Newsletter · · Ashutosh Agarwal

Burry Shorts the AI Accounting While JPMorgan Triples Its Micron Target - The AI Capex Tracker - Week of August 18, 2026

AI capex and semiconductors newsletter for the week of August 18, 2026. Two forensic accounting shops put hard numbers under the bear case, Michael Burry's final 13F alleging roughly $176 billion of hidden depreciation and New Constructs timing a dilution clock in months, while JPMorgan tripled its Micron price target to $1,550 and NERC issued only its third-ever Level 3 grid alert.

The AI Capex Tracker

Week of August 18, 2026: Burry Shorts the AI Accounting While JPMorgan Triples Its Micron Target

Issue: Tuesday, August 18, 2026

TL;DR

  • The bears finally put a name and a hard number on the depreciation worry. A forensic breakdown of Michael Burry's final public filing, his last 13F before he shut his fund and went dark, lays out the whole short in plain English: the big cloud companies quietly stretched how long they pretend an AI server lasts, from three years back in 2020 to nearly six by 2025. That one accounting tweak is hiding roughly $176 billion of depreciation across the tech industry and overstating tech earnings by 20–27%, the argument goes, even as Nvidia makes last year's chips obsolete every single year. Burry's bet is small and cheap (put options costing ~$9–10 million a name), and it isn't that AI fails, it's that the spreadsheets are lying about the physics. (Deep Values, Aug 17)

  • A second forensic shop put a clock on the same worry, and it's ticking in months. New Constructs' David Treanor says Oracle, Amazon, Microsoft and Meta are each burning $200 billion or more of cash over the past year on AI, and each will soon have to dilute shareholders, cut staff, or drop out. His countdown to that squeeze: Oracle ~6–7 months, Amazon ~10–11 months, Microsoft ~1.5 years, Meta ~2 years, with Alphabet the healthy outlier at five-plus years. Oracle alone carries $249 billion of off-balance-sheet debt and, on his math, a negative economic book value. (Money Life, Aug 17)

  • The bulls didn't blink, they raised the target. JPMorgan tripled its Micron price target from $500 to $1,550, memory was again the best-performing corner of the trade, and one widely followed macro shop said flatly that the cycle "is wiping the floor with Michael Burry and his ilk" because the hyperscalers' assumptions are still conservative, not stretched. Two specialist tech analysts piled on: the weekend panic that "memory goes to zero and everything goes to optics" was "one of the dumbest threads" they'd seen, memory is still the single biggest wall to scaling AI. (Real Vision, Aug 17; The Six Five, Aug 17)

What's new

A note on the window: this is the Tuesday issue, sweeping roughly the last 24 hours (Aug 17–18). Yesterday's issue was about the split over Nvidia's $500 billion financing plan and the first hard dollar figure on hidden leverage, ~$1.5 trillion of off-balance-sheet commitments. This cycle the bear case got sharper in a very specific way: it stopped being a vibe and became two independent forensic-accounting shops, each with a number and a mechanism, Michael Burry on depreciation, New Constructs on the dilution clock, while the bulls answered not with hand-waving but with a tripled Micron target and a fresh reminder that memory, not optics, is the real bottleneck. Ranked by where the dollars of profit-and-loss land hardest for a book.

1. Michael Burry's final short, decoded: he isn't betting AI fails, he's betting the accounting does. Deep Values, Aug 17: a forensic walk-through of Scion Asset Management's last public 13F (the Q3 2025 filing, right before Burry deregistered with the SEC in November 2025 and moved to his private "Cassandra Unchained" commentary). This is the clearest articulation yet of the depreciation short that's been hanging over the whole trade.

The structure is a barbell. On one side, a small, boring value book (~$68 million) in unloved names like Molina and Lululemon. On the other, a $1.38 billion notional derivatives book betting against AI, specifically 5 million put contracts on Palantir and 1 million put contracts on Nvidia. The scary-sounding $1.38 billion is just the notional value the SEC makes you report; the actual cash at risk is only about $9 to $10 million of option premium per position. So this is a defined-risk, asymmetric bet, risking ~$10 million to make a fortune if tech re-rates.

