, running since 2021; he describes it as long biased, above 50% net exposure, with very low leverage)
Salesforce and MongoDB (Where his operating background comes from, and both examples in the conversation — Salesforce as a low-to-moderate grower unlikely to blow up short-term numbers, MongoDB as the database layer under an explosion of AI-generated software)
CrowdStrike (His candidate for the real AI winners, "rather than maybe some of the semis" — while warning that cyber multiples are loftier than AI's)
ServiceNow (Named with Salesforce as the stocks that rise several percent on bad news for AI, evidence of the negative correlation between semis and software)
Workday (The potential acquisition he cites as part of application software's repricing, which he says was never about AI disrupting the whole business)
Fastly and Elastic (The usage-based and infrastructure names that spike on more compute, and — with MongoDB — the subjects of the paid newsletters he remembers from the 2020-21 retail wave)
Wix (His other example of software less likely to blow up short-term numbers than infrastructure)
Nvidia and Micron (The well-known semiconductor names he says are not extrapolating: "In fact, they're telling you we're at peak.")
Google (The first negative free cash flow quarter among the hyperscalers, and the company whose consensus estimates Wiethe pulled up mid-conversation; also where Sergey Brin and Larry Page have come back to work on AI)
Meta (The compute overbuild bear case: punished for building for itself, rewarded for offering to release capacity, which Keller reads as signaling to investors rather than a plan)
Amazon and AWS (His measure of how the megacap profile has changed — "up something like 10x since 2016 or 2017" — and of moved expectations, from praying for 18% growth two years ago to the 40s now)
Apple (Named in the bull case for base-business durability: services still growing double digits)
Microsoft (Cited for the powered-shells framing on its latest call — capacity bought without chips in it, as a couple of years of optionality)
Oracle, CoreWeave and Nebius (Among the nine or 10 scaled providers of compute that replaced a three-vendor cloud market)
Anthropic and OpenAI (The ARR leaks the market is trading off, the S1 he expects within weeks, and the labs now building their own power and data centers)
SpaceX (Both a source of the capital being raised this year and the precedent for a listing that drains money out of the also-rans; he would "certainly quibble with the SpaceX valuation")
Tesla and GameStop (His 2021 timeline: the meme names cratered around the first quarter while the real businesses peaked in late 2021)
ARK Invest (Wiethe's reference point for the 2021 deleveraging, alongside Archegos, and the case he says is not comparable to today's AI businesses)
SOX, IGV and Nikola (The semiconductor and software benchmarks whose correlation he called negative one, and the 2021 capital sink he contrasts with today's equity issuance)
Books & Resources Mentioned
The Infinity Machine – Sebastian Mallaby (The book on Demis Hassabis that Keller cites for a passage on Zuckerberg being equally enthusiastic about AI, crypto and NFTs)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.