AI build-out — shortage vs bubble Contested — capex-ROI bears vs contracted-demand (AWS backlog) bulls

Baker (Sohn, May 15): AI's binding constraints are physical — watts and wafers. TSMC's flinty capacity discipline (guarding Morris Chang's legacy) rations wafer supply (~5%/yr vs the "double or triple" Jensen wants), a real-world brake past manias lacked — so the build-out may avoid a classic bubble ("smoother for longer"). Memory may be in its first true capacity cycle since the mid-90s (the one cycle you don't sell). The watt shortage gets solved by orbital / space-based compute toward ~2030; the painful late-decade short is terrestrial power/cooling industrials that over-built. Custom silicon: Amazon's Trainium > Google's TPU; quality neoclouds (CoreWeave / Crusoe / Nebius) earn a durable GPU-utilization premium. Rule (Jun 21): reframes AI as a physical, not digital, problem — there isn't the power, water or copper to build all slated data centers, so the most-bullish AI case can't happen on schedule (it arrives "20 or 30 years late" or gets rationed by price). If the resources don't show up as analysts assume, the AI complex's earnings come down — and since AI is "the tail that wags the dog," the broad market with it. Singh (Jun 21): the physical-shortage framing is now confirmed in the data-center pipeline — Jefferies/JPMorgan find ~50% of 2026 and ~80% of 2027 announced builds haven't broken ground (SemiAnalysis pushes back to "under 50%" but agrees delays are large); the binding constraint is physics not capital (transformer lead times 12–18→30–36 mo, grid-interconnect queues 3–5 yrs, capacity-clearing prices +1,037% to $329/MW-day), so whoever already has live capacity and secured power sits in a sellers' market with premium lease rates (CRWV, NBIS, DGXX). App Economy (Jun 30, Cerebras): the binding constraint is now physical real estate — Feldman's first-earnings mantra "demand is not the constraint, supply is not the constraint, the constraint is data centers" reframes an AI-chip maker as an infrastructure-buildout company (capex $132M/qtr, FCF −$120M, ~$2.3B of off-balance-sheet future data-center leases), racing to add US / Europe / Bell-Canada capacity while renting chips back from a customer. The bull note: Cerebras shows genuine non-Nvidia AI demand — its wafer-scale engine sidesteps the HBM / CoWoS / 3nm bottlenecks everyone else fights over (plentiful SRAM + mature 5nm) and clears ~1,000 tok/s (~10× a GPU cluster), with speed as the moat. App Economy (Jul 29/31): demand still exceeds capacity through 2027 on both sides of the duopoly — Jassy: "[at $220B] we will still not have enough capacity to meet all the demand we have in 2026… I believe this dynamic will also be true in 2027"; AWS +37% to $42.2B (fastest in 18 quarters) at a 39% margin with a $496B backlog; Microsoft's Azure +43% above its own 39–40% guide, guided to ~45% next quarter with H1 accelerating, commercial RPO +84% to $678B (+25% even excluding OpenAI). Total cloud-infrastructure spend +43% to $143B — an 11th straight accelerating quarter — with GenAI cloud services +165%. Share: AWS 28% · Azure 20% · Google Cloud 15%. Carlson (Jul 31): the serial-recognition thesis got hard evidence against the unmonetizable-capex read — MSFT beat EPS by 13% and rose 15.5% in a single day on a $3.4T cap after being −17% YTD, and AMZN's AWS grew 37% to $42B in a quarter while margins rose, with order visibility out to 2028. His frame: the four capex vendors (GOOGL/MSFT/AMZN/META) are "the entire distribution layer of all of AI" and the market re-rates each one's spend serially on its own earnings report — Google first (15→25 PE), now Microsoft and Amazon, leaving META "the last holdout" — which is why he added $4,000 to META into its post-earnings drop. Ives/Luria (Eisman Ep 70, Jul 27): Ives — "year 3 of an 8-to-10-year buildout" (the Vegas strip, 1955), chip demand-to-supply 15:1 on his Asia trip, "gut-check moments three to four times a year" the norm. Luria's demand anchor: OpenAI + Anthropic run-rate is already >$75B (>$100B adding Gemini/Meta/xAI) "from what was zero a couple of years ago" — real economic activity underneath the capex, whoever wins the model layer. App Economy (Aug 