Here's the mechanism, in plain language. When a hyperscaler buys billions of dollars of AI servers, it doesn't expense the cost right away, it spreads it over the server's "useful life." The forensic red flag: to make current profits look bigger, the big clouds "quietly extended the estimated useful life of these servers... from three years back in 2020 all the way out to nearly six years by 2025." Stretch the schedule, lower the annual expense, and profits magically rise. The problem:

"You have corporate spreadsheets claiming these massive server farms will be economically viable for six years, while the physical reality of the silicon market renders them obsolete in 12 to 18 months."

Nvidia ships a faster, far more power-efficient chip every year, Hopper to Blackwell to Rubin, so it "becomes a financial liability to run the older, power-hungry servers." The forensic estimate is that this mismatch is "hiding roughly $176 billion in depreciation across the broader tech industry" and "overstating corporate tech earnings by roughly 20 to 27 percent." When reality forces write-downs of those "ghost assets," the earnings hit is "violently sudden,"an explicit parallel to the Cisco dark-fiber bust of 2000. Why it matters: this is the running depreciation debate finally attached to a named investor, a dollar figure, and two specific tickers (PLTR, NVDA). It's also cheap enough to be wrong for a long time and still survive, which is exactly why it's dangerous to dismiss.

2. A second forensic shop puts a stopwatch on the cash burn: dilution in months, not years. Money Life, Aug 17: David Treanor, founder of New Constructs, which reads filings footnote-up rather than headline-down. This is the most actionable new bear input of the cycle because it turns "unsustainable" into a calendar.

Treanor names four AI cash-burners, Oracle, Amazon, Microsoft and Meta, each "burning $200 billion or more over the trailing 12 months in cash." His point isn't that they go bankrupt; it's that they will soon have to do something shareholders hate, issue stock, cut staff, or exit the race. His countdown to that forced move:

"Oracle, we show, has about six or seven months. Amazon, about 10 or 11 months. Microsoft, it's got a year and a half. Meta's got two years. Alphabet's got over five years."

The tell he saw early: off-balance-sheet debt. He flagged Oracle back in November (the stock is down 38% since), pointing at a $249 billion off-balance-sheet figure people waved away as "just a commitment." His retort: it's "ante-up money, they've got to be in to play the game," which is why the weakest balance sheets are the ones whose stocks fall when they announce a big AI number. On his model Oracle's economic book value has gone negative (a "no-growth value" of about -$27 a share). His broader warning is the one PMs should sit with: every AI name is "priced as if they're going to be the winner... there's no way they all win," and the lasting profits will accrue to users and to what Chamath calls "the harness" (helping companies wring value from their own data): not to the chips, memory, data centers, energy or the models themselves, which he calls commodities. He thinks Meta and Oracle are the two "most in danger of going to the great beyond." Why it matters: pair this with Burry and you have two unrelated forensic desks arriving at the same destination, one via depreciation, one via free cash flow, and Treanor is the one holding a clock.

3. The bulls answered with a number of their own: JPMorgan tripled Micron. Real Vision, Aug 17: Andreas Steno and colleagues on Macro Mondays, one of the loudest memory bulls of the cycle. Rather than argue with the bears, they raised the stakes.

The headline: JPMorgan raised its Micron price target from $500 to $1,550, more than triple, and memory was again "probably the best performing trade over the past week." Their framing is a direct shot at the short sellers:

"Once again, the AI cycle is wiping the floor with Michael Burry and his ilk... the assumptions made by the hyperscalers, the assumptions made by market analysts around the memory stocks, they're incredibly conservative still, in sharp contrast to the run-up to the year 2000."

