Joseph Carlson — research hub

Stock & name index

▲ Positive

► Neutral / referenced

▼ Negative

Overall thesis

• The 4-phase AI roadmap. Phase 1 (scarcity) — sellers into demand win on bottleneck pricing power; Phase 2 (normalization) — supply catches up and only durable sellers survive; Phase 3 (buyers) — hyperscalers take power and monetize AI across their own products for 10–20 years; Phase 4 (software bifurcation) — distribution/data winners rerate up while commoditized single-feature SaaS rerates down. He classifies every AI name by where it sits.
• Concentrated in the phase-3 buyers — META now a top-3 conviction position. "Amazon and Meta are my two top picks this year. Not Tesla, not Nvidia." He holds Amazon ($188k), Google ($130k + $90k), Microsoft ($70k + $21k), and a brand-new Meta stake he's built to $180k — now his third-largest. He keeps adding to Meta into a ~$28k drawdown ("my biggest loser by far"), staying "fully bullish."
• Own SOME phase 1, but only a durable seller — and trim it when it gets stretched. His pick is ASML (his best performer, +177% / +$103k) — an EUV monopoly selling decades-lived machines + service. But at a near-50× forward PE / 1.3% FCF yield he did his first-ever ASML sale, a 10% trim, to fund the new buys — "I am not bearish," still ~$150k held. TSM/Nvidia are "becoming less cyclical," but the parabolic cyclical sellers — memory (MU/SanDisk), construction/equipment (CAT), networking/power/cooling (VRT/ANET/CRDO) — he won't chase; they rerate lower when supply catches up.
• "AI is a commodity" is only half right. The models are commoditized, but the moat is the wrapper — distribution, integration, trust, full-stack control. So the hyperscalers that aggregate AI demand extract the pricing; spending on capex to stay full-stack (not "renting" tools from OpenAI/Gemini/Anthropic, the bind Apple is in) is strategic, not waste.
• Culture / recency-bias discipline. Bad-culture and "lost users" narratives track the stock price and reverse on recovery (NFLX, SHOP, GOOGL all did) — he discounts them and separates one-time, explainable metric drops (Meta's 20M = an Iran outage + Russia WhatsApp block) from real deterioration. He cites ARK Invest / Cathie Wood as the cautionary momentum-chasing case (Morningstar's top wealth-destroyer, ~$14.3B erased).
• Phase-4 software, picked apart. Distribution/proprietary-data winners rerate up — S&P Global ($100k+), Moody's ($47k), plus AI-insulated SHOP/ABNB and the debated-but-kept SPOT/DUOL (learned-behavior lock-in). Commoditized single-feature SaaS at bundling/AI risk — DocuSign, Zoom, and the "less predictable" CRM/Adobe/Intuit/Autodesk — he avoids.
• The rotation into left-behind quality (the rest-of-2026 game plan). "While everybody's zigging, I'm going to be zagging." With S&P breadth negative ex-AI/energy, he's buying forgotten non-AI compounders the crowd orphaned — initiating Uber and DoorDash ($10k each, building to $30–40k via weekly adds): dominant US share (~70% rideshare / 60–70% food delivery), Costco-like membership engines (Uber One 50M / DashPass 35M), inflecting free cash flow; AV/Waymo risk he treats as already-priced and being aggregated away (Houston robotaxi w/ Lucid + Nuro). Cites Peter Lynch's "dull, out-of-favor" rule; the orphan list keeps growing (Spotify, FICO, Copart, Netflix, Meta). By early July he formalizes it into a "9 best stocks to buy in July" list — META, AMZN, MSFT, NFLX, UBER, DASH, Copart, Constellation Software and Mastercard — grounded in a sector-concentration screen: semis are now ~20% of the S&P 500 (vs ~8% at the 2000 peak and never above ~5% for two decades), "everything else" has fallen from ~98% to ~80% of the index weight, and since the short-term market "is mostly a reallocation game," that everything-else bucket is being sold indiscriminately to fund the semi trade.
• The first crack — AI cost crossing into the consumer economy. Micron's blowout quarter is >90% price, not volume, and that pricing now forces Apple (Mac/iPad +15–25%) and Microsoft (Xbox +$100–150) to raise consumer prices — "strictly inflationary," dragging units and the tech sector, and the cycle's "first major warning sign." He separates such flow-driven drawdowns from real deterioration (Netflix's chart identical to Spotify's = rotation, not fundamentals) and discounts price-tracking media narratives (the decades of wrong Meta/Facebook headlines).
