Jeff Keller — founder and portfolio manager of Capeite Partners, a technology-sector-focused long-biased hedge fund (est. 2021). Ex-operator (Salesforce, MongoDB); themes-first tech investor known for navigating the 2022 drawdown.
Korea took the July deleveraging — hedge-fund degrossing plus a notable share of retail margin-called — inside a secular trend still working for "basically two large companies in that country." He wants to be the buyer of a forced liquidation.
His infra example (and he flags the ex-employee bias): AI proliferates software, and "every software application has a database line underneath it" — an exploding end market should explode database revenue, though models now pick the database.
Named with Nvidia as the well-known AI name whose numbers already embed peak rather than extrapolation — the cheap way to own hyperscaler capex + lab ARR.
The AI complex all trades on the same two drivers (hyperscaler capex + lab ARR) at wildly different multiples; Nvidia "is really not extrapolating" — it's telling you we're at peak, which is why he prefers the low-multiple end of the trade.
The bull's durability data point — "Apple services is still growing double digits today" — behind the case for 10–15 years of double-digit growth lifting all boats; excluded (with Meta) from the hyperscaler-cloud engine that drove Mag-7 returns.
Up ~10x since 2016–17 — that return profile is behind us. AWS reaccelerating from a hoped-for 18% to the 40s while the stock barely moved is the market's queasiness about ROI and customer concentration. Still 15–20%/yr, just a different profile.
The S-1 lands within a week or two; little new information, but public labs replace third-hand ARR leaks with metrics and damp narrative volatility. On the listing: "I think it's going to trade at a crazy price. I'm not saying I'll buy it, but there's going to be a lot of enthusiasm for it."
A former employer, and the standard-bearer for application software. The crash was a rational multiple recalibration to 5–15% growth — "it's really hard to dream the dream" — not the market pricing AI destroying the business.
"The real AI winners are CrowdStrike and some like that rather than maybe some of the semis" — but at 25–30× revenue, with the labs themselves eyeing cyber for disruption, it's "lofty territory… more so than AI, frankly."
His CDN example of usage-based uplift: more compute flows straight into the numbers, so short-term spikes are quick but the long term is "more under question." Also a 2020–21 retail-mania marker (paid newsletters were written about it).
Search still growing double digits and the stock reasonably cheap, but the first negative free-cash-flow quarter is real and consensus's 2028 snap-back presupposes ROIC visibility nobody has. Sergey and Larry returning is his proof the spend is existential, not financial.
"Socks [SOX] and IGV basically have a negative one correlation" — his cleanest evidence that the software rally is substantially the unwind of long-AI / short-software pairs, not a fundamental verdict.
Biggest relative spender; the sublease messaging is signalling to buy investor rope, not a profit lever. Rational spend, but Zuckerberg's pattern of going big on every tech trend makes it "a more binary question as a shareholder. It's not one I can underwrite."
The worked example of capex as an option, not concrete: buying "powered shells" before the chips is literally purchasing a couple of years of ramp optionality — rational even at a low-but-positive ROI.
Exhibit for the technical read on the software rally: "some bad news for AI leads into ServiceNow and Salesforce stocks go up several percent" — the unwind of paired long-AI / short-software books.
With Anthropic, the ARR the whole complex trades off — "people are latching on to every ARR leak" in an information vacuum. Both are also building their own power and data centers, blurring where the profit pool lands.
In the roll-call of scaled compute providers that took the cloud market from three vendors with thousands of customers to nine or ten providers with fewer customers — structurally why hyperscaler multiples don't re-rate on growth.
Q2's 100% gain in the SOX with heavy retail involvement made the subsequent chop unsurprising — but late June was not the end of the mega trend. Realized vol ~55 on a one-month look-back (peak ~185).
The capital-vacuum precedent (space also-rans sold to fund it), which he thinks won't repeat for AI: issuance into real businesses at realistic prices "is less capital sucking" than 2021's SPAC era. He'd still "quibble with the SpaceX valuation."
2021-analogy reference: the SPACs and GameStops cratered in Q1 2021 while "stocks like Tesla, software eventually kind of hit their peak in late 21" — it was inflation and rates that did the 2022 damage.
Grouped with Salesforce in the application-software recalibration — and the potential acquisition is what happens to a low-growth software franchise once the premium multiple resets.
Elon "is entering that space" and pivoting all his attention to building data centers — new supply that could move the market to a compute glut faster and shorten the window on today's pricing.
Where he "starts with a negative": extrapolating today's extremely high price of compute is the biggest risk in the AI trade, and Elon entering the space "might move us quicker to the glut" — into a market that went from three compute providers to nine or ten.
The other named neocloud under the same caution — value rests on today's very tight compute pricing, exactly the short-term dynamic he'd be most careful extrapolating as new supply keeps arriving.
A themes-first, long-biased tech investor with an operator's background (Salesforce, MongoDB). The core claim: AI is a genuine open-ended growth story still climbing the adoption S-curve, so drawdowns like July 2026 are deleveraging events, not tops — and the way to be paid is to be in the big theme early, own the cheapest expression of the factor everything trades on, and keep enough optionality to be the buyer when leveraged holders are liquidated.
Be in the big theme early; pick names later. "In technology a lot more money gets made being in the correct major theme than it does picking within those themes." Sector exposure wins in the open-ended phase; dispersion and sustainability decide returns only once the numbers reach GDP scale — where AI is starting to arrive now.
Short-term certainty trades at a huge premium. When long-term uncertainty is this high, anything with knowable next-two-quarters (cyber, usage-based infra) gets bid to 25–30× revenue. He underwrites a 2–3 year through-cycle earnings multiple with downside protection instead, and leaves the one-to-two-quarter beat game alone.
One factor, many multiples. The whole AI complex trades on hyperscaler capex and lab ARR. Names telling you "we're at peak" (Nvidia, Micron) should beat the ones priced for an extrapolated future — same exposure, lower bar.
Heavy retail involvement means the clock is ticking on a trade — not that today is the top. Paid newsletters about single names, ticker cheerleading, and big announcements that stop moving stocks are the tells.
Never short an open-ended growth story — there's no catalyst and no end to the narrative. He likes short selling and still won't short AI; the trade only breaks when adoption slows.
Three gauges on the AI trade: (1) lab ARR / the adoption S-curve, (2) the capex trajectory (levelling off at a high level is fine), (3) the forward price of compute and data-center build times — where an overbuild shows first. As of AUG-2026 all three read strong.
Hyperscaler capex is optionality, not ROIC. Defensive (not playing could kill the business) plus offensive (open-ended upside), structured to stay optional — powered shells before chips. So a low-but-positive return is rational, the spending continues regardless of the share price, and it needs "religious analysis from the West Coast" more than financial analysis.
Buy the forced liquidation, but only with all the pieces. Identify the levered seller, the geography where the pain concentrates and evidence of margin calls; require the secular story to be intact; demand a 3–4 standard-deviation extreme — then be the buyer.
Optionality > being fully deployed. Cash, low gross, low leverage — because "three standard deviation events seem to be happening every several months," and you make your money on one or two ideas a year. Leverage is "a great way to reduce optionality."
Behavioral inefficiency is rising in the AI era. He expects faster and more crowded narrative cycles, LLM echo chambers manufacturing consensus and overconfidence — "markets probably getting less efficient over time," which is the case for an unconstrained, long-duration, old-school long/short book.
Appearances
One dated page per appearance — each has its full stock table, talking points, and the saved transcript. Newest first.