Their supporting evidence on the depreciation fight is the same real-world data point the bulls keep returning to: on its latest call, CoreWeave said it had re-rented its A100 (Ampere) GPUs for another three years even though those chips are already five to six years old, i.e., renting them out longer than the depreciation period the bears assume. They add a genuinely new risk to watch, though, which cuts the other way (see the memory read-through): a US–South Korea rift over Korean high-bandwidth memory allegedly leaking into China via Malaysia, which could snag Samsung's and SK Hynix's Chinese-fab licenses right as HBM becomes "maybe more important than GPUs." Why it matters: the bull case is no longer just conviction, it now has a Street price target three times higher and a concrete supply-side scare to trade around.

4. The price-war math nobody wants to run: a real fight would gut the frontier labs in an hour. AI to ROI, Aug 18: the only Aug 18 episode in the sweep, and it does the arithmetic the "China caught up" headlines skipped. A frontier model costs roughly $6–8 per million output tokens to actually serve. Anthropic charges about $25 per million for its Opus model, a ~70% gross margin. Drop the price to the $4–6 range where Meta's open model and the cheap Chinese open-weight models sit, and:

"If Anthropic actually cut their Opus pricing to match... the margins would flip from that 70% [positive] to... a negative 65%."

In other words, a genuine price war "would gut the business model for Anthropic and OpenAI almost overnight, probably in an hour." Neither lab can absorb it: Anthropic's revenue run-rate is ~$47 billion (up from $9 billion at the end of last year) but it has raised the equivalent of 2.8 years of today's revenue in cumulative capital and owes $350 billion+ of data-center commitments over three to five years, on a gross margin running nearer 40% than the 50% it forecast; OpenAI is at ~$30 billion of revenue but a projected $14 billion operating loss this year and data-center commitments "approaching $1 trillion." The episode's constructive twist is a metric worth stealing, cost per completed task, not price per token: on an agent job, a top model that succeeds 90% of the time can be cheaper to finish the work than a model priced 75% lower that only succeeds 75% of the time, once you add ~$17 to fix every failed attempt. Why it matters: the bear case now has a demand-side detonator (a Chinese-triggered price war) on top of the supply-side one (depreciation), and it explains why the labs "basically have to raise money forever."

5. America's grid regulator just issued its third-ever red alert, and named the number. POLITICO Energy, Aug 17: Jim Robb, CEO of NERC, the body responsible for North American grid reliability. NERC just issued a Level 3 alert, its highest, and only the third in ~30 years, an "essential actions" order telling utilities exactly what they must do to safely connect giant data-center loads.

The scale: NERC's aggregation of utility plans points to about 200 gigawatts of new capacity needed over the next five to seven years, roughly half of it data centers. Robb's yardstick: "200 gigawatts is 200 San Franciscos,"and a single gigawatt is "the size of a nuclear plant." The reliability problem isn't just how much power; it's how these loads behave. Unlike a city of people flipping switches independently, a data center runs flat-out 24/7 and is hypersensitive to power quality, so a grid disturbance can make many of them trip offline at once:

"We had a big one two or three weeks ago in Virginia where there was a voltage fluctuation on the line and we lost three gigawatts of data center load... And that wasn't a single three gigawatt data center. That was multiple smaller data centers... They all made the same decision and they all said, I'm out."

There have been three or four such events in two to three years. Robb's fix leans on gas ("the great thing about a gas turbine is that they can be made quite large... but supply chains are pretty well congested"), with solar-plus-batteries as a complement ("batteries aren't power plants, they're time machines"). His kicker for the AI race: "the only thing that China is ahead of us in the AI race is power." Why it matters: this reframes the power story from queue delays (a scheduling problem) to grid reliability (a systemic one), it's now backed by a federal-facing regulator, and it hardens the bull case for firm on-site generation and grid-reliability equipment, while flagging a fat-tail risk (correlated data-center trips) that could hit uptime for everyone on a circuit.

6. Behind-the-meter is now the base case, but the smart-money version is "flexible," not "off-grid." The Infrastructure Investor, Aug 17: EQT's AI-infrastructure lead, who manages a business with ~250 gigawatts of power across its strategies. Where yesterday's story was Amazon building a fully islanded 7.6 GW gas plant, EQT lays out the more nuanced model the infrastructure funds actually prefer:

"What's happening much more right now is either fully behind-the-meter gas complemented with batteries... often a bridging solution to the grid, and then the gas acts more as a backup solution eventually."