• The "stay the course" discipline (late-June portfolio update). Both funds are down ~8% YTD and ~15% behind an S&P pulled up almost entirely by AI/semis + energy — but he diagnoses it as multiple compression (≈ −20% on a flat ~16% growth engine), not deterioration. His toolkit: a Peter Lynch anchor (long-run "100% correlation" between EPS growth and price; the McDonald's 1985→2026 chart), a look-through table treating the whole book as one company (~16% rev / ~15% EPS, "earnings power owned" $53k→$61.6k), and a three-way moat test (an AI model, his own view, Morningstar) that lands stable-to-strengthening on nearly every holding — META/GOOGL the clearest improvers, DUOL the lone moat "under test." Verdict: hold patient, no high-grading; "over time the price will follow the earnings."
• Anti-perma-bear, pro-conviction. He methodically dismantles Jeremy Grantham's CNBC 70%-crash call — the Buffett indicator (market-cap/GDP) is outdated (global revenue, higher margins; Buffett dropped it 25 years ago), the "2 weeks to 2 years" timing is unfalsifiable, and Grantham was a decade-plus wrong (called the 2000 top in 1995, bearish through the 2010s) — and applauds Joe Kernan's rare on-air pushback. The flip side is reaffirmed conviction in Meta & Amazon as his two favorite 2026 buys (Meta's glasses + super-intelligence capex = a moat increase), while the Polen Capital collapse (~$85B→~$33B AUM) is his standing case for a nimble solo book over big committee-run firms.
• Mastercard — his largest position & deepest single-name conviction (Dec-2025 Deep Dive). A ~$200k stake, "even bigger than Google," framed as an "asymmetric bet." The reframe: MA is not a credit-card company but a "protocol of trust" selling an equilibrium to every transaction's "trust deficit," walled by a chicken-and-egg network moat and a rewards-as-hostage hold on high-end consumers. He waves off crypto/stablecoins ("efficiency = finality," which strips consumer float/chargebacks/recourse) and the real threat — government A2A rails (UPI/Pix/FedNow, domestic-only & fraud-ridden) — because MA sells fast-growing Value-Added Services (~20%/yr, $12.5B TTM, "faster than Google") on top of any rail. He's more bullish on MA than Visa (Visa "doubles down on the core rail" the government will "squeeze like a python"); worst case is a slow regulated-utility drift — "no bad story," far more upside than downside.
• The AI "motivation framework" — bet on the commoditizers (Jul-2026). He sorts the model-owners by whether they want the model layer commoditized or premium: Meta (maximally — monetizes via 3.5B-DAU distribution), Google (in between), Amazon/Microsoft (cloud commoditizers) all want models cheap/interchangeable; OpenAI and Anthropic need them premium ("the model IS the monetization layer"). His call: in the tug-of-war "the hyperscalers win" as models grow "cheaper and less differentiated," so Meta's cheap-model push (Musepark 1.1, +21% in 15 days on a $176k position) is strategic, not waste. Same lens makes him lean negative on the pure-model labs.
• Netflix reframed — playbook, not panic; and the semi trim continues. Into Thursday earnings he treats NFLX's ~50% fall to ~$74 as a media "story," not deterioration: adding podcasts/live-channels/bundles is Netflix's lifelong content-expansion pattern (the always-on-channel idea is 5+ years old), the engagement dip is real but "addressing issues is what Netflix does" (325M subs, low churn) — so any slide "from the 70s into the 60s" is a buy. He also made a second pair of ASML trims (45 PE) while "still very bullish fundamentally on ASML and TSM," reaffirming the semiconductor-trade-slowdown call and the rotation into quality-at-low-valuations (Uber, Netflix, Meta). Into earnings he's bullish the big banks (JPM/BAC/GS — trading, credit cards, rates held up) and expects the 12-state Paramount–WBD antitrust suit to fail (DOJ already cleared it).
