1The AI drawdown stayed inside the build-out while the index printed records
The semiconductor index fell about 29% from its June peak while the S&P 500 sat near all-time highs. The drawdown never left one end of the AI trade, and Ross Givens, in a video titled "The Biggest Wealth Transfer of the Decade Starts NOW", argues that is the shape of a handoff rather than a bust: money left the companies building the capacity and entered the ones using it, with nothing fundamental broken in between. The single-session tape carried the same shape on 4 August, when fell more than 8% on record revenue — a multiple-compression reaction, not a miss — while rose about 30% after beating and raising its outlook.
His case is historical rather than modelled, and he says so. Railroads: 20% of US track in receivership by 1877, 40% of bonds in default, the stocks down 60%. Fibre: more than $500bn laid, less than 2% of it lit, Global Crossing bankrupt in January 2002. Airlines: cumulative industry profit of roughly zero through 1992, and still earning about 6.8% on an 8.2% cost of capital in 2026. In each case the builders financed the capacity and the users kept the profit.
The test he runs on any AI holding is three questions. Does it own the customer, the data or the standard, or is it renting a commodity anyone else can rent? Must it borrow to grow? And when compute gets cheaper, is that a headwind or a tailwind? Levered neoclouds, retrofitted bitcoin miners and memory fail all three, which is where his sell list comes from. is the worked long: 19 of the top 20 drug companies run FDA regulatory paperwork on Veeva, a licence costs roughly $1,800–2,000 per sales rep per year against about $130,000 all-in for that rep, and the stock fell about 50% in the first half of 2026 on a "SaaS apocalypse" read. Veeva is giving its AI agents away free through 2030 — his point being that the moat was the regulated record, never the intelligence.
One accounting detail carries more weight than the analogies. Every large hyperscaler has extended the assumed useful life of its servers — three years to four, four to six — except , which shortened it in its January 2025 10-K at a cost of $1.4bn of extra depreciation, and Givens notes the stock peaked around that disclosure. On the application side he points at crossing 30 million paid Copilot seats, 's Agentforce annual recurring revenue going from $800m to $1.2bn in a single quarter, and down 31% year-to-date into July and turning around fast. sits on the other side of his ledger: cash flow gone negative, long-term debt nearly doubled.
Two readings compete for the same fact. The hosts of Animal Spirits, on The Compound, treat the divergence as healthy rather than significant — the "legitimate bull market" at all-time highs while the "degenerate bull market" of meme baskets, 2x levered semis and a DRAM ETF is in shambles. That is a leverage story, and it implies the build-side names come back once the forced selling is done. Nicholas Owens of Morningstar, on Prof G Markets, lands closer to Givens from a third direction, arguing that value in Nvidia's "five-layer cake" should accrue upward to products and applications over time. What none of the three offers is a date. Givens concedes he cannot time it, and hedges that he is not telling anyone to sell their AI holdings — only that owning build-side names alone is "one position ten different ways."
2The AI stack's demand side runs on borrowed money, and its route to consumers is now in court
If profit is moving up the stack, it matters who sits at the top and how they are paying. Anthropic is in early talks for roughly $36bn of Blackstone debt to fund leasing Google AI chips — on top of about $132bn of venture money raised and roughly $35bn already borrowed months ago, per Bloomberg reporting relayed on Prof G Markets. That is revenue funded by its customer's balance sheet, which is a different quality of demand from a cash-funded order.
The concentration shows up one layer down. reported its first post-IPO quarter: revenue up 92% year on year with space, connectivity and AI all beating, plus an partnership to put datacentre compute in orbit. Shares rose about 10% into the print and fell about 8% after hours. Nicholas Owens of Morningstar came in roughly $200m light of his own revenue and operating-margin forecast and counted about $11bn more capex than he had modelled; his fair value stays $62 a share against a price already halved from $225. The disclosed number that carries his call is customer concentration — 38% of total revenue from two customers, at 18% and 12%. Owens and the host both infer one of them is Anthropic, and both state that as an assumption rather than a disclosure, which would make a single counterparty renegotiation worth about 19% of revenue.
Then the legal front. asked a federal judge on Monday to bar OpenAI and two former employees from using claimed trade secrets, and to fast-track the case citing an "imminent threat"; the hearing is set for 1 October. Alex Heath, of the Sources newsletter, reads it as a fight for delay rather than damages — an injunction would freeze OpenAI's first hardware device, an ambient speaker due to be unveiled later this year, with earbuds, a wrist device and glasses to follow and explicitly no phone. He notes that OpenAI's public reply called the suit "careless, aggressive, and oddly personal", said it had no interest in using trade secrets rather than denying that it had, and never rebutted the specific coercion allegations. Apple separately confirmed John Ternus as incoming chief executive.
Prof G Markets is structurally bearish on AI, and its conclusion that the unwind is a question of when rather than if is editorial, not evidence. The arithmetic underneath it is not. If the labs' compute bills are debt-funded and their route to consumer revenue is contested in court, then the revenue those labs generate for their suppliers is more conditional than a backlog makes it look.