1): Samsung's Q2 extended the memory shortage horizon again — constraints worsen in 2027 and now persist through 2028, with this year's unmet demand deferred to next; 60–70% of HBM capacity locked under multi-year contracts (five AI data-center customers signed, five negotiating), first HBM4E samples shipped, DRAM/NAND guided up 15–20% sequentially into Q3. Record revenue (+130%) and profit (18x) still met a stock down 40% from its June peak in a chip-sector rout — "a low P/E means little when the market is betting the peak is near." 2026-AUG-03 (David Hay / Haymaker portfolio update): the rising-margin defence of the megacaps is retired — margins came from "the Magnificent Seven-type names with their network effects and capital-light business models," yet "many of these have become highly capital intensive due to the escalating 'arms race' to build out data centers." Hyperscalers are projected to invest "some $700 billion in AI capex this year, with estimates for $1 trillion next year," and "serious questions are being raised about the ultimate returns on these vast sums." 2026-JUL-31 — Dan Niles (panel clip): the AI money goes "into infrastructure which goes right back to chips," and the shortages "are going to continue for the next couple of years" — but the binding link has already moved: "the bottlenecks have switched. You've got CPU bottlenecks now which is why you've seen Intel up over 200% versus an Nvidia that's up like 20% or so." He reads the 200%-vs-20% spread inside one theme as the market marking where scarcity sits, and warns the framework is non-permanent — "the bottlenecks are going to keep switching." The destination of the capital is the physical layer: "the physical infrastructure is where all of this money is being spent, including Google that raised $85 billion recently." 2026-AUG-03 (Joseph Carlson): the capex-bear case breaks on evidence. Amazon's Q2 print — AWS +36.7% YoY to $42.2B, a fifth consecutive quarter of accelerating growth and the fastest in 18 quarters, a $169B run-rate, backlog +154% to $496B with capacity coming online through 2028 already committed, and $16.6B operating income at ~39% margins (expanding on custom silicon/network gear) — is direct counter-evidence to the "capex is destroying hyperscaler economics / they're the new telecoms" case (Michael Burry et al.). Andy Jassy's first-ever "AWS is likely to become a trillion dollar revenue business" claim is credited because commitments, growth rate and secular trend back it. The build-out is contracted demand, not speculative capacity. 2026-AUG-08 — John Polomny (AIA Weekly): the Apollo arithmetic — data-center capex adds 1.7pp of GDP in two years (1.4%→3.1%, ~0.85pp/yr) vs housing's fastest 0.5pp/yr and telecom's 0.15 — "close to twice the pace of the housing boom at its fastest," and "the same arithmetic runs in reverse" (housing's 6.2%→3.0%-of-GDP unwind is what made 2008 severe). Polomny: "I see no path to recovering the capital that's being invested… it's eventually going to bust" ('27–'28 his guess) — a sword of Damocles; "I don't want anything to do with this." The booming ISM (>50) is largely this capex plus reshoring. 2026-AUG-02 — Mohnish Pabrai (New Money): "Two or three players make out big time and then there's a lot of carcasses on the roadside. And we don't know who the carcasses are and we don't know who the winners are." Hyperscalers are forced players — he cites Zuckerberg: "whether the bet works or not, we don't know. But we have to play." Key calibration: headline capex overstates real volume because input prices are 4–5× higher — "when Google spends a hundred billion in 2027, that's like the equivalent of spending 20 billion five or six years ago." His verdict on the whole complex: the too-hard pile. 2026-AUG-08 — App Economy (PRO): two-sided evidence in one issue. Bull: Cloudflare accelerated to 36% growth through a 20% workforce cut with non-human traffic exceeding human traffic for the first time; DigitalOcean AI ARR +212% with RPO up 12× to $894M. Bear: Datadog's largest customer — "a leading AI company" on 17 products that just signed a nine-figure renewal — has begun cutting usage, the decline is baked into H2 guidance, and the stock fell 17% on