• The serial capex re-rating — and buying the last holdout (Jul-2026 earnings week). The thesis stops being a forecast and starts printing: Microsoft (−17% YTD on capex fear) beat EPS by 13% and rose 15.5% in a single day — "moving like it's a penny stock" at a $3.4T cap — and Amazon followed (+4%, then +9% after hours) on AWS +37% to $42B in a quarter, at rising margins with demand visibility out to 2028. That's the same arc Google walked over the prior year (15→25 PE, +100%), so the pattern is explicit: the market resolves the capex controversy one name at a time, on earnings evidence — which makes Meta, the last holdout, the place to add. He bought another $4,000 of META down 22% / −$38k, decomposing its 14% EPS miss into a $2.4B legal charge + $1.18B of May severance + faster-than-peer depreciation ("taken out, their earnings per share were in line") against +27–28% revenue to $60.8B and a record 3.6B DAUs that buries the "peak social media" narrative. He grants Meta is genuinely later (no public cloud to make the spend billable) but not different in kind. The allocation rule sharpens: keep a leg in ASML but never concentrate in semis — "commodity-like… scarcity-driven," running on "one-time scarcity-driven volume" supplied to the aggregators he owns. Apple (unowned) reported "fine"; fail of the week is Leopold Aschenbrenner's Situational Awareness, the ~$20B fund that levered a concentrated bottleneck-AI bet and sold its book to Ken Griffin — the mirror of his own "I don't have any leverage and I have a lot of time to wait."
• The capex bear case breaks — and the thesis gets graded (Aug-2026). "Sentiment is changing rapidly for the big hyperscaler stocks"; the bears' telecom analogy (Michael Burry et al.) is abandoned in a week. The cause is specific and he attributes it by sequence — "it wasn't after Microsoft reported that Google and Meta went up. It was after Amazon reported": AWS +36.7% to $42.2B, the fifth straight accelerating quarter and fastest in 18, a $169B run-rate (3× Netflix, ~70% bigger than all of Tesla, half of Berkshire and closing by 2029), backlog +154% to $496B with capacity through 2028 already committed, and $16.6B of operating income at ~39% margins — enough that he takes Jassy's "trillion dollar revenue business" claim at face value. He grades his own Sept-2025 call (AWS reaccelerating to ~25%, stock to 240–260) against a 36.7% print and a $285 stock: "I underestimated it to a huge extent." The book runs to all-time highs ($1.03M / +$374k) with GOOGL $223k and AMZN $207k "in a neck-and-neck race," both deliberately overweight. New this episode: FICO (never owned) as the mirror-image case — dominance is not enough once a 100× multiple meets real fear, the regulatory moat erodes and a distributor war drags Equifax down with it — versus S&P Global, "not as reliant on one thing going well" (11% organic, ratings + indices + Market Intelligence). Into the week's prints he's positive Uber/DoorDash (the equal-sized split buy, judged on scaling not profits) and Texas Roadhouse (beef + labour inflation normalising = "an incredible margin expansion story"), and holds every share of Duolingo at −$13.4k on the asymmetry of stocks (Peter Lynch's 6-of-10) while conceding "I'm not 100% confident." He declines to borrow Tom Lee's bull case (S&P 7,800 this month, 2027 "one of the best years") — "more just looking at the bright side of it" — grounding his own 2027 optimism in hyperscaler demand, though he does carry Lee's SpaceX unlock overhang as real downward pressure on the market this year.
• Rotating the buy list forward — three new dislocations (Aug-2026). With the S&P and QQQ near all-time highs, he makes the discipline explicit: a name you hold is not a name you're urging capital into. MSFT ($507, +~120% off the Jan-2025 $220 highlight), ASML ($1,750, +~150%, a $124k position) and AMZN (+22% YTD, a $200k position, +$77k) come off the buy list precisely because the dislocation closed — "I'm not highlighting this one as great value… cuz it's already gone up" — leaving the rule intact: quality is only cheap "at time periods where there's something wrong." The three replacements: UBER (24 PE / ~4% normalized FCF yield on ~20% revenue growth, priced for AV destruction — but Waymo's 500k weekly rides is 17 minutes of Uber volume, "600 times bigger" and widening, and because Uber owns per-city network density everywhere it "can be late to the game with AVs and still be in the lead"; $24k bought, targeting $150–160 from ~$70); META (−21% on the year, where the legal risk is "the smaller of the bear cases" and the capex fear is the whole story — answered not with better-ads-via-AI but with Zuckerberg's same-day essay: never be "reliant or bottlenecked," the Apple App Store toll repeating with Anthropic/OpenAI as "the Apple and Google of AI" — with a neocloud salvage floor that "would pay for the investments that Meta's already made"); and NFLX (−37%, where the engagement narrative is refuted claim-by-claim — season-1→2 drop-off improved, subscribers grew, total engagement +2% YoY, and the per-user decline is a mix effect from expanding into low-TV-consumption markets). He also takes the other side of SemiAnalysis' "Gemini is cooked" — the same shop called Gemini state-of-the-art in November, so odds of "zero" are indefensible — while agreeing with its second half: Google is fine at second tier because it is redirecting spend from model training into Google Cloud (possibly +100% growth, operating margins now 35%). Fail of the week is Keith Rabois/Opendoor: stock comp called "fake" in 2021, rebutted by Uber's own CFO, settled by a chart that has "gone nowhere" for five years — his standing screen against adjusted-earnings storytelling.