The clever part is flexibility: a data center that can pop off the grid on command "effectively [becomes a] very large battery," which lets it add huge base-load without driving up everyone's bills, "data centers should be reducing the average utility bill, not increasing it." West Texas (land plus cheap gas) is ideal but not tier-one for latency; internationally, EQT sees the build concentrating where firm power exists, the Nordics, Spain/Portugal (solar), Italy, and India ("building power generation at unprecedented scale"). Training goes where the power is; only inference has to be local. Why it matters: it tells you the durable winners in power aren't just gas turbines, they're the integrated developers who can co-locate generation and load and offer grid flexibility, and it's a reminder that the "just skip the grid" narrative has a more bankable, grid-friendly cousin.

The debate

Steel-manning both sides of the hyperscaler capex thesis (call it ~$800 billion in 2026 heading toward ~$1.5 trillion in 2027). Yesterday the fight was whether Nvidia's $500 billion financing plan is clever plumbing or hidden leverage. This cycle it sharpened into a cleaner question: now that two forensic-accounting shops have each attached a number to the bear case ($176 billion of hidden depreciation, and a dilution clock ticking in months), do the bulls have anything harder than conviction to answer with? They do: a tripled Micron target and the reminder that memory, not optics, is the wall.

Bull: the demand is real, the assets last longer than the bears claim, and the money keeps showing up. Nvidia is throwing off roughly $1 trillion of projected free cash flow over three years and recycling it into the ecosystem; the $500 billion institutional-credit plan (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR, with Nvidia backstopping 25%) is "another $500 billion in sales" that quietly locks customers into Nvidia and "won't be the last one" (The Six Five, Aug 17). On the depreciation scare specifically, the bulls' rebuttal is empirical: CoreWeave is re-renting five-to-six-year-old A100s for another three years, evidence the chips out-earn the schedule, not the reverse (Real Vision, Aug 17; The Six Five, Aug 17). And the memory bulls now have a Street catalyst: JPMorgan's Micron target tripled to $1,550, memory "wiping the floor with Burry," assumptions still "conservative" versus 2000 (Real Vision, Aug 17). Even a skeptic like Steve Eisman won't short it, "we're market weight technology... I do think the demand is exceeding the supply" (The Real Eisman Playbook, Aug 17).

Bear: the accounting is flattering the earnings, the cash is running out on a clock, and China can break the price. Michael Burry's final 13F targets the depreciation mismatch directly: schedules stretched from three years to nearly six, ~$176 billion of depreciation hidden, tech earnings overstated 20–27%, and a Cisco-2000 ending when the write-downs come (Deep Values, Aug 17). New Constructs' Treanor supplies the timer: Oracle/Amazon/Microsoft/Meta each burning $200 billion+ a year, dilution forced in 6–24 months, Oracle's economic book value already negative and $249 billion off the balance sheet (Money Life, Aug 17). AI to ROI adds the demand-side detonator: match Chinese open-weight prices and Anthropic's ~70% gross margin flips to about -65% overnight, on top of OpenAI's ~$14 billion operating loss and ~$1 trillion of commitments (AI to ROI, Aug 18). And even Eisman concedes the concentration bomb: ~70% of hyperscaler AI revenue traces to just OpenAI and Anthropic, "if they ever get in trouble, the ecosystem is in trouble" (The Real Eisman Playbook, Aug 17).

