The AI trade Crowded — downside ▼ into the Nov-3 midterms, but the bear window is dated, not open-ended

Woo: the AI rally (Nasdaq 100 +20% since the April Claude release) is "very crowded." Q1 hyperscaler CapEx fell vs Q4 (more in volume terms given memory-chip costs), frontier capabilities are "plateauing," competition is closing in on Claude Code (Cursor/Copilot/Codex; MSFT canceled most licenses), and "token maxing" is yielding to budget-constrained usage. Chips (INTC/AMD/AVGO) lost momentum; balance of risk favors the downside — long-dated gold calls as the hedge. Gromen: US equity valuations are in "La La Land" — his adjusted Warren Buffett metric (equity mkt-cap minus federal debt / GDP) is the highest in 65 years (above 1Q-2000 & 4Q-2021) — "a terrible risk-reward." Muir: sentiment is maxed (lowest FMS institutional cash ever; record AAII actual exposure) while QQQ/S&P go sideways — Kovner's "consensus not confirmed by the market"; Asian-retail AI buying is the final gasp, so buy your straw hats in winter (cheap hedges now). Singh: FCF has died as a metric (Mag 7 FCF ≈ zero on capex, SPV debt) and circular financing (Nvidia→neoclouds→GPUs; OpenAI↔Oracle; the $2B Marvell pump) inflates demand — the unwind cascade runs opto/memory → GPUs → power, triggered by a hyperscaler capex cut; only TSMC/Korea/ASML capacity discipline is prolonging the cycle. Expect 10–15% index "hiccups" within two years. Singh (Jun 14): the circularity is levering up — the $35B Apollo/Blackstone "Big Sky" SPV has Broadcom backstopping $30B of Anthropic chip-collateral debt whose proceeds flow to Nvidia (Google co-makes the chips and owns 14% of Anthropic; lenders shown no financials), while Citadel's tokenomics work shows inference costs biting (Microsoft canceling Claude, Uber capping its AI budget) — frontier models concentrate in the few biggest balance sheets, the rest trade down to cheaper models. Rasgon (bull): the semi rally is all earnings — SOX multiples actually compressed YTD, so a bear must explain why earnings are unsustainable; Burry's depreciation case fails (old-GPU rental prices are rising, fully-depreciated GPUs very profitable), capex keeps being revised up (Google $180B / Amazon $220B / Meta $135B — existential for hyperscalers), and monetization is visible (Anthropic ~$1B→$44B annualized run-rate in a year; agentic coding at "takeoff velocity"). The real bear case — no/poor return on AI — can't be front-run via capex cuts (too late by then); watch token-usage proxies. Meanwhile the Nvidia/Broadcom lag vs the bottleneck names "has to normalize," likely upward. Paulo Macro: the "Great Circularity" — even the picks-and-shovels hardware players' "cashflow" is VC-sourced cash funneled through the LLM labs via capital raises and relaundered back in a flywheel; "it all works splendidly until VC flips from buyer to exiter and someone asks for their money back." Gundlach (Jun 12): a late-cycle top on concentration — the Shiller CAPE is near its all-time high (matching the dot-com peak) and the "AI Big 10" are 41% of the S&P in just 10 names (the Nifty-50 '72 / Japan '89 / dot-com analog), and it's "already reversing"; the Philadelphia semi index's record April monthly return rhymes with February 2000, and a coming data-center over-build (construction has crossed above office) sets up the bust. "A very bad time to be doubling down on this concentrated portfolio." App Economy (NVIDIA, May 22): the bull counter — NVDA Q1 FY27 +85% Y/Y with Data Center 92% of revenue and a 75% gross margin held through the Blackwell ramp ("margins are the lie detector"); demand is broadening beyond hyperscalers (the new ACIE segment +74% nearly matches Hyperscale +115%), China is modeled at zero (a free call option), and "compute is revenue" — the cycle "is still expanding, not digesting." Eisman (Jun 12): the deeper problem is no moats — despite ~$1T of hyperscaler capex this year, the LLMs are undifferentiated ("one week Gemini is on top, the next Anthropic"; users switch constantly) and OpenAI is reportedly cutting token prices, so "trillions are being spent for a product with no moats and prices already being cut" — the capex builds a commodity, not a franchise. Larson (Jun 16): not bearish but defensive (~20% cash as "optionality / bullets") — own the physical backbone of AI (power, copper, cooling, energy — "halo" assets with low obsolescence), not the chips; semis (SOXX) are "the muscle" and a "gambler's paradise" of 2–4× levered derivatives he's trimming. The melt-up "runs on gamma" (dealer call-hedging — "when the Hulk runs out of gamma he's just Bruce Banner") with the Buffett indicator at 240% of GDP and a SpaceX/Anthropic/OpenAI issuance wall reversing 20 years of shrinking float. Software is where AI is a "wrecking ball" (avoids Adobe); even defense is "rotational money leaving" after the war — wait for the washout. Currie (Jun 17): tech now "looks like commodity guys" — AI compute is a cyclical commodity (an oil refinery: data + electrons in) and the hyperscalers are "putting steel in the ground," yet trade ~30–40× when capital-intensive cyclicals get 10–12×; capex is nearing 120% of cash flow (free-cash-flow yields "likely to go negative even at Google") — the same signal oil producers flashed at the 2014 top, in reverse. "Rising ROC will be the oil guys, not the tech guys." Polomny (Jun 17): "Datacenters Are the New Shale Oil" — AI datacenters are "deeply negative-ROIC investments," and the funding ladder is the tell (cash flow → debt → now equity: GOOG's first major issuance since 2004, META reportedly next); MAG7 are "no longer the high-margin, asset-light businesses they once were" — profitless growth in a revolutionary industry that "goes on until it can't," then the crash and a capital reallocation "into our favored sectors." Eisman (Jun 18): the bear case has "sharpened" into six legs — (1) capital intensity (software cash-machines now must raise equity), (2) no moats (weekly Gemini/Anthropic/OpenAI leapfrogging — "trillions spent for what looks like a commodity"), (3) price cuts (WSJ: OpenAI mulling lower token prices), (4) token-pricing pushback (usage pricing; MSFT Copilot Jun 1; Uber burned its annual AI budget in 4 months), (5) no ROI evidence, (6) power/water limits + community pushback; the bull leg holds only while Nvidia's revenue (+85%) stays elevated. His tell: own the suppliers (power, semis, networking — Arista/Cisco), not the "airline"-like hyperscalers. Sacerdote (Sohn, May 12, the structural bull): we're in "the first batter's box" — only ~10 bips of ~1B white-collar workers use AI the advanced (agentic) way and Claude Code is ~14M DAUs heading to ~500M, so adoption is "an L-curve, straight up" with "half of what we need" in compute. Own the model layer (Google + private Anthropic/OpenAI ≈ $200B combined run-rate revenue at "staggering," compute-locked-in margins, ~18× earnings) and the chip layer; bifurcate software — bearish horizontal apps, bullish data/infra (DataDog, "Anthropic uses their tools"). Carlson (Jun 17 / May 26): a 4-phase rotation — scarcity-phase sellers (NVDA/MU/TSM/ASML/AVGO/VRT/CAT) lead now on bottleneck pricing power, but over ~1–2 yrs power shifts to the hyperscaler buyers that own the customer and monetize AI for 10–20 yrs (GOOGL/MSFT/AMZN/META); own SOME durable phase-1 (ASML's EUV monopoly) but avoid the cyclical memory/power/cooling names (MU/SNDK/VRT) that rerate lower when supply catches up. META is the left-behind value — ~18× fwd (vs S&P 21.5×, QQQ 27×) growing rev ~25–30%, the "broken-company" culture/users/capex FUD misleading — META & AMZN his "two top picks… not Tesla, not Nvidia." Singh (Jun 21): the pricing power is cracking — Microsoft moving Copilot to usage-based pricing and adding a self-hosted DeepSeek v4 (~⅓ the input / ~1/7 the output cost of Claude) is "an absolute watershed" that strips OpenAI/Anthropic of pricing leverage (JPMorgan blocked Claude in Hong Kong on cost), so he's heavily short the listed pre-IPO wrappers (VCX, DXYZ) holding their secondaries; offsetting near-term, confirmed data-center delays hand live-capacity operators (CRWV, NBIS, DGXX) premium lease rates, so the bubble has "at least a year" left. Paulo Macro (Jun 23): Korea is "ground zero" for the narrow semis bubble — a KRW collapse despite a skyrocketing Kospi, explained by foreign EWY outflows in May (foreign exit vs domestic bid); Chosun reports SK Hynix slowing its HBM4 ramp to make higher-margin general-purpose DRAM, and the "Great Circularity" plus weak AI ROI (Uber) leave Mag-7-vs-SPX "feeling very heavy" as the stress boomerangs back onto the top-heavy indices. Eisman (Jun 26): the rollover went live in the Q2 report card — Micron's blowout sparked only a chips/power Thursday pop (GE Vernova) while every hyperscaler (AMZN/META/ORCL/MSFT/GOOGL) fell and "net net NASDAQ was down"; capital intensity is tolerable only with a moat — without one it's "a race to the bottom… price wars… low returns on capital" (Google's Gemini co-head left for OpenAI, a DeepMind engineer for Anthropic — no moats, no non-competes; software CRM/ADBE/INTU −18/19/39% and consulting ACN/IT the AI victims, memory MU/SNDK and AI-power GEV the scarcity winners). Halftime committee (Jun 25): the same read from the desk — "the AI trade is the buildout, not the hyperscalers"; eight of the day's top-ten S&P names are AI-infrastructure (semis/equipment, industrials/power/grid) while the Mag-7 are now net S&P detractors YTD (equal-weight Mag-7 −8%, the ex-Mag-7 "S&P 493" +14%), the spenders "footing the $800B→$1.6T capex bill" getting no shareholder credit (Microsoft's worst month since Dec 2000). Singh (Jun 28): the pricing-power erosion is accelerating — UBS finds ~60% of US enterprises (e.g. Uber) routing to cheaper open-source Chinese models (DeepSeek/Qwen/MiniMax/GLM/Kimi) via "model routing," Bloomberg shows the US-frontier enterprise share falling ~70%→~30% in a year as Chinese open-source rose ~15%→~45%, and DeepSeek V4 runs at <1% the cost of Claude Opus 4.8; OpenAI is delaying its IPO to 2027 (it is ~45% of Microsoft's cloud backlog). Carlson (Jun 26): the 'first major warning sign' for the AI cycle is that AI hardware-cost inflation has broken out of investor-funded capex into the consumer economy — Apple raising Mac/iPad prices 15-25% and Microsoft hiking Xbox $100-150 on memory costs — and the Mag-7 sold off on the read (AAPL −5%, MSFT −3.7%, AMZN −2.7%, META −2.2%). Sohn (Jun 29, Strategas chartist): semis (SOXX, now 18-19% of the S&P, up from ~2% a decade ago) + hardware/equipment + power (GE Vernova) are the ONLY leadership while software is the laggard and 'a real problem' — 'the market lives and breathes by' the semi index, and 'if it ever reverses there's no place to hide.' The mega-cap charts split: GOOGL best, AMZN 'an eh,' META/MSFT/ORCL messy-to-short (200-day slopes flattening or rolling over). Halftime committee (Jun 29) on the Mag-7 'lag-7' pullback: Lebenthal — it's margin pressure from chip-input inflation (good for chipmakers) and over-high expectations, not a fatal flaw, so the dip is 'a setup for the next leg higher' (AMZN, GOOGL). Link — the multiples are reasonable vs cloud growth (MSFT 22x/Azure +38-40%, AMZN 24x/AWS +28%, GOOGL 24x/Cloud +63%), but the real worry is free-cash-flow burn (some go negative FCF this year) and no operating leverage in the capex cycle; Meta is the only one 'with complications.' Terranova — watch the momentum factor: 'the first crack in the bull trend will come from the momentum factor itself.' Harrington — a WSJ 'earnings forecasts on steroids' piece flags Mag-7 EPS flattered by private OpenAI/Anthropic stakes — 'a little pause on future valuations.' Halftime committee (Jul 9): Brown's answer to "the only question that matters" — where the CapEx value finally accrues — is the consumer gateway: Apple's 2.5B devices become the distribution layer of consumer AI ("bring your own LLM" — layer your existing ChatGPT or Claude onto agentic Siri; Apple gets paid through the App Store either way) while sitting out the CapEx race, the way Instagram (7 employees, $1B to Meta) captured the internet AT&T and Verizon built; Alphabet gets a "double benefit" as spender and Gemini owner, while how much accrues to Microsoft's buildout "I'm not sure the market is as confident." Talkington inverts it to losers: consumers don't want another device (no OpenAI "walnut-shaped box"), and the enterprise is murkier — she owns Microsoft but "the money they're spending, I just don't see the output" (Copilot mediocre vs the other LLMs; Teams so weak Zoom "should not exist as a publicly traded company"), with Oracle execution-risked via its OpenAI tie. Eisman (Jul 10): the market IS the AI trade — info tech is 38% of the S&P 500 (50%+ with Google/Amazon/tech-related), and per