• The Ackman letter as independent confirmation — and the "zero test" for AI fear (Aug-2026). Pershing Square's Q2 report adds six new positions and three of them — NFLX, MA, SPGI — are already big Carlson holdings, joined by V (bought with MA as one payments trade), ICE and ALC. He is careful about the direction of causation: META/GOOGL/MSFT/AMZN/NFLX/MA/SPGI were all his before Ackman's disclosures; only CMG and UBER were follow-ins. What matters is the common shape — each name is a quality compounder marked down on an AI-disruption story, and the method for sizing that story sharpens into a repeatable test: name the product actually threatened, pull its share of revenue, then assume it goes to zero. S&P Global fell 25% on Anthropic's Claude Co-work launch (25×→19×, a five-year low) even though Capital IQ is under 7% of revenue and 40%+ of Market Intelligence revenue is proprietary data embedded in workflows — annihilate the whole contested line and the forward P/E is still 26. Netflix's "engagement problem" decomposes into watch-time quality (live drives sign-ups and retention out of proportion to hours) plus a geographic mix shift into low-TV-consumption markets. Mastercard/Visa's two feared disruptors both grow the pie: stablecoins win "where cards are not the incumbent" (cross-border B2B, remittance corridors, dollar savings in weak currencies) and agentic commerce adds transactions because agents "adopt, not replace, consumers' existing payment preferences." Meta's capex gets a resale floor — unused compute "can be profitably monetized in a highly supply constrained compute market." And the frame for all of it is Ackman's breadth statistic: semis + tech hardware, 8% of companies and 22% of market cap, delivered ~85% of the S&P's first-half gain, so 90%+ of the index contributed under 2% — dispersion Ackman used to deploy $5B of cash into exactly the bucket Carlson has been buying.
• Uber, fully argued — density, peak demand, and why the hybrid wins (Aug-2026 deep dive). The AV bear case gets a structural answer rather than a valuation one. The reframe: Uber is not a ride-sharing company but "a physical asset allocator" — it entered every city where roads, fuelling, cars (idle ~95% of their life) and willing drivers already existed and simply made them communicate, which is why it scaled globally with almost no capex. The unit of competition is per-city density, not scale: enough local supply that any rider is picked up in minutes and any driver stays utilized. Lyft never cleared that bar ("they're out of it"), and — decisively — "nowhere has Waymo knocked them out of critical density," even in its most mature cities. Waymo is credited as smart, Google-funded and the only real AV threat, but it is scaling a physical business: ~$100k lidar-equipped vehicles, 3-D city mapping, permitting, parking and maintenance, then years of waiting for the flywheel — and its San Francisco proof point is a demographic outlier (≈2× LA's density, 4–5× the US top-50 average, ~2× as affluent, ~half the trip distance) that shouldn't be extrapolated. The knife is peak demand: Friday spikes ~6×, so an AV-only operator "must hold significant… underutilized supply… or deliver worse consumer experience and leave significant demand on the table," while humans supply the peak for free because they park their own cars. Hence both conclusions — Uber can plug any number of AVs into a dense network at 100% utilization from day one ("Uber can be late and still be first"), and a hybrid human+AV network is a better product than any 100% AV network even at the finish line. AV software is explicitly not a durable moat — "it's no ASML machine" — so supply commoditizes and, as with NFLX/YouTube/BKNG/the app stores, "the aggregator usually wins." Tesla's robotaxi is the one theoretical fleet that could cover the peak, but "so far Tesla's not using lidar… the whole dream of it seems somewhat fanciful." The honest bear case he states himself: a full decade of Uber inaction would let Waymo take critical share — so the real deadline is 5–7 years, "not nearly as severe as being priced into the stock," and the next five years of Waymo expansion headlines are inside the thesis, not evidence against it.