The synthesis. The two camps still agree on the facts and disagree on the timing, but the bear side got more precise this cycle, and precision is what starts to move books. Both accept the chips out-earn the crude three-year scare (a point for the bulls). Both accept that free cash flow has gone negative, the financing is increasingly debt, and the enterprise ROI is unproven (points for the bears). What's new is that the bears now hold a number (Burry's $176 billion) and a clock (Treanor's 6–24 months), while the bulls hold a catalyst (JPMorgan's tripled Micron target) and a rebuttal (CoreWeave's aging-GPU rents). Eisman's is the most useful middle: he's market-weight, won't short it, is actually more bullish on the semiconductors than the hyperscalers, but warns it "could be a little bit of a race to the bottom." The real question has narrowed to two triggers: does the enterprise revenue (and a clean Nvidia demand quarter) show up before Treanor's dilution clock forces an equity raise, and does anyone actually start the token price war that would vaporize the labs' margins?

Sell signals to watch:

  • The dilution clock starting. Treanor's near-term names are Oracle (~6–7 months) and Amazon (~10–11 months); the first surprise equity raise or "efficiency" headcount cut from a burner is the tell that the cash math bit (Money Life, Aug 17).

  • A real token price war. If Anthropic or OpenAI actually cuts headline model prices to match open-weight/Chinese models, the margin flips negative "in an hour,"watch published API prices, not press releases (AI to ROI, Aug 18).

  • An AI write-down. Burry's thesis only pays when a hyperscaler actually shortens a depreciation schedule or writes down "ghost" servers; the first such disclosure validates the $176 billion estimate (Deep Values, Aug 17).

  • Correlated data-center trips. Another Virginia-style event, multiple sites dropping ~3 GW together on one disturbance, is both a reliability and an uptime/SLA risk for the whole complex (POLITICO Energy, Aug 17).

  • A US–Korea HBM rupture. If Washington actually squeezes Samsung/SK Hynix China-fab licenses, the memory bull trade gets a supply scare that cuts both ways, bullish price, bearish availability (Real Vision, Aug 17).

Stocks in play

NVDA. Bull: it is generating roughly $1 trillion of free cash flow over three years and using it to lock the ecosystem in, the $500 billion institutional-credit plan is "another $500 billion in sales" and "won't be the last," a widening moat rivals can't match (The Six Five, Aug 17). Bear: it is one of Burry's two named shorts (1 million put contracts), and its own annual chip cadence, Hopper to Blackwell to Rubin, is the thing that makes six-year server depreciation a fiction (Deep Values, Aug 17). Next catalyst: its quarter in late August (~Aug 26–27), the demand read for the whole complex, now the referee between Burry's short and the bulls' "another $500B in sales."

AVGO. Bull: the custom-silicon-away-from-Nvidia theme keeps compounding, this cycle even Cisco underscored it, running almost its entire portfolio on in-house Silicon One and touting AI attach across hyperscaler and on-prem switches, the ecosystem where Broadcom sits (The Circuit, Aug 17). Bear: no fresh Broadcom-specific print this cycle; it rides the complex's beta into a jumpier credit-and-accounting backdrop. Next catalyst: hyperscaler ASIC ramps and read-through from Nvidia's late-August quarter, otherwise quiet in the podcasts this cycle, no direct signal.

AMD. Bull: Eisman is actually more constructive on the semiconductors than the hyperscalers, on the logic that the clouds "are going to continue to spend money" regardless of their own returns (The Real Eisman Playbook, Aug 17). Bear: same Eisman flags it "could be a little bit of a race to the bottom" for some players, and there was no AMD-specific product catalyst (MI450X/Helios) in this window. Next catalyst: its second-half data-center step-up; Nvidia's late-August read-through.

MSFT. Bull: it's the strongest-balance-sheet counterparty of the group and its cash generation is still the best of the four, ~$19 billion in the quarter, with the best five-day price move "since the IPO in '87" (The Real Eisman Playbook, Aug 17). Bear: that $19 billion was down ~25% year-on-year, and New Constructs' clock still gives Microsoft only "about a year and a half" before the AI burn forces dilution or cuts (The Real Eisman Playbook, Aug 17; Money Life, Aug 17). Next catalyst: whether cloud demand broadens enough to reverse the year-on-year cash-flow decline.