Apollo's Torsten Slok AI capex contributes ~100bps of 2026's ~2% GDP growth (reshoring +30bps, tax refunds ~90bps), so "AI better succeed and succeed big — if it fails or takes a lot longer to become profitable, the US is going into a recession and the market is going straight down"; his subprime lesson applies — arguments evolve as facts emerge, and the cracks arrived sequentially (Oracle's OpenAI-heavy backlog, the 2026 capex guides, then the capital raises), with even semis now wobbling: Samsung printed +1,800% operating profit and fell 7% on hyperscaler-slowdown fear while semi pricing still rises. Polomny (Jul 10): 7th/8th inning — Mag-7 forced capex is going FCF-negative under "creative accounting" and when it blows up "the NASDAQ could go down 80% like it did then" (1999/2000); position to benefit from many possible futures, not the one consensus future. Hickey (Thoughtful Money, Jul 14): the greatest stock and earnings bubble in US history (agreeing with Grantham on the bubble, unsure on the 70% crash) — Buffett indicator 241% of GDP vs 160% in 2000, price/sales 3.7× vs 2.3×, and the reported CAPE of 40 is really ~67 adjusted for the AI earnings distortion: suppliers (NVDA/MU/WDC/STX) book data-center orders as instant revenue while the buyers depreciate the same spend over 5–6 years, and one-time OpenAI/Anthropic/SpaceX stake markups were 40% of Q1 earnings growth — strip both and S&P Q1 growth falls from ~28% to single digits. The pricing is being gutted underneath: Chinese open-source models (DeepSeek/Zhipu/Qwen) hit ~46% token share (vs an 11% prior-12-month average) doing ~90% of tasks at ~1.5% of the cost, Meta slashed prices 75%, OpenAI/Anthropic discount, and the token pullback is broad (Microsoft canceling most Claude Code licenses and weighing DeepSeek, Tesla's $200/week caps, Coinbase halving AI spend, Amazon's $500M one-month bill; Chamath's CTO: costs doubling every 45 days for ~5% return) — "the economic modeling, if there was any to begin with, is disintegrating"; it was billionaire FOMO, never a model, and only ~5% of planned capacity is even turned on. Expects a 2000-style bust where everything gets smashed (MSFT fell 60% then as the BEST performer) before the survivors — MSFT, GOOGL, META with real fallback cash flows — become the buys; ORCL the most stretched (100% of revenue into data centers, debt 2.5× sales, −60% off highs) and some AI names go bankrupt, with OpenAI/Anthropic ~50% of hyperscaler backlogs the reflexive risk. Timing unknowable (nobody rang a bell on March 10, 2000) — watch exhaustion signs; sitting on record cash waiting for the fat pitch. Halftime (Jul 15): the unwind broadens even as mega-caps bounce — capital rotates out of memory/semis into the trillion-dollar names (AAPL/MSFT/META/AMZN/GOOGL ~+3% on the day; Buffett reveals Berkshire's new Alphabet stake, calling not buying Google "a mistake"), momentum −4% unlevered (Santoli: the AI-hardware bar is "incredibly high" — Micron, TSM, then ASML all sold off good prints; ASML's +15% guidance raise wasn't enough). The memory fever "broke" — SK Hynix's US listing −12% on "too much supply," and Weiss exits Micron: "closer to the end of the AI trade, not AI fundamentals" (Tudor Jones's another-year-or-two is now front-run). IBM's worst-ever day vs CrowdStrike's record high = the software haves-and-have-nots as finite IT budgets redirect to AI and cyber. The other side: Baruch holds MU through it (largest position), and Sechan says semis are "innocent until proven guilty" — price led fundamentals by a year in 2000, so buy the dips with the hyperscalers as the hedge. Smead back-fill: Nov-26-25 (Squawk Box) — the AI-capex mania is the late-90s telecom capex boom rerun: even the SURVIVORS of a capex mania compound ~5% for decades (AT&T/Verizon since 2000); "you don't want to own the people that do capex" — Google won in 2002 precisely because it didn't have to build the internet; the Altman-Nadella interview read as the scale-tipping moment. Feb-05-26 (with Bill Smead) — hyperscalers are "this generation's Nortel": Meta's incremental ~$45B of capital earned ~3% on net income (negative on FCF), Microsoft the same — asset-light businesses turning into the most capital-intensive in the world means returns on capital fall; the top-10 market-cap list turns over almost completely every decade. Pabrai (jun-30, rec. jun-8): AI is "very deeply loved" — so avoid it; risk-free bargains only live in the hated-and-unloved. The crowd rotates shiny objects: Bitcoin (100k→70k) was dumped to buy AI, and AI will next be dumped for the SpaceX IPO — "deeply loved" is a sell signal, not a buy list. Finucane (Jul 16/17): Oxbow's CIO calls the reversal "inevitable" and is playing defense — AI beneficiaries are only ~5% of his growth book while ~45% of S&P market cap (and ~70% of the Nasdaq-100) sits in one AI trade; after a bubble peaks these groups historically take ~2.5 years to bottom (Cisco/Nvidia post-2000), and 2027 brings hostile base effects (cycling H1-2026's high growth/inflation prints makes growth look decelerating — poor for risk assets). Singh (SSR, Jul 19): Dan Niles frames the AI pullback as a speed bump (token minimization, cheaper Chinese/open-source models), not the bust; the open-source-AI case is that a ~32% reinvestment rate is hyperscaler breakeven and Jevons's paradox means cheaper models drive MORE compute/memory demand (hyperscaler G-spreads above 150 for the first time the borrowing tell). ~70% of 2027 data-center builds delayed pushes the cycle into 2028; AI names "tradable, not buy-and-hold." Hay (Jul 26): earnings quality is the crack — of Alphabet's very good quarter, "almost two-thirds… have come from the gains on their venture capital investments in these entities which are big customers," booked in other income, properly but unrepeatably: "it is a circular financing type of situation… they are very unsustainable." Add the depreciation on hundreds of billions of capex that "has really yet to hit the P&L" and (per Fred Hickey) a teens/20s P/E becomes "like 65." The tape is the tell: GOOGL fell on good earnings because it raised capex — "what happened in the late 90s… companies like Cisco reporting great earnings and the stocks went down. So maybe this is the big shakeout." Same skepticism inverts for cheap-looking semis (MU: single-digit P/E but extreme price-to-sales, margins from scarcity not