• The 13F season, read as method (Aug-2026). Four managers' quarterly filings, examined for how to read a filing rather than what to copy. Valley Forge (Dev Kantesaria) shows an all-reductions quarter — FICO, SPGI, MA (−28%) and V (−22%) sold into falling prices during "their longest streak of underperformance in their fund's history" — which Carlson diagnoses as redemption selling, not conviction selling (the confirming fact: Kantesaria is "on the record saying that they hold almost no cash"), with the uneven cut sizes as the residual signal that even a forced seller reshapes. His one disagreement is trim-selection: sell Moody's, which is holding up better on a higher multiple, and keep the already-marked-down S&P Global. The untouched name, ASML (+62% YTD at a stretched multiple), earns its exemption as "the one confounding variable in his portfolio" — reinforcing Carlson's standing risk-factor audit: five names, one rate-sensitive Achilles' heel, versus his own deliberately uncorrelated book (MA without V, plus META, ASML, GOOGL, SPGI, COST, MSFT, TXRH, DASH, UBER). TCI (Chris Hohn) supplies the episode's sharpest disagreement: a 17% Microsoft position taken to zero on a moat mechanism — Claude disintermediating Office, with Azure as the domino behind it — rotated into Google (+12%, ~8% combined), the exact reverse of his 2018 trade. Carlson concedes the direction ("certainly transitioning some of the value to Claude's interface") but not the magnitude — "I'm not so concerned about Microsoft that I'm going to be selling my position… In fact, I become more bullish on Microsoft over time" — and sidesteps the pair trade by owning both. Pershing (Ackman) makes UBER its top holding at 12.72% with META added 20% to 9%, which Carlson reads as the two-name recommendation implicit in weighting: "if Bill Ackman was to tell you to buy two companies, it would be Uber and Meta, which I agree with both. I've been buying both." Dorsey independently buys UBER, adds 24% to META, builds AppLovin (47% growth) and Booking, and trims ASML — where Carlson discloses he is trimming his own on a stretched valuation while drawing the counter-lesson from Dorsey's $1,300 trim before the run to ~$2,000: "you can be very slow to take gains in stocks like this." The news half applies the same first-principles discipline to the 29-state Meta trial in Oakland (COPPA + consumer protection, framed by the press as social media's "big tobacco moment" after a ~$1B New Mexico loss): under-18s are 2–4% of DAU and less of ad revenue ("it's the parents that buy things"), five years of teen-account tightening are already shipped, so the weighted outcome is a fine of "multiple tens of billions" plus stricter teen controls, with a forced algorithm overhaul reaching adult accounts at "around 10%, maybe less." Cambridge Analytica ($130 → ~$560, a ~400% return for buyers of that panic) is the base rate — so he keeps holding and keeps buying META.
• The buy-target board — the whole portfolio, priced in advance (Aug-2026). The most complete disclosure of the book he's given: 14 positions, ~$1.5M, +$540k, and — importantly for how he grades himself — a time-weighted 14.8% CAGR rather than the flattering "+50%" simple return ("it doesn't account for how much money you had in the portfolio at a given time"; a third of the portfolio's life was under $100k). Every holding gets a hard, pre-committed buy price strictly below the market, and $10,000 goes automatically to whichever hits first — the emotion is removed by fixing the board in advance rather than by choosing well in the moment. Each target is derived the same way: take the company's actual EPS growth and cut it with a named cause, haircut the exit multiple, then solve backwards for the entry that still compounds at ~13–21%. The board: GOOGL $250 (his largest position and explicitly not a buy at a 26 forward PE / 1.2% FCF yield), MA $450 (the one name whose 29 multiple he refuses to haircut — "capital-light… not doing any big capex"), AMZN $200 (19.3%), META $450 (17.2%), SPGI $350 (a target he expects to miss, and says so), ASML $1,200 ("one of my favorite stocks ever… But I have to set a buy-in target"), NFLX $60 (19.9%, the best on the board), MSFT $400 (simply giving back the 25% earnings pop), COST $600, TXRH $140, MCO $375, DUOL $90, DASH $120, UBER $60 (20.9%). Two second-order lessons fall out. First, the required discount is a valuation verdict on your own holdings: 18–25% is his normal band, so Costco at 38% below is an admission — "it is an overvalued company that I continue to hold," with the special dividends redeployed elsewhere — and DoorDash at 47% below is the same signal on his fastest winner. Second, portfolio-level odds beat single-name odds: each dip looks unlikely alone, "but when you start adding all of them together… the odds dramatically increase," and he expects a trigger within six months. Standing views reaffirmed: DUOL's AI bear case rejected on focus ("that's all they focus on"), and META defended twice — the drawdown as a rerun of "the whole ChatGPT scare," and the fail of the week going to Gene Munster's bearish CNBC take, graded against Munster's wrong ChatGPT-kills-Google call at $170 ("now you own Google… after Google's stock price went up 100%") with Meta now at an 18 forward PE and the litigation already public. The news half is a long defence of Uber against the More Perfect Union and Business Insider "personalized pricing" documentaries: they collapse dynamic (situation-based) into personalized (person-based) pricing — the seven-phones experiment is explained by request timing and driver availability, and Uber's audited filing enumerates the denial — and, his real charge, both cite Uber's technical rebuttal in a single clause and never show it. "Journalism shows both sides."