GOOGL. Bull: it is the balance-sheet winner of the capital war, New Constructs gives Alphabet "over five years" of runway, far more than any peer, and its ability to float an $85 billion raise means "the weaker players aren't going to have it" (Money Life, Aug 17). Bear: that same $85 billion raise "is shocking", Google hadn't tapped the equity market since the early 2000s, a sign of how capital-intensive the build has become; the stock is back at the highs, pricing in a lot (The Real Eisman Playbook, Aug 17). Next catalyst: whether Cloud growth converts to real free cash flow rather than more commitments.

AMZN. Bull: management just gave what one client called "the most bullish conference call in my life," signaling it will spend "way more" on AWS capacity into strong demand; new all-time highs (The Real Eisman Playbook, Aug 17). Bear: Amazon has been free-cash-flow negative for the trailing 12 months and the burn is "accelerating," which on New Constructs' math leaves it ~10–11 months from a forced dilution/cost-cut decision (The Real Eisman Playbook, Aug 17; Money Life, Aug 17). Next catalyst: evidence AWS AI demand is broad and third-party, not concentrated in a few frontier labs.

META. Bull: it retains the most optionality on ad-funded AI, and it isn't near the front of anyone's dilution queue, Treanor gives it ~2 years of runway (Money Life, Aug 17). Bear: it's the clearest capex-detached-from-core case, quarterly cash flow of just $785 million, and Treanor names Meta (with Oracle) as "most in danger of going to the great beyond... least likely to be around in five years" (The Real Eisman Playbook, Aug 17; Money Life, Aug 17). Next catalyst: whether ad-revenue growth keeps outrunning the capex line.

Read-throughs

  • Power/grid: the story graduated from queue delays to grid reliability, and it's now a federal alert. NERC's Level 3 "essential actions" alert (only the third ever) frames ~200 GW of new demand over 5–7 years, half data centers, as a reliability problem: giant 24/7 loads that trip together, as when a Virginia voltage wobble knocked ~3 GW of data-center load offline in one event. Gas turbines are the fast fix but supply chains are "congested"; solar-plus-batteries complements but "batteries aren't power plants" (POLITICO Energy, Aug 17). The actionable posture: bullish firm on-site generation (gas turbines and the integrated developers who co-locate load and generation, per EQT's "flexible, not fully off-grid" model with ~250 GW behind it, The Infrastructure Investor, Aug 17) and grid-reliability equipment; and a fat-tail risk, correlated trips, that could dent uptime for the whole complex. One caveat on siting: ERCOT's "connect and manage" approach has brought ~102 GW of solar/wind/storage online while MISO and PJM are still working eight-year-old interconnection queues, which is exactly why data centers keep coming back to Texas (Factor This, Aug 17). No direct Vertiv (VRT) or Eaton (ETN) commentary surfaced by name this cycle, flagged as a gap, not a signal.

  • Memory (Micron, SK Hynix): still the loudest bull leg, now with a Street target and a geopolitical wildcard. JPMorgan tripled its Micron target to $1,550, memory was the week's best trade, and two specialist analysts called the "memory-to-zero, everything-to-optics" weekend narrative "one of the dumbest threads" they'd seen, memory remains "the single biggest and most significant wall for scaling AI," and the point that Rubin uses less memory per GPU is misleading because it needs more chips, so total memory demand rises (Real Vision, Aug 17; The Six Five, Aug 17). The new wildcard: a possible US–South Korea rift over Korean HBM allegedly rerouted to China via Malaysia, which could threaten Samsung/SK Hynix Chinese-fab licenses into the Sept 24 Fed meeting and a Trump–Xi summit, a supply scare that is bullish price but bearish availability (Real Vision, Aug 17).