innovation, overordering → "air pockets") and cyclicals swept into the mania (CAT ~32× PE / 6× sales — "take some partial profits"). A Meta capex cut "might [send] that stock straight up." Singh (Jul 26): Max7 multiples have compressed 32× (Oct-2025) → 22× while reported earnings "soared" on one-time gains and capex-depreciation assumptions; the basket fell 4.8% in a week (~$750B of market cap) and token costs have erased the entire agentic-AI bounce. The trade instruction: wait for sell-the-news on the AMZN/MSFT/META capex guides, then buy the recipients — "when the market realizes all the Max7 guys are raising their capex spending, you'll see another bid for the semiconductor names." Hyperscalers are ~16% of the S&P vs semis ~19%. Eisman (Jul 27): acted on it — sold his long-held Google "a couple of months ago… because I felt I wanted to reduce my exposure to AI" and is sitting in cash, with no replacement ("buying something defensive is kind of a waste of time… they don't go up… people either want to buy AI or they don't, but they don't want to shift out of it to buy Clorox") and no short ("I would not. I've lightened up"). The reason: "it's all one trade. It's literally one" — even 60/40 holders "are actually not diversified": more than 50% of the 60 is tech/AI-related and "most of the new issuance of bonds is AI related." So the outcome is binary — "it better succeed"; if it doesn't, "I think we have a big correction." He does not expect the pop from disappointing capability ("it's going to be something really good — that doesn't mean everybody succeeds") and won't redeploy soon ("I don't think this debate is going to get settled within the next two weeks"). The cost-side bottleneck: Anthropic and OpenAI drive much of the spending and their closed models are "now much more expensive than the competition… a potential bottleneck." Eisman (Jul 31): the debate is now cut in two — "the amount of money it takes to be a hyperscaler is insane, and that expenditure itself is a moat… there are only going to be a few," so GOOGL/AMZN/MSFT/ORCL "have real businesses here," while the LLM providers have "no moats or at best… shallow" ones, with enterprises switching models and using cheaper open-source Chinese alternatives. The named selloff catalyst: "a key thing to monitor… is the health of Anthropic and OpenAI" — so much of the hyperscaler backlog is these two (roughly half of Oracle's $600B+ is OpenAI) that trouble there corrects the whole ecosystem. Damage so far: NASDAQ −7% from its Jan-2 high. 2026-JUL-31 — Dan Niles (panel clip): applies a base-rate ceiling before any company analysis — "you only have a certain number of winners… who's the winner in search? That's just Google. What about in e-commerce? That's just Amazon. Are you going to have five different guys win in AI? No." His count is two, split by customer type: Google in consumer ("they have the complete stack. I think they win in AI overall") and Anthropic in corporate ("they got to profitability in Q2 and their revenues are ramping like nothing we've ever seen in history for a company of that size") — leaving OpenAI "stuck between the two of them… jammed between the other two guys." Asked whether OpenAI and Anthropic interest him as public-market stocks, he takes only Anthropic. 2026-AUG-03 (Joseph Carlson): a one-week sentiment reversal in the hyperscalers, attributed by sequence — "it wasn't after Microsoft reported that Google and Meta went up. It was after Amazon reported." AMZN +5% on the day, +23% in five sessions, through a $3T cap; MSFT +16% in a single day at a $3T+ cap, erasing all 2026 losses; GOOGL +3% and META +5.5% purely on read-across after both had been sold on their own good reports. "People are tripping over themselves to buy shares today." Carlson grades his own Sept-2025 AWS-reacceleration call (~25%, stock to 240–260) against a 36.7% print and a $285 stock — "I underestimated it to a huge extent." He declines Tom Lee's top-down bull case (S&P 7,800 this month, 2027 "one of the best years for the stock market" as the SpaceX unlock, new-Fed testing and the leverage unwind clear) — "more just looking at the bright side of it" — grounding his own 2027 optimism in hyperscaler demand and monetization instead. App Economy (Aug 4): the beat-and-fall regime persisted — AMD beat and guided Q3 ~$0.5B above consensus and fell ~9% after doubling YTD; SpaceX beat and fell ~6%. "The numbers were strong, but expectations have moved even faster" — the move is positioning, not the print. Palantir is the counter-case: a 12th consecutive quarter of acceleration (+93%), Rule of 40 at 155, 63% FCF margins, record $2.13B TCV / $6.24B RDV — carrying ~80x FY26 EBITDA. 2026-AUG-10 — Paulo Macro & Le Shrub (Fly on the Wall): "claudification" — retail and pods asking the same AI models the same framing question ("find me the bottleneck in photonics… power… memory"), getting the same names back, "make new baskets and just buy the same crap": "they think they're having alpha, but actually they're just doing momentum investing in a different word." First casualty: Leopold Aschenbrenner's fund at 4× leverage — copied wholesale, "they all blow up together" (their earlier call: the blow-up was a "situational-awareness mean bottom — you got to buy the AI trade"). Institutionalising, not peaking: IBKR adding a ChatGPT portfolio function, Millennium working with Anthropic — "early innings or maybe mid… there's no free will out of all these guys." The passive/mechanical bid and "the active slash retail monkey" now herd identically. Singh (Aug 9): an Nvidia-employee note says Meta has ~400,000 excess GPUs it must place on the rental market — "they have no other way but to enter the game" — the specific mechanism that would depress GPU lease rates for NBIS and CRWV, and the one condition under which Morgan Stanley's "Neoclouds win in every scenario" framework fails. Separately, LLM token costs are back to start-2026 levels: Alibaba's Qwen 3.8 Max at ~$2/M input and $6/M output, "a tenth of what Claude costs." Carlson (Aug 10): takes the other side of SemiAnalysis' "Gemini is cooked" — the same shop called Gemini state-of-the-art in November, so putting Google's odds of a repeat at literally "zero" (their word, not "practically zero") is indefensible. But he accepts the report's second half: Google is fine even at second tier because it is deliberately putting more money into Google Cloud than into training models — cloud that "could grow above 100%" with operating margins ticking up every quarter, "now up to 35%." Given the choice, "a lot better of an opportunity to build up Google Cloud" than to own the best model. 