• The beaten-down screen — temporary or permanent? (Sep-2026). Three dips, each a different test of the same question. LULU is the headline case and the cheapest thing he's discussed all year: −51% YTD / −76% over five, forecast −11% EPS, and an 11.9 forward PE on shrinking earnings — "well below commodity multiples, meaning that Lululemon is currently being priced for capital destruction." That is the screen: not cheap versus its own history, but cheap versus the multiple the market reserves for businesses with no pricing power. Michael Burry (a person, not a position) supplies the thesis — Lululemon is "a much better brand than a management team," so the damage is marketing and merchandising, which is fixable, while the product is still wanted (Alo Yoga and Vuori are copying it; Costco sells dupes) — and the conviction: a 32% fund stake at a $177 average, personal buying all the way to $100, now his largest personal holding. Carlson adds the mechanical leg most miss: at this multiple the >5%/yr buyback becomes an engine, pushable "to 7% to 10%… upwards of above 10%" on almost no dilution with $1.39B of cash — the CRM leveraged-buyback playbook — so a merely stabilising business still compounds EPS. Precedents cited: ANF ($30→$185→$83→~$150) and RL. ISRG is the mirror image — no valuation cushion (31 forward PE, 2.56% FCF yield) but a model he rates highly enough to break his own no-medical rule and plan a deep dive: "a printer and the ink type of company" where the system sale drags perpetual single-use instruments plus service revenue, keyed off 2,000 new installs TTM; the derating is deceleration (20% → ~13%), not deterioration. ADBE is the AI-disruption test at a sub-10 forward PE — "basically being treated like a car company, like Ford… getting close to the airline category. And it's a high margin, faster growing software company" — with a >10% FCF yield (8.5% after stock comp), 12.5% growth, and a prediction into Thursday's print: 14 straight quarters without a miss, so "beat by 2 to 3%," and sentiment low enough that a glimmer of guidance moves it hard. The news half converts the Aug deep dive into a one-paragraph rebuttal of the post-Cybertruck "UBER is cooked" consensus (Ross Gerber): the missing variable is highly variable demand — a fleet owner must either under-serve the rush-hour spike or hold vehicles idle all day, while Uber's drivers supply the peak from their own garages for free — so the hybrid network is the endgame and Uber already runs it (driverless vehicles on the network now, against TSLA's 40 on the road). Fail of the week starts as Gemini and reverses on air onto the Mount Shasta hikers who missed the noon turnaround and then blamed the AI — with the conflict declared first: "the Gemini from the Google that I know and love. I've made a lot of money with the investment."
• The "AI slowdown" — Meta as the beneficiary, Anthropic's pacing call as regulatory capture (Sep-2026). After a departing Anthropic researcher's extinction warning and Dario Amodei's We Must Pace the Frontier essay (a coordinated, regulated, industry-wide slowdown, partly echoed by Sam Altman) knocked ASML/NVDA lower, he runs two analyses. Who wins a forced pause: screen for the heaviest AI spender with the least direct AI revenue — AMZN (AWS), GOOGL (Cloud) and MSFT (Azure) rent their capacity out, so META is the one whose capex falls without its ad monetization slowing; margins and FCF expand, the "Zuckerberg will spend endlessly" discount lifts, and a 21 PE on 27% TTM revenue growth (flat all year) rerates. Via Colossus he adds the heads-I-win-tails-I-win frame: if models commoditize, value flows to Meta's distribution/attention/commerce; if not, Meta sits within six months of the frontier with its own model rather than "in his competitor's prison" like AAPL — so the spend reduces risk. Execution evidence: the agentic Muse app (OpenTable bookings, 10–15-minute shopping runs checking out via Stripe, 4.9 stars), MSL rebuilt in a quarter around the $14.3B Scale AI stake, and a founder-controller's record of winning every platform transition. Why Dario wants it: the unilateral-action test — if the models are that dangerous he could slow Anthropic "without an act of Congress," yet the ask protects "commercial advantage." With ~$500B of commitments, promised profitability and an IPO ahead, Anthropic and OpenAI need to cut training spend anyway, and a regulated pause freezes cash-rich followers (Meta, Gemini, xAI) behind them: "the regulatory capture will be complete." He also rejects the Manhattan Project analogy (a weapons programme; AI is the internet) and the Hugging Face agent story (a containment failure), and notes that slowing a technology Dario says could cure most diseases within a decade has its own death toll. Fail of the week: the Hunter Biden laptop memecoin (~$190 → under $2 in minutes, ~80% of buyers down, 10 wallets up big) — "zero tears shed by me."