a raised outlook. Single-customer AI concentration is now a live growth-curve risk for the picks-and-shovels cohort. Carlson (Aug 10): reframes the META capex fear as mispriced by locating a salvage floor under it — the spend buys re-deployable compute, so even a lost race can be "simply turned into a neocloud" that "would pay for the investments that Meta's already made with an attractive return." The stated motive is strategic, not ROI-of-the-quarter: Zuckerberg's same-day essay says the point is not to be "reliant or bottlenecked by other companies," with Anthropic and OpenAI cast as "the Apple and Google of AI." Verdict: "huge ambitions, huge potential upside, but a lot of the downside is already priced into the stock and largely limited." 2026-AUG-14 (Haymaker): His calibration of build-out risk: it "could slow meaningfully, but it is almost certain to continue at a rapid clip, just not as fast as is now projected — the hyperscalers view not keeping pace with their peers as an existential risk." 2026-AUG-04 — Dan Niles: the June-quarter hyperscaler prints cleared his two-part test — acceleration and margin expansion in the same report: AWS +9pt to 37% growth with operating margins +1pt; Azure +~4pt to 43% with margins +1pt; Google Cloud the standout — growth accelerated 19pt, margins +3pt, "and by the way, they're growing 82%… you're seeing both growth and profitability at that layer." Overhang clearing: "the situational awareness getting taken out their public positions… solved my issues with the speed bump, which I saw coming." AUG-16 (Jay Singh, SSR): the depreciation bear case is failing on observable prices. One-year H100 rentals rose $1.70 → $2.35/GPU-hour (Oct-25 → Mar-26), cross-provider on-demand medians $2 → $2.70, B200s clear at $5.30-7/GPU-hour, and a 2020-vintage A100 remains in commercial AI-training use six years later with rising lease rates — "the installed base remains productive well beyond its initial depreciation period, which is one of the bear's main theses." The positioning consequence: Nebius printed +454% revenue growth and rose 34% in a day, and the shorts "completely lost their shirt… including Michael Burry." His rule: price the scarce input in the secondary market before shorting the equity on valuation. 2026-AUG-15 — App Economy Insights PRO (Cisco Q4): a subtraction test for whether the build-out is broad or concentrated — product orders grew 35% but +25% excluding hyperscalers, i.e. ordinary enterprises are refreshing networks for AI, not just the clouds. $4B of Q4 hyperscaler AI orders took FY26 to $9.3B against a ~$9B target and FY27 revenue was guided ~15% higher — yet the stock fell ~4% because only $7.5B of FY27 AI revenue was guided against that backlog, and gross margin fell 2pp to 66% on hardware mix and memory cost. "The debate now shifts from AI demand to conversion." 2026-AUG-17 — Jay Singh (David Lin Report): durable demand for HBM, grid connection, liquid cooling, advanced-GPU premiums, ASICs and Broadcom products — against "a symmetric risk of the cost of overbuilding," with data centres running one-to-two-year delays leaving "a lot of unused GPUs." Anthropic IPO expected September–October at a ~$50B run-rate from single-digit billions a year ago. 2026-08-18 (App Economy Insights, Q2 13Fs): the buildout kept spilling into the physical economy — GE, CRS, LIN, BE, CEG, EQT, UNP, CRH, GEV and BKR all made the quarter's top-buy list, a full stack of generation, power equipment, materials and freight bought "as investors sought ways to participate in the massive data center buildout beyond GPUs." Continuity rather than novelty (the theme "continued to spread"), which is what a durable capex cycle looks like in filings — with the caveat that late-stage diffusion into the most tenuously connected names is also how a theme gets crowded. Noble via Polomny (Aug 19): "the entire AI buildout is shaping up to be housing bust 2.0" — housing, not dot-com, as the analogue: the financing breaks, not the utility. (George Noble, reprinted by John Polomny / AIA weekly, 2026-aug-19.) App Economy (Aug 21): Alibaba runs the same capex playbook as the US hyperscalers — ~$10B quarterly CapEx (+75%), FCF a −$6.6B outflow, group operating margin cut from 14% to 6% — but supplies the disclosure the US names mostly withhold: a stated ~3-year breakeven on AI compute assets, "comfortably within their expected useful life," trending toward ~2.5 years as utilization, Cloud margin and in-house silicon mix improve. The corroboration is segment-level: Cloud revenue +45% against adjusted EBITA +133% to $830M (~12% margin) — profit compounding three times faster than revenue while capacity scales hardest. (App Economy Insights, 2026-aug-21.) 