2026-AUG-04 — Dan Niles (CNBC): open source "cuts your cost to produce a lot of these tokens by 90%" and most workloads never needed a frontier model, so the clouds will "route whatever you're trying to do to the best model available for the task" — "the model layer becomes more of a commodity… Anthropic and OpenAI become more commoditized over time, and the value accrues to more of the infrastructure players, which includes the cloud platforms and semiconductors." Positioned that way "100%." AUG-16 (Jay Singh, SSR): the post-liquidation bounce is exhausted, and the tell is beat-and-raise results selling off. Measured from the July 29 Situational Awareness forced sale: CSCO −8% on the week (+45% YTD) after an 8% bounce, AMAT −6% (+97% YTD) after 24%, COHR −14% (+77% YTD) after 71% — "negative stock reactions to the headline beat and raise earnings on both revenues and EPS… the biggest problem was arguably their recent bounce." Conclusion: "the easy money on the AI technical rebound from oversold levels is probably over" — add hedges on further gains and get more selective. 2026-08-17 — Strategas is bearish on the funding and still market-weight the trade. "One thing that is not arguable is that the business has become incredibly capital intensive"; the published line "it's going to be hard to continue to grow AI investments without further dilution of shareholders" got Trennert fired by a client — the episode's sharpest crowding datapoint. Yet: "I wouldn't short it. We're market-weight technology, and I do think the demand is exceeding the supply," with a preference for "the semiconductors" and the caveat the hyperscaler layer "could be a little bit of a race to the bottom." Verrone's sentiment read: negativity "reached a fever pitch… when Oracle got to like 120," and since then "Google's basically back at the highs. Amazon just made new highs. The 5-day change in Microsoft is the best we've seen since the IPO in '87" — "at a minimum I think they're tradeable." What people are struggling with is method, not fundamentals: these were "cash flow monsters" and investors are "searching for the right way to analyze them today cuz it's different… mistaken for some deterioration in AI fundamentals." (Trennert/Verrone on Eisman Ep 73, Aug 17) 2026-08-18 (App Economy Insights, Q2 13F round-up): the trade broadened beyond NVIDIA. Across a fixed sample of 20 alpha-screened funds, AMZN and TSM tied as most-held (top-five at 9 of 20), GOOGL at 8, NVDA at 5 — but NVDA was a top-five buy for only one fund, while TSM and Cerebras (CBRS) were top buys at four each and AMAT/MU/AMD at three each, with INTC, ALAB, QCOM, NBIS and CRWV also appearing. Core exposure kept, marginal dollar moved down the chain. Absences carried information too: MSFT fell to a single top-five appearance after −20% YTD, and AAPL and TSLA were absent from every fund's top five. Halftime panel (Aug 21): Brown's framing into NVDA earnings week — "Jensen Huang is the AI Fed chair… he is to the earnings growth of the S&P what Kevin Warsh is to interest rates" — with the read-through set re-rating off 2027–28 guidance (Dell $300B, Arista $235B, GEV $270B, AVGO $1.7T) and NVDA's expected $92B quarter = +97% y/y sales at a near-market multiple after this year's multiple contraction (why confirmation suffices). Raskin's unmet ask: "the productivity story is still to come… what are the killer apps besides coding?" — and OpenAI's growth "a little disappointing." (CNBC Halftime panel, 2026-aug-21.) 2026-AUG-26/28: App Economy: Jensen Huang — “when the world goes to agentic, fully agentic systems, you are going to have agents running all the time, working with other agents” — inference demand decoupling from human usage. The cost side of the same mechanism: CrowdStrike's AIDR ending ARR nearly tripled Q/Q because “every agent effectively becomes another identity and another attack surface,” a category management believes could eventually exceed EDR. (App Economy Insights 2026-AUG-26/28) 2026-AUG-30 (Jay Singh, SSR call): the largest print in market history, immediately given back. Nvidia delivered $96.2B revenue (+106%), $89B data centre, ~$60B net income, commitments $119B → $279B, 75% gross margin, Q3 guided $108B, and a CFO guide of ~70% growth through FY28 with Vera Rubin in full production and cloud backlog above $2T — adding a record $453B of market cap in a single session. It then fell on Warsh, on the chip-tariff proposal, and on self-financing scrutiny. Ecosystem: AWS committing 2 million more GPUs for 2027-28, the reported $12.9B Hugging Face acquisition, SpaceX's Vera Rubin NVL72 in orbit via the Starmind satellite, and OpenAI's Broadcom-built Jalapeño/Palomino inference chip beating the GB300 on throughput-per-watt and latency at 700W (benchmarked against GB300, not Vera Rubin). The frontier-lab economics: Anthropic and OpenAI's combined run rate went $29B → $105B in seven months, and Anthropic is marketing an IPO at a $2T valuation raising up to $100B on a $30T TAM — "which is bananas because US GDP is only $32 trillion." Valuation is not the constraint: Raymond James notes Nvidia trades below 15× CY27 earnings against the S&P at 18.6×. Steve Eisman (New Money clip, 2026-AUG-15): the diversification audit, stated as arithmetic — "even people who think they're diversified because they own 60% stocks and 40% bonds are missing the fact that they're actually not diversified because of their 60% more than 50% of it is tech and AI related. And of the 40% of bonds, most of the new issuance of bonds is AI related." The conclusion is about correlation rather than price: the classic hedge leg no longer diversifies, so "what scares me is that it's all one trade. So it better succeed." (2026-AUG-28, The David Lin Report): the same test applied to a professional book, via the fund that blew up in July — imagine 1905, buy every auto and auto-parts company and short every horse-carriage company: "what you've done is one trade… your longs and your shorts are not independent of one another. They move exactly in unison." A press-driven scare reverses both legs together, and "if you're levered 4:1, you get put out of business really quickly." Directionally right and fatally positioned are not mutually exclusive. Macro restatement: "as long as AI CapEx keeps powering higher, the economy and the market probably goes higher. It's all one trade." 