  • Optics/networking: the prints split, and the attach-rate hype is running ahead of the manufacturing. Lumentum beat across the board and guided up; Coherent's revenue guide was fine but margins disappointed, and the two stocks diverged sharply on the week, the tell is in what they sell, visible in the gross-margin gap. Coherent is pushing into six-inch indium-phosphide wafers (theoretically better margins) but there's only one immature supplier, so it isn't paying off yet; Lumentum is executing on four-inch and can reuse that IP for near-packaged optics (The Circuit, Aug 17). The bigger read-through for anyone modeling optical content per rack: it's "the wild, wild west,"everyone's on custom designs, and the market is confusing "scale-up" (within a rack, barely happening) with "rack-to-rack" (happening now at 800G–1.6T). In-rack co-packaged optics (CPO) is far off and can't scale to the attach the bulls assume because the manufacturing bottleneck is too deep. Net: favor Lumentum's execution over Coherent's near-term margin drag, and fade the most aggressive optical-attach models. Cisco, meanwhile, quietly became an AI-networking story again on its own Silicon One custom silicon (The Circuit, Aug 17).

  • Frontier labs and China: the concentration bomb now has a price-war fuse. Roughly 70% of hyperscaler AI revenue traces to OpenAI and Anthropic, so the labs' solvency is the ecosystem's solvency (The Real Eisman Playbook, Aug 17). The new risk is that a Chinese open-weight model (Kimmy K3 "nearly matched" the frontier at a fraction of the price) forces a price war that flips Anthropic's ~70% gross margin to ~-65% and can't be absorbed by either lab's cost structure (AI to ROI, Aug 18). The read-through: this is the single cleanest way the whole capex thesis unwinds from the demand side, and it makes the NeoClouds and any name over-indexed to a single lab the most fragile.

  • The dilution clock is the quiet macro tell. Oracle (not one of our seven, but the canary) is down 38% since New Constructs flagged its $249 billion off-balance-sheet debt, and now carries a negative economic book value; the four big burners are each torching $200 billion+ a year, and the healthiest (Alphabet) has ~5 years of runway while Oracle has ~6–7 months. The first forced equity raise or "efficiency" layoff from a burner is the moment the accounting debate becomes a price move (Money Life, Aug 17).

What changed vs last issue

Yesterday's issue (Monday, Aug 17, "Nvidia becomes AI's central bank. The debt hits $1.5T.") crystallized the $500 billion financing plan into a single disagreement, Gavin Baker's "central bank of AI" bull framing versus Robin Wigglesworth's "this is a debt cycle,"and delivered the first hard off-balance-sheet number (~$1.5 trillion, up from ~$1 trillion a quarter earlier), plus OpenAI's own study finding no revenue link to usage. This cycle the debate advanced on five fronts:

  • The depreciation bear got a name, a number, and two tickers. New this cycle: a forensic breakdown of Michael Burry's final 13F, ~$176 billion of hidden depreciation industry-wide, tech earnings overstated 20–27%, schedules stretched from three years (2020) to nearly six (2025), 5M Palantir puts and 1M Nvidia puts at ~$9–10M premium each, Cisco-2000 as the template. Yesterday depreciation was still a framework (Kedrosky) plus Jensen's rebuttal; now it's a specific short with a dollar figure.

  • A second forensic shop added a clock. New: New Constructs' David Treanor timed the cash burn: Oracle ~6–7 months, Amazon ~10–11 months, Microsoft ~1.5 years, Meta ~2 years, Alphabet 5+ years to forced dilution, with each of the four burning $200 billion+ a year, Oracle at $249 billion off-balance-sheet and negative economic book value. Two unrelated forensic desks (Burry on depreciation, Treanor on cash) now converge.

  • The bull case answered with a Street catalyst. New: JPMorgan tripled its Micron price target to $1,550; memory was the best trade of the week; the bulls argued the cycle is "wiping the floor with Burry" and assumptions remain conservative. Yesterday memory's rebuttal was qualitative (Gil Loria's "not a commodity"); now there's a hard number on the tape.

  • The price-war/China threat got quantified. New: match Chinese open-weight prices and Anthropic's ~70% gross margin flips to ~-65% "in an hour"; Anthropic ARR ~$47B (from $9B) but 2.8 years of revenue raised as capital and $350B+ commitments at ~40% gross margin; OpenAI ARR ~$30B with a ~$14B operating loss and ~$1T commitments; Kimmy K3 "nearly matched" the frontier cheaply. This is a demand-side detonator to sit alongside the supply-side depreciation one.