2026-AUG-15 (Mark Newton, Fundstrat, Jimmy Connor): not a bubble, but the thing to watch — Google "already spending over 200 billion on CapEx" while raising another "25, 30 billion… these are extraordinary numbers." Verdict: "I don't sense that it's a bubble. These companies are making really an extraordinary amount of money. We're almost in a time like the late '90s" — with 2028–2030, not now, the window he'd worry about. 2026-AUG-19 (Ted Oakley, Oxbow Advisors, The Real Story): back out the hyperscalers' new debt and the depreciation still to land and "I don't think those earnings will hold up" 24 months out. The template is Cisco 1999–2002: the paradigm arrived, the fibre/router capacity overshot it — "it ended up in different ways, not what everybody thought it was in '98 and '99." He declines the trade entirely and owns the physical inputs. 2026-AUG-27: Hay: no precedent at this scale — per a Columbia Business School chart relayed by Luke Gromen, “prior eras of excessive spending to finance efficiency/technological breakthroughs have never represented such a large share of GDP.” The AI-productivity rebuttal “may well turn out to be the case,” but both the federal and AI spend risk disappointing. (David Hay, Haymaker Daily 2026-AUG-27) 2026-AUG-28: App Economy: NVDA revenue +106% to $96.2B (DC +117%), Q3 guided to $108B assuming zero China compute, FY28 guided ~+70% while still supply-constrained; Rubin lifts the take per gigawatt ~$18B (Hopper) → ~$25B (Blackwell) → ~$40B (Vera Rubin); gross margin bends to ~71-72% in Q4 on memory scarcity — supply-driven, not demand-driven. (App Economy Insights 2026-AUG-28) Kedrosky (Meb Faber #648, Aug 28) supplies the framing case for the bubble side: across 200 years, mega-bubbles combine some of loose credit, a genuine technology story, a real-estate component and a policy angle — "this moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history," which is why single-lens analysis keeps mis-sizing it. He is explicit that the technology works ("they're wildly useful" — required setting for a great bubble; "if it was shitty technology… we wouldn't be having this conversation"). Two specific tells: utilization at 35–40% against a scarcity narrative (hoarding, double/triple ordering), and the original sampling sin — the whole build-out was sized from coder usage, an expansive, tight-grammar, fast-gradient-descent domain that "could hardly be less representative of AI's future," whereas most white-collar use is compressive (40 pages in, five bullets out). Also, capital allocated over 60 years in electrification is being allocated in about four here. Steve Eisman (The David Lin Report, 2026-AUG-28) puts a number on every link of the chain, entirely from public disclosure — "I don't think it's hidden… it's not like I discovered the Rosetta Stone": US GDP grows ~2% this year and "half of it is from AI CapEx"; Nvidia's revenue more than doubled but "70% of its accounts receivable was from five companies"; "Nvidia sells chips to hyperscalers, 70% of hyperscaler AI revenue is from Anthropic and OpenAI, and that equates to about 25 to 35% of their cloud revenue"; "if you look at just Oracle, 50% of its $600 billion backlog is just from OpenAI." Therefore "the entire AI ecosystem food chain is dependent upon the future health and success of Anthropic and OpenAI… that also means the US economy hinges upon" them — and, at its limit, "if tomorrow OpenAI failed, the US economy I think would go into an immediate recession, and the market would have a massive correction." The bubble tell he nominates is second-half revenue growth, not profitability — "eventually profitability matters, but near-term these stocks don't care" — because "token maxing… ended" around late June/July, so H1 measures a spending mood that no longer exists and "the third and fourth quarters are going to be much more interesting than the first half." 2026-AUG-31 — Jeff Keller (Other People's Money): the biggest single risk he names is extrapolating today's extremely high price of compute, "certainly maybe with neoclouds" — and Elon entering the space at scale "might move us quicker to the glut." Structurally the compute market went from three vendors with thousands of customers to nine or ten scaled providers (hyperscalers, Oracle, SpaceX, labs procuring directly, CoreWeave/Nebius) with fewer, more concentrated customers — why AWS reaccelerating from a hoped-for 18% to the 40s barely moved the stock. Contrarian counterpoint: stocks pricing a deceleration in data-center starts (labor, power applications, pre-midterm state moratoriums) are wrong — "we've seen various technology moral panics over time and usually the market is strong enough." Greg Abel (CNBC) 2026-SEP-02 adds an unusual demand data point from inside the economy rather than from vendor revenue: Berkshire underwrote its Alphabet add partly on what its own ~100 operating companies see — "we have a lot of visibility from within our companies as to how we're using AI, what type of benefits it's delivering. So that brought incremental interest." A conglomerate treating internal adoption as the demand channel is evidence the spend is grounded, though he is explicit it sizes the category and does not rank the players ("we saw Google as a significant player… there's a lot more to Google than what I just said"). On the physical side he is a build-out bull with a throttle: data-centre load is "a significant opportunity for Berkshire and Berkshire Hathaway Energy," but rationed by site preparation, not by capital or capacity. Slegers 2026-MAY-10: states the capex return chain as arithmetic anyone can check — hyperscalers are expected to spend $700bn on AI infrastructure in 2026 alone; a 10% return on that requires $70bn of profit, which at a 10% margin requires $700bn of revenue — against total AI-related revenue estimated at roughly $40bn in 2025. Explicitly not a claim of failure ("AI revenue is growing quickly, but it still is very early in the journey"), but a statement of the burden of proof embedded in index prices, alongside a Shiller CAPE over 40 that "approaches the valuation we saw before the 2000s dot com crash." 2026-SEP-02 (CNBC Halftime — Sam Altman / Lebenthal): Altman supplies the bubble-side quote — "I'm not worried about our compute build out plans. I am worried about the world's compute build out plans… the first signs of what feels to me like unsustainable silliness of random new neoclouds popping up, people claiming that they're going to build gigantic massive compute next year that I think they don't have the revenue to support or a buyer." Both Weiss and Liz Thomas discount it on incentives — he "doesn't want more going up because he's already locked in what he needs"; "spoken like a true CEO." The same-day counter-datapoint is order-book evidence from a non-owner: Dell lifts its forecast (Citi and BofA both to $600), which Lebenthal reads as "for all of those people to say AI is a bubble, the CapEx is going to turn off — I don't think so," and carries straight to Oracle ("we are compute constrained and Oracle will benefit"). 2026-AUG-28 (McCracken, Value Hive): an AI PhD's not-yet-a-bubble verdict, with a structural end-date attached. "I still don't think it's anywhere near close to popping. I think this is the last part of the tech tree for this era" — what we have is "virtual intelligence," good at finding where things break (mathematical counterexamples, cryptographic loopholes) but incapable of posing its own conjecture: "absolutely not." He grounds the trillion-dollar capex in the architecture's crudeness — skip connections plus Transformers plus brute scale, "the core idea is incredibly simple… we honestly don't know what we're doing" — which is why the compute bill is a gigawatt-week per hard answer. Investment conclusion: don't bet on the frontier ("Gemini, GPT, Claude, all doing the same thing"), bet on the standardised inputs they