2026-AUG-31 — CNBC Halftime committee (off the WSJ's Zuckerman/Banerji piece): the momentum index is down >9% since July 1 vs the S&P's +2.8% — the biggest quarterly momentum underperformance in 25 years; July was the second-worst momentum month in ~40 years (BofA), beaten only by April 2009. NVIDIA's blowout guidance failed to lift the group (Marvell either) → the desk treats it as a mechanical unwind, with capital relocating from semis into software (IGV near its highs). Talkington's flag: 60% of NVDA's last-quarter revenue sat in accounts receivable. Counterweight (Terranova): the print may re-concentrate the market rather than broaden it. 2026-AUG-20 — AI must destroy employment to be worth it, which halves the tax base. "I think AI is going to be revolutionary… but before it does, I think it is going to undermine the tax base of the United States that already has a fiscal problem, because half of the tax base comes from jobs." The specific exposure: healthcare administration is the largest employer in 39 US states, "uniquely suited to being disintermediated massively by AI. And all those people have mortgages, cars, student and consumer loans that they will then default on." The bind is symmetric — if AI doesn't create unemployment it isn't productive enough to justify the trillion in AI debt, the trillion in lease commitments and the trillion in semiconductor commitments, "so that hits receipts. A snake eating its own tail." Positioning that follows: take profits, don't short ("I wouldn't short them here") — the bubble runs while new credit arrives (SEC loosening AI securitization rules, PE syndication, a compute derivative), the pattern he took from reading about the South Sea and Mississippi bubbles. Gromen (Monetary Matters, 2026-AUG-20) Pernas (Monetary Matters, 2026-AUG-07): recorded into a 40–50% one-month drawdown across the memory/AI complex (SanDisk, SK Hynix) with the market "really questioning if ROI is there anymore." Their counter-framing: model AI as four layers — chips → infrastructure (clouds) → frontier models → applications — and apply "commoditize your complement". Open-weight models (Moonshot's Kimi, which they believe is substantially distilled from frontier models rather than pure architectural innovation) are bearish for frontier labs only; every other layer gains, because cheaper tokens mean more enterprise AI, more cloud consumption and higher GPU rental prices — with Jevons paradox keeping total spend rising. Enterprises had been blowing budgets "in a span of a month" at ~$10 per million input tokens (Uber and Amazon publicly pulling back), so the frontier demand air pocket is simply re-hosted on the hyperscalers. Index vulnerability: the top 10 names are ~30–40% of the S&P, are "existentially forced" into a step-function capex with uncertain ROI, and average roughly 10× EV/sales while spending all their free cash flow and more — "a lot is going to have to go right for that multiple to be roughly correct." Conclusion: "the leaders of the index are being forced to kind of subsidize this spend for the rest of the economy… a very special time to be an active manager… a pretty exciting time to be a stock picker." They hold no semiconductors at all, citing China — already dominant in EVs and drones, "going after ASML" and the memory makers, with CXMT's Shanghai STAR IPO at roughly a half-trillion market cap. (Pernas, Monetary Matters, 2026-AUG-07.) Wigglesworth (Monetary Matters, 2026-AUG-16) — pricing power is a wasting asset. Asked what a millennium of financial history says about the pricing power of a new industry: "it tends to erode." What is priced into NVDA is the assumption "that nobody else can create GPUs at scale and quality of an Nvidia ever," while Nvidia itself depends on a supply chain it does not own — "monopolist-like pricing power tends to not last very long. Sometimes it can last for a few years. But it never lasts forever as far as I know." He discounts the ~$500bn Nvidia/alt-manager financing MOU as marketing until the paper exists ("take press releases with a pinch, maybe a fistful of salt"), and separates the majors from the outliers: not worried about META/GOOGL/AMZN going bust because a real product survives ("people are still going to be going on YouTube even if Google wastes a few hundred billion"), but ORCL is "clearly the weakest of the litter" and CRWV is "the real king of debt… I've never seen a bigger gap between EBITDA and net income loss," with no fallback product at all. On Fink's "AI securities": plausibly compute futures eventually, "but just because you say something is an asset class doesn't make it so." (Wigglesworth, Monetary Matters, 2026-AUG-16.) 2026-AUG-31 — Jeff Keller (Capeite Partners, Other People's Money): July 2026 was a deleveraging event, not the blowoff top — Situational Awareness LP's margin calls plus a Korean retail/hedge-fund deleveraging, not a fundamental break. His three health gauges all read strong at end-August: lab ARR (adoption still climbing the S-curve, "no crack in there yet"), capex (locked strong for '27, '28 in view; levelling off at a high level still works), and the forward price of compute plus data-center build times (very tight, "not suggesting we're at a glut yet" — though Elon entering neoclouds could hasten one). He likes short selling and still won't short AI — "shorting an open-ended growth story can be deadly because there's no catalyst." Expects chop, not new highs, until the next breakthrough (recursive self-improvement "floating out there"); prefers the low-multiple names (NVDA, MU — "telling you we're at peak") over the extrapolating behind-the-meter/services winners, since the whole complex trades on the same two drivers. 