• Guest thesis — Zeta Global as "the Palantir of the demand side" (Aug-23 Investor Exchange). Not Carlson's pick: a member with 42% of his portfolio in ZETA pitches an agnostic marketing cloud whose moat is an owned 240M-American identity graph (Disqus, LiveIntent) matched to clients' data to predict ad returns before spend, growing through wallet share inside clients it already has (1.7% of Fortune-100 marketing budgets, NRR 120–128) and a PLTR go-to-market partnership. Carlson interrogates the privacy-backlash risk and the "next Palantir" label ("there hasn't been any next Amazon") but calls it "a really good pitch" at ~26 PE on 36% growth — and restates his own discipline: trimmed ASML ~20% near $2,000 to fund DASH/UBER, holds COST regardless of valuation.
• Levered buybacks, AppLovin and the streaming hub (Aug-27 six-stock update). CRM's +22% day is decomposed — ~8–9% organic, EPS flattered by Anthropic gains — and its real story, ~$25B of ~5% debt retiring ~15% of shares in a year, is worked through to show EPS rising (~$12.50 → ~$13.16) and improving as the business grows: "smart financial engineering," though he keeps the Meta he sold it for (3× the organic growth) and wants every stock he's sold to do well. NVDA's supply-constrained 70% guide: "I far underestimated" demand, still prefers GOOGL/AMZN durability. First looks: APP (−60% on +47% growth; setup liked, researching) and SPCX (unmodellable at $1.84T — fine as a small bet). Positive RDDT ("a good investment at a good valuation") and NFLX (a hub selling other streamers' subscriptions for a cut).
• AMA answers (Sep-1). UBER vs Waymo: Waymo is "the real deal" but Uber keeps peak density even in LA — erosion at the edges, both can win; TSLA's robotaxi "hasn't proven anything." GOOGL held at ~$342 ("a normal valuation"), add around $250–260. The AI build-out is not a scam — OpenAI/Anthropic revenue is paying demand; crypto is the overhyped one. He deliberately doesn't flood the public book with Qualtrim income (the goal is showing returns, not deposits) and parks extra cash in SCHG, now VTI.
• The expected-return dashboard (Sep-5 portfolio update). $1.48M, $528k lifetime gains, 200% time-weighted, flat YTD vs S&P +12% (Berkshire beat the index in only ~66% of years). A five-phase AI cycle — build → monetize → distribute → commoditize → owners of the customer — makes phase 4 "the most dangerous" for bottleneck sellers and tilts the book to aggregators. Only trade: $1,000 of UBER at $67.66 on the Aug-5 low. Each holding gets a five-year return (EPS growth × exit multiple): above the 15% hurdle META 18.14%, UBER ~17.5%, NFLX 15.77%; close MA (31× — the highest multiple), AMZN 13.48%, SPGI, DUOL, MSFT 12.23%, DASH; below GOOGL 9.6%, MCO, TXRH 8.86%, ASML (fair value ~$1,300–1,400), COST 5.47%. New money follows the ranking rather than churning the book; a clean-slate $10k goes META 30 / UBER 20 / NFLX 15 / MA 12.5 / AMZN 10 / SPGI 7.5 / MSFT 5 (15.95% weighted). Uber aggregates autonomy (Wayve, Zipline) while DoorDash builds it (Part 135 drones, Dot robots).
• Against the AI-doom moment (Sep-10). Jacob Coxin's viral Anthropic resignation, a media tour and Bernie Sanders' "corporate death penalty" bill are "highly manufactured and contrived": the Hugging Face "hack" was instructed agents, extinction-level bioweapons and infrastructure takeovers are exaggerated (materials, skills and fragmentation are the real barriers), and the remedy looks like Anthropic regulatory capture that China won't follow — the setup for his Sep-14 Meta-benefits-from-a-slowdown call.