all consume. Nine of the 48 names in the AUG-03 SSR compilation sit in the AI physical layer — memory (Samsung, SK hynix ×2, Western Digital), components (Samsung Electro-Mechanics MLCCs), tools (Disco), silicon (Broadcom), switching (Arista, Accton) and cable (Sumitomo Electric) — which is itself a positioning fact about where bottom-up money was concentrated in the June quarter. Baron's Broadcom write-up argues the vertical-integration bear case "remains only a bear narrative at this stage" because ASIC complexity now demands "extreme co-design of compute, memory, input/output on dies, and networking fabrics," citing the Google TPU agreement extended to 2031 and OpenAI's inference chip taped out in nine months. Crossroads adds the power constraint from the other end: FTAI's 25MW gas-converted jet engines exist because "'time-to-power' is the key constraint for data centers facing multi-year turbine backlogs," and one cloud provider signed a $1.465bn opening order. (fund pitches compiled by SSR, AUG-03) (2026-SEP-03, Dan Niles, Excess Returns) Overinvestment is the definition of a real revolution, not evidence against one — "by definition you have overinvestment going on… it means it's a great time to invest. The problem is when that eventually breaks." His measurable test reduces AI revenue to price per token times tokens produced: since end-May price is −50% (open source / open weights) while volume is +2.5×, helped by an agentic phase that "uses 10 to 100 times more tokens than the chat-based AI phase." The profitability gate also passes — the big three clouds went from 35% to 43% revenue growth March to June "and more importantly, the operating margins also expanded about 2 percentage points." CNBC Halftime — Jim Lebenthal (2026-SEP-04): the capex-is-profitable case argued entirely from the buyers’ numbers rather than the sellers’ — “Andy Jassy… it takes less than three years to recoup the cost of a new data center… Alphabet growing its web services at 80%, Microsoft Azure at 40%.” He reads NVIDIA’s print as “the clearing event,” with a week’s lag before the read-through landed outside chips (“it’s things like Microsoft, it’s things like Meta”). Sechan leaves the counterweight standing rather than dismissing it: questions remain around “the durability of the earnings, the durability of the spend, the continued ability for markets to finance that.” 2026-SEP-07 (Excess Returns — Dan Niles clip, + hosts): the clearest dated clock yet offered for the demand leg. Niles: AI has run distinct phases — "training initially, then inference, and then on January 30th of this year the agentic phase" — and "the agentic phase uses 10 to 100 times more tokens than the chat-based AI phase," so "you've got a long way to go, at least another year for stocks to go higher… but you're going to have to get more selective and watch the data like a hawk." He rebuts the smart-money appeal with a filing, not an argument: pull Cisco's May 2021 release, where bookings went "from 70% year-over-year growth in several months to negative 30%" at what had been the most valuable company in the world — "do massively smart, big companies get it wrong? Yes, all the time." Zeigler adds ground truth for the durability side: clients "are just now feeling the agentic lift… that happened literally earlier this week," and a buyer who has just measured a return "is willing to spend more budget," implying a second spending leg from the late adopters. Both hosts cap it: "it doesn't mean we're not in a bubble." 2026-SEP-10 — a dated peak-capex call from a Canadian long/short manager (JF Tardif, Timelo). He takes Nvidia's own guide as the bear input, not the bull one: 70% sector growth next year "will make about close to $1.7 trillion of spending overall," which he does not believe is sustainable — "we're very much coming to the top, the peak of spending." The framing is a mountain, not a plateau: after the peak "spending is going down" and "I can imagine that those stocks would go down a lot." Timing is left open, the direction is not: "it could be in 27, maybe it's in 28. That's a real debate. In my mind it's not if." App Economy Insights (2026-SEP-11): the contracted-demand side got its conversion