2026-SEP-01: Trader Ferg flags Kedrosky's overdetermined-failure framing as "the best articulation of the danger of high multiples/valuations I've heard": "at high valuations, failure is overdetermined in a statistical sense… so many ways to fail, all of which are low likelihood—maybe 5% at most—that when you combine them all… given 20 different ways this could fail… there's greater than a 60% chance of failure. So what looks not particularly risky or unpredictable is actually highly predictable… when your PE is 70… the market increasingly looks at it and says, 'How can this thing continue to outperform expectations?' And the answer is, whoops, it can't." The multiple is treated as a forecast rather than a risk to size around. (Ferg's Finds, 2026-SEP-01.) Slegers 2026-AUG-02: first concrete crack in the leadership — “While the S&P 500 was flat, the iShares Semiconductor ETF declined by 20% in mid July,” alongside an unusually wide VIX/VIXEQ dispersion gap. His reading: “This could be a sign that the AI and semiconductor trade is cooling off. In the meantime, investors rotate into other parts of the market.” Separately (2026-JUL-19) he treats AI-disruption fear as the source of the month’s opportunities rather than a thesis — Gartner -27.8% and EPAM -27.0% lead the losers, and the Paychex spotlight is explicitly built on the market being “very fearful that AI will disrupt” a business whose customers “almost never leave.” Slegers 2026-MAY-07: declines to pick a winner and says so — "as Buffett would say, predicting the winners of the AI race is in our 'too hard pile'" — and substitutes a pick-and-shovel strategy: during the 1848-55 gold rush "most miners never found gold… who did make money reliably? The guys selling picks, shovels, pans, and boots." Three layers named as the modern equivalent: cloud compute (AWS, Azure, Google Cloud), energy and power grids, and "cooling systems, fiber optic cables, data storage — boring but essential." Slegers 2026-MAY-14: runs the full 15-step quality template on Arista Networks as the pick-and-shovel AI name and still declines it — 14 of 15 checks pass (63.5% gross margin, 41.0% ROIC, net cash, CAPEX 1.1% of revenue, FCF 46.7%/yr over a decade, 36.7% CAGR since IPO) but the price fails: 37.7x forward against a 32.1x ten-year average, with the reverse DCF requiring 18.9% annual FCF growth for ten years just to deliver 10%. Verdict: "a great business but we don't like the valuation level … happy to stay patient." The competitive claim worth tracking is a standards fight: Arista connects Nvidia GPUs over Ethernet while Nvidia pushes its own more expensive InfiniBand, and 35% of Arista's revenue comes from Microsoft and Meta alone. Pieter Slegers (Compounding Quality), 2026-JAN-04 supplies the 2025 scoreboard from outside the mega-caps, and the pattern in it: of the ten best-performing reader picks, the two AI-infrastructure names re-rated on a single hyperscaler contract each — Nebius +174.4% on "a multi-year AI infrastructure deal worth over $17 billion" from Microsoft (Meta following), and IREN +261.1% on "about $9.7 billion", also Microsoft. Slegers notices it in print — "do you start to see a correlation between AI and the best performers of the year?" — and draws no conclusion. The list's other members (Olema +348.8%, ASTS +235.6%, MU +226.8%, KTOS +187.8%, RKLB +179.5%, UUUU +156.0%, EOSE +107.2%) are disowned in the same issue: "these are not the type of companies we usually cover… most businesses on this list are cyclical by nature." The unmade structural point: two of the year's biggest winners are single-customer concentration stories described as pick-and-shovel plays. 2026-AUG-26 (Dowd/Phinance, WTFinance): the June rally to new highs "was a very narrow rally… all AI or AI adjacent" and peaked with the semiconductor index. "I'm not calling a top, but I'm watching the structure" — if semis go to new lows, "the bubble may have already burst, but people won't realize that until a couple months from now." The capex cycle stops one of two ways: credit markets getting too expensive, or simply not enough grid power for the hyped data centres. When it slows, the pick-and-shovel beneficiaries roll over and "it becomes a reflexive problem." 2026-SEP-02 (CNBC Halftime — Terranova): the momentum factor broke below the situational-awareness lows. Mizuho: the momentum baskets "have now completely round tripped and broken below their July lows… increasingly self fulfilling as systematic strategies de-risk, forcing additional selling into weakness"; Goldman's TMT-and-momentum pair had its worst month on record in a series dating to April 2011. Terranova keeps the bottom call but detaches it from the entry — "I think it's washed out… but the other side of that is that I don't think you step in here until that factor proves itself, because in the wake of situational awareness you had that false start… you have to have a degree of apathy towards it." Broadcom was the live test and he refused it: 21× forward vs 32× in June and an 18× 10-year average, "sometimes there's no trade… but if they deliver 200% earnings growth and 85% revenue growth and that's not enough, then that AI momentum trade is going to continue to sit silent." (2026-SEP-03, Dan Niles, Excess Returns) He splits the trade from the thesis — "between now and the midterms on November 3rd, I'm pretty negative," then "at least another year for stocks to go higher." Being early is expensive: the NASDAQ rose 86% in 1999 and another 24% into 2000, so "you can know we're in a bubble but still have a lot of opportunity to make money before that bubble breaks." Dillian (2026-SEP-03): “Do you think that AI is a bubble? I do. I do. I’m not bashful to say that” — sourced to the balance sheet rather than the multiple. On sentiment, “we had four magazine covers on AI in the last week,” one from Barron’s along the lines of the bubble never ending, which he reads as “a little brutal on the sentiment side” and as classic both-sidesing (“you’re calling it a bubble but you’re still bullish … make a call”). He also reads an AI podcast forecasting “double digits or maybe even triple digits” bond yields as “a pretty strong sentiment indicator” — domain experts making market calls. Crucially he refuses to convert any of it into timing: “there’s nothing in the S&P chart that leads me to believe we’re going to crash tomorrow, next week, next month. The chart is pretty neutral.” The one trigger he is waiting on is the second derivative of growth — Nvidia’s guided 70% stepping down to 60% and then 50%: “that’s when the stocks are going to top.” CNBC Halftime — Steve Weiss (2026-SEP-04): the trade did not come back, it got selective — “people believe in the AI trade, but now they’re being more selective. So they’re going to the companies that have shown that they can generate revenue and ROI.” What is being paid for is safety, “the known factor,” cheap debt and defensible valuations; what is being discounted is the build-out layer — “Cat goes here and there, but still it’s well off its highs. GE Vernova, well off its highs… much more volatile than these names.” 