  • Power went from a queue story to a national reliability alert. New: NERC's third-ever Level 3 alert; ~200 GW needed over 5–7 years (half data centers); a Virginia event where a voltage wobble tripped ~3 GW of data-center load simultaneously; gas-turbine supply chains "congested." Yesterday's power story was Amazon routing around ERCOT with a private 7.6 GW gas island; this cycle the regulator quantified the whole-grid risk, EQT laid out the "flexible behind-the-meter" model (~250 GW behind it), and Factor This contrasted ERCOT's 102 GW "connect and manage" with MISO/PJM's eight-year queues.

  • Eisman surfaced first-hand and stayed on the fence. Yesterday he was second-hand (via RiskReversal) as "newly skeptical but not yet alarmed." Now, in his own words: market-weight, won't short it, more bullish on semis than hyperscalers, but calling it a possible "race to the bottom." His fresh cash-flow stats: Meta $785M quarterly, Microsoft ~$19B (down ~25% YoY), Amazon negative for 12 months; Google's $85B raise "shocking."

  • Fresh, concrete data points: Burry ~$176B hidden depreciation / 20–27% earnings overstatement / depreciation 3yr (2020) → ~6yr (2025) / 5M PLTR + 1M NVDA puts / ~$9–10M premium each; New Constructs dilution clock, Oracle ~6–7mo, Amazon ~10–11mo, MSFT ~1.5yr, Meta ~2yr, Alphabet 5yr+; each of Oracle/Amazon/MSFT/Meta burning $200B+ TTM; Oracle $249B off-B/S, no-growth value ~-$27/sh, down 38% since November; JPMorgan Micron PT $500 → $1,550; Anthropic ARR ~$47B (from $9B), ~40% gross margin, $350B+ commitments, price-war margin ~+70% → ~-65%; OpenAI ARR ~$30B, ~$14B operating loss, ~$1T commitments; frontier serving cost ~$6–8/M output tokens, Opus ~$25/M; NERC ~200 GW / 5–7yr (half data centers), Level 3 alert (3rd ever), Virginia ~3 GW simultaneous trip, 3–4 such events in 2–3yr; EQT ~250 GW; ERCOT ~102 GW connect-and-manage vs MISO/PJM ~8-yr queues; Micron memory best weekly trade; Meta quarterly FCF $785M, MSFT ~$19B (-25% YoY), Amazon FCF negative TTM; Google $85B raise; CoreWeave re-renting 5–6yr-old A100s for +3yr; ~70% of hyperscaler AI revenue from OpenAI + Anthropic; Lumentum beat/guided up vs Coherent margin miss.

  • Quiet vs yesterday: no fresh Broadcom (AVGO) or AMD company-specific print; no Vertiv (VRT) or Eaton (ETN) commentary by name in the power sweep; no new off-balance-sheet total beyond yesterday's ~$1.5 trillion; no fresh update on the $500 billion plan's mechanics beyond The Six Five re-affirming it.

  • Still open: whether Burry's depreciation short or the memory bulls' "conservative assumptions" is right; whether Treanor's dilution clock actually starts ticking (an equity raise or layoff from a burner); whether a China/open-weight token price war actually breaks out; whether Nvidia's late-August quarter shows demand it isn't itself financing; and whether the US–Korea HBM tension escalates into a real supply scare.

Next catalysts to watch: Nvidia's quarter in late August (~Aug 26–27), the demand read for the whole complex and the referee between Burry and the bulls; the September 24 Fed meeting under the new Warsh regime, which also frames the US–Korea/HBM geopolitics; any forced equity raise or "efficiency" layoff from Oracle/Amazon as the first crack in the dilution clock; NERC's year-end filing to FERC codifying data-center interconnection standards; and a Trump–Xi summit as the wildcard for both HBM supply and the open-weight price war.