proof. Oracle delivered 850 MW of new capacity in Q1 FY27 (almost triple Q4) and 300,000+ GPUs; OCI growth accelerated 93% to 121% ($7.4B) and RPO still rose $26B Q/Q to $664B, with demand for training and inference 'still exceeding available capacity.' Verdict: 'Oracle no longer needs to prove that AI demand exists'; the open question is whether the economics justify the capital. Shares remain ~50% below the Sept-2025 peak. Astrid Wilde (Value Hive, 2026-SEP-11), founder of robot-data firm Inheritance AI: shortage side. The binding constraint is now manufacturing capacity 'for a lot of things' - sensor vendors that shipped single test units 18 months ago now require 100-1000-unit orders with 3-9-month lead times, 'not yet in public markets'. Hyperscalers model at least ~20% near-contracted returns on data centres, so spending continues 'forever'; asked whether future compute is malinvestment if open-source models are 'enough', her guess is 'probably not'. Dan Niles (2026-SEP-15): every revolution overbuilds, but this is a deceleration, not a collapse. Hyperscaler capex was up ~75% last year and is pacing ~100% this year; it should "slow down dramatically, but still grow" next year, like internet traffic that doubled yearly instead of every three months. Agentic AI (dated to OpenClaw, Jan 31) uses 10–100x more tokens and is still early. "I don't think this is a crash but I do think spending… will slow down next year." Mike Taylor (Hedgeye, 2026-SEP-15) — bubble side: 'three minus one' ($3T installed base needing ~$1T/yr vs ~$100B of lab revenue). Top-50 S&P companies generated ~$5T FCF in 2025 and will give AI 3-4% of it, not the 10% ($500B) needed; stocks up while bonds say no is 'exactly what happened in 06.' Joseph Carlson (Aug-27 / Sep-1, 2026): Nvidia's ~70% FY2028 growth guide is supply-constrained (customers would support ~100%) — "I far underestimated the continued demand"; the build-out is "definitely not" a scam, because end customers keep paying OpenAI/Anthropic and expanding usage after trial, unlike crypto. Peterffy, 2026-SEP-16: bullish on AI earnings (Nvidia 27× vs 70% growth is "relatively low"; open-source adoption alone lifts productivity everywhere), bearish on the spenders: hyperscalers must buy all the compute they can to be one of the one or two winners, so compute prices fall and they "will have to write down much of the compute that they bought." Casey (2026-SEP-05): "if you own any computer stocks or AI stocks… dump them. We're at the top of the bubble"; Nvidia, the largest market cap, is "part of the AI bubble" and the stock market's meltdown will hit pensions. Eisman (CNBC, 2026-SEP-17): "a little bit more nuanced" than an eight-bubble-signs checklist, but the chain from Nvidia through the hyperscalers "all depends on the future health of Anthropic and OpenAI" — "if something were to happen to one of those two companies, then the chain would really fall apart," and OpenAI is "the weaker company," one reason it postponed its IPO. He has "taken some off the table" and isn't adding. 2026-SEP-18 (CNBC Halftime): Amy Raskin calls it "a financially unsustainable arms race for AI dominance that is not going to end well" - roughly $200B of cumulative AI revenue (per Roger McNamee) against ~$1T of capex this year, with higher yields, component prices and power costs all raising the return bar, and too many players for anyone to stop. Rob Sechan concedes "we will overbuild" but "that is not today's problem" - mega-caps de-rated from 34x (2025) to 24x, some to 18x, data-center demand "insatiable." Jason Snipe: AI capex $900B this year, $1.6T next; Goldman says AI drove nearly half of S&P earnings growth this year and the tailwind fades in 2027. George Noble on The Real Eisman Playbook Ep 76, 2026-SEP-21: "show me the ROI." Julien Garran (MacroStrategy Partnership) puts AI malinvestment at 24x the dot-com's; the market may be worth only $50-100B, with Chinese models "good enough for 95% of us at 95% less cost." Peter Berezin's lesson: the 1999-2000 internet traffic bulls were right (43%/yr for 25 years), and it "didn't stop Global Crossing... from going bankrupt." Eisman: ~70% of hyperscaler AI revenue is OpenAI + Anthropic; OpenAI is "in trouble."