2026-SEP-02 (John Feneck): expects a 15–20% drawdown in the S&P by Q1 2027 and a sector rotation “between now and March — that's my call” — out of tech, AI and momentum and into value, the examples he names being Newmont and Barrick. What breaks it: “the AI tech trade for sure”; the March-2009-to-now rally is “long in the tooth.” 2026-SEP-07 (Excess Returns): two dissents worth logging against the crowded-downside read. Niles dates a bull runway of "at least another year" off the agentic phase, and Forehand names the two things genuinely unlike 1999 — "there is massive demand for it now, very different in the late '90s where we were building in advance of demand," plus physical "governors that are limiting it." The discipline he attaches: "concede the idea that there might be an intelligent person on the other side of this and they might be right." Zeigler's counterweight is the Bob Farrell read — the contrarian itch is everyone "saying the same version of the same thing" across S-1s, IPOs and earnings calls at once. 2026-SEP-10 — Tardif is positioned, and names the instruments. Since 2021 he has cut single-name shorts and hedges with puts instead: "mostly S&P 500 but sometimes others like right now we have puts also in Nvidia," bought into rallies and sold into declines, taking a net-long equity book "well below zero on a net basis." The carry argument is why it is permanent rather than a timing bet — "if we're wrong on the options it's a very small cost every month. But if we're right, and faster the market goes down, better it is for options." He also flags that the trade is now one factor: on a Monday when ~75% of stocks fell with semis, "TransAlta was up, Capital Power was up, and copper stocks, many of them up 10%." 2026-SEP-09 (Steve Eisman / Ed Zitron): Eisman platforms the overheated-AI-trade case rather than rebutting it — the objection is revenue quality, not valuation: no multiple or target is argued, only that ~70% of hyperscaler AI revenue traces back to two loss-making labs the hyperscalers themselves fund. Jay Singh (SEP-13) on the Amodei "pace the frontier" letter and Altman's "ill-advised moment to go public": "sentiment risk, not fundamental risk" — no hyperscaler guide has moved — reclassifiable only on a delayed model, pushed compute deployment, lower accelerator demand or an AI-spend guide cut. The real risk is funding: if Anthropic/OpenAI can't raise, "how are they going to support Microsoft, Amazon, and Google's cloud backlogs?… who's going to buy Nvidia GPUs"; exited SKM (Anthropic proxy) on it. David Woo (2026-SEP-14): "the AI trade is the US economy" (US ~60% of world market cap, Mag 7 ~30%, tech ~40%); without it the market would already have collapsed on oil and rates. Washington will do anything to protect it — cap long rates, pre-empt state data-center laws, ban Chinese models — but "cannot have his cake and eat it too." Short via QQQ put spreads and OTM index puts expiring before the midterms; traded in and out. Dan Niles (2026-SEP-15): cash is now his top pick #1 — "between now and midterms, I'm very negative on the market overall, which means I don't want to be involved in anything." Semis and "the CoreWeaves of the world": "you should have been pulling back a few weeks ago." The semiconductor index fell 29% from end-June to the Situational Awareness bottom in about a month, which he calls a warning shot because today's triggers (data-center pushouts, labs slowing training) are worse. 2026-SEP-14/15 (CNBC Halftime): the Amodei essay selloff (SOX -5.3%) split the trade into spenders vs vendors — Santoli: "maybe the spenders are going to be OK, even if the vendors have to do with less"; MSFT/GOOGL green, Weiss calls Broadcom a falling knife. A day later Brown's selection rule: "there is no AI without the appropriate amount of cybersecurity" (CRWD +104% YTD, FTNT/PANW doubled); Link cites accelerating Azure/Google Cloud/AWS as "real confirmation" of ROI (not Meta). Gundlach (2026-SEP-16): out of all AI equity exposure "starting last week" — losers in "the AI race for the holy grail" will lead the next significant drawdown; model portfolio now holds zero AI (equities 30% in equal-weight Fortune 500 DFVE). Deluard (2026-SEP-17): the AI narrative is "starting to weaken" ~4 years after ChatGPT — Nvidia's growth rests on hyperscaler capex (>$1T across five companies) that can't keep doubling ("you run out of economy to eat"); data-center backlash, pulled/delayed IPOs and bailout talk (Oracle, OpenAI) say "we're closer to the end"; the ~40 AI-adjacent stocks are ~50% of US market cap and bonds are "a derivative of the AI trade" — insure the fault line; too big to fail is not too big to lose. Jay Singh (2026-SEP-20): the AI pacing debate, read three ways. Amodei, Altman and Nadella back 'deliberate pacing' and an Anthropic/OpenAI/Google standards body. Michael Burry calls it 'self-serving': LLMs aren't AGI, Chinese competition erodes incumbents, 'IPOs need hype and puffery,' and it could be 'a cover for a real uncontrollable slowing' as the IPOs slip to late-26/27. Trump calls it a hoax. Singh: 'the AI companies are self-serving, but… there needs to be… cybersecurity oversight.' Goldman PB had hedge funds buying TMT in 10 of 11 sessions, at the 97th-percentile pace, 'right into this slowdown.' Meanwhile Nvidia raised GPU prices 20% from Oct-1 and semis printed a record $425bn quarter. 37% of BofA FMS respondents see AI over-investment, and Goldman says fully expensed capex would cut 2027 EPS about 22%. 2026-SEP-18 (Eisman, mailbag): diversifying away from AI is hard because utilities (power), industrials (build-out) and large banks, investment banks and alternative managers (financing) are all AI-correlated; only staples and healthcare - "only 14% of the S&P 500" - are not. His fix: sub-sector and low-volatility ETFs (LVHD, SPLV, KBWP for P&C), and short-against-the-box to hedge large embedded tech gains without triggering tax. Chris Whalen (2026-SEP-19): the AI boom that carried the market for two years is "kind of petering out". Corporate issuance is slowing and banks are warning on capital-markets revenue.