• Live AMA — asymmetry, scale and risk factors (Sep-15). META: "very asymmetric" — $900+ on a rerating from ~20×, Muse "wowed" him, Meta One diversifies ~98% ad revenue, and $180–200B 2027 capex is a rational land-and-permits grab. UBER: completes Waymo's weekly paid trips in ~13 minutes, lead widening, executive insider buying — the Google-during-ChatGPT setup where the bear case is priced in before it happens. AI and software: most names are messy "both" stories; consumer INTU looks dicey, enterprise NOW/CRM safer, and disruption forecasts keep missing (NICE revenue still growing). Watching HOOD (deep dive, not a buy), ISRG, SN; CASY's −30% "a bit of an overreaction" but not cheap. Risk is managed by factor, not count (~7 names removes ~90% of specific risk; AMZN "almost like an ETF"), and he now buys VTI so his ETFs don't repeat his stock book's single opinion.
• The agentic-takeover map (Sep-18). Seven months after Citrini Research's 2028 Global Intelligence Crisis, Meta's Muse (OpenTable bookings without the app, a whole allergy-aware Smith's grocery cart at the lowest prices) convinces him agents are real, so he re-runs Citrini's exercise on his portfolio. The risk is "not automation" but losing the customer to an agent that becomes the demand gatekeeper; companies are scored on how much value comes from removing tedious steps versus how much scarce execution remains. Most exposed: BKNG (sold; marketing ~30% of revenue, so the threat is a gatekeeper swap rather than a volume collapse; OpenTable "super vulnerable"), INTU (sold; TurboTax "just a big form," filing becomes ambient), LZ, SLQT/GOCO/EVER, Z — the Citrini idea of habitual intermediation. Mixed: DASH (kept — it executes delivery, OpenTable only coordinates), UBER (density wins the agent's pick), AMZN, DUOL, and his largest holding GOOGL (search ads monetize declared intent an agent compresses, but distribution, intent data and the Universal Commerce Protocol defend it). Resilient: MA ($189k, still adding — more agents mean more need for trust at the transaction), V, ASML ("the apex predator"), SPGI/MCO (authoritative, licensable output), NFLX, COST, TXRH, CAKE, FICO. Beneficiaries: META (feed ads create intent, so it's more insulated than Google; "I've been buying Meta like crazy"), MSFT ("the agentic operating system," defended by permission context), SHOP. The falsifier he names: agents that never beat the incumbents' own interfaces.
• The agentic wars start: Amazon blocks Muse (Sep-21). META jumps ~9–10% on the day and 33% in a month ($534→$722), taking his stake from −$30k to +$13k. He calls it the textbook case of his setup: a low forward P/E, fast growth, one of the widest moats and the attorneys-general suit settled, so all it needed was a story change. Muse supplied it: #1 on the App Store, 4.9 stars, generous free usage, and 3.6B daily users to push it to. Wells Fargo raised its target from $640 to $796, and "Meta still has a lot more room to run." AMZN's block ("unauthorized AI agent") fails his control test, because a logged-out Muse is blocked too, so the credentials excuse is "bogus." The real motives are agents devaluing the sponsored listings behind a $76B high-margin ad business, disintermediating the customer touch point, and bypassing Amazon's Rufus assistant. He expects a negotiated, certified-agent deal, since a stand-off grows costlier for Amazon as Muse scales and WMT could open up instead. He reads the block itself as validation, "a massive victory" for Meta. GOOGL ("a massive Google shareholder") gets called "embarrassing" for letting Meta ship first despite owning Gmail, Docs and Maps. He also takes a victory lap on the Sep-14 call (via Yahoo Finance), grades Gene Munster's month-old bearish Meta call, and makes Andrew Yang's unsourced "self-replicating code polluted the internet" CNBC story the fail of the week, faulting its sourcing chain, its mechanics and Yang's track record.

Transcripts

One dated page per appearance — each has its full stock table, talking points, and the saved transcript. Newest first.

To process — backlog

Joseph Carlson appearances discovered via YouTube search (Joseph Carlson), not yet processed — verify publish dates & channels, newest first. Limited to the last ~2 years. None queued yet.

For personal study — not investment advice. Source material © The Joseph Carlson Show.