3A $45bn fund unwound in 20 sessions, and leverage only explains the speed
Situational Awareness, Leopold Aschenbrenner's fund, went from $9.3bn of assets in March to $20bn in June to $45bn in July, sent an investor letter on 24 July framing the selloff as a buying opportunity, then sold its entire public levered book to Citadel at roughly a 10% discount on 30 July. It ran at about 4x leverage. It is down around 70% month-to-date and still up 80% on the year, carried by an early Anthropic stake.
The damage was measurable across the tape. Warren Pies, cited on Animal Spirits, put semiconductor single-stock volatility at 75% during the liquidation — only the third time ever, after March 2020 and April 2025. Six of the seven Mag-7 names were in drawdowns of 15% or more on 29 July; two days later was down 15% and up 20%, and had fallen 10% at one point before recovering. In Korea the KOSPI fell about 40% in five weeks and is still up roughly 50% on the year, with a reported 3.5% of South Korea's population receiving a margin call. Retail sold a net $243m of single stocks in one session last week, the largest one-day outflow since the COVID crash — and it marked the low.
Where the two accounts split is cause. The Animal Spirits reading is the standard one: leverage is the thing that blows funds up. Givens rejects it explicitly — 4:1 leverage set the speed, 20 sessions instead of two years, but the direction came from being wrong on both legs, long the physical build-out and short application software. That distinction decides what the episode teaches. If it was leverage, the lesson is about crowding and sizing, of the kind visible in a DRAM ETF that gathered $26bn since its April launch and returned 102% while the money that arrived late still lost about $1.6bn on a dollar-weighted basis, or in up roughly 1,000% in a year and 25% in a single day after the liquidation. If it was direction, then the same trade is still on for anyone who did not have to sell.
4The September hike stopped being the base case on a single payroll proxy
Polymarket's September Fed contract flipped in a day. No change now prices at 52.5%, up three points; a 25bp increase at 45.5%, down two; any cut at about 2.6% — on $14.7m of total volume, $1.17m of it in the last 24 hours, against $2.1m of liquidity. Yesterday the hike led at 55.5%.
The mechanism was one release. ADP total employment rose 44,000 month on month in July, against 95,000 in June and 122,000 in May, on the not-seasonally-adjusted level series — so the deceleration is the signal and the exact figure is not comparable to the headline private-payrolls print. Slower hiring weakens the case that the Fed still needs to tighten, and the hold edged ahead.
It settled nothing. Inflation is sitting exactly where it makes the decision hard: headline CPI ran 3.46% year on year in June and core PCE 0.13% month on month, while Polymarket's July CPI contract drifted higher on the day — 3.4% modal at 45.5%, up 1.5 points, with 3.3% at 31.5% and 3.5% at 16.5%. At $351k of volume that last market is thin enough to treat as weak evidence. Against a 3.63% fed funds rate, a 3.4% print leaves the real policy rate near zero, which is why hold-versus-hike stays close to a coin flip and a cut is barely priced at all.
The long end is telling a separate story. The 30-year sits at 5.23%, the 10-year at 4.70% and the 2-year at 4.25%, with the 10-year-minus-2-year spread flattening to +0.43 from +0.47 on 31 July, and the 30-year mortgage at 6.66% and rising. Bank of America, relayed on Animal Spirits via Joe Weisenthal, described the reaction to new Fed chair Kevin Warsh as resembling an emerging market facing a credibility shock — steeper curve, weaker dollar, equities down, forward guidance withdrawn. Note that the curve part of that description does not match this week's prints, which flattened slightly; what does match is the currency leg, with gold posting its largest daily gain since February and the dollar near a six-week low.
5A ceasefire is priced at 90%; tankers actually flowing is priced at 61.5%
The two Iran contracts moved in opposite directions on the same day. Polymarket's US–Iran effective ceasefire market, on $12.9m of volume, went to 86% by 14 August (up 5.5 points) and 90% by 31 August (up seven) — a large, liquid, one-day move that corroborates the Reuters reporting of Qatari-mediated progress rather than fading it. The Strait of Hormuz normalisation contract, resolving 31 December on $7.1m of volume, went the other way: 61.5%, down a point.
The gap between those two numbers is the whole story. The proposed terms reportedly give Iran control of inbound traffic, Gulf oil exports remain roughly 40% below the pre-war mark, and Houthis attacked a Saudi tanker in the Red Sea this morning. A deal and restored shipping are two different events, and the market is pricing them apart. Equities have taken the first of them: the Dow and S&P 500 closed at record highs on 4 August on hopes the strait reopens, Brent fell below $80 a barrel and the 10-year yield declined. That is a peace trade rather than a supply-restored trade, and a refiner's quarter shows why the distinction bites — beat estimates because the Iran war lifted US refining margins, which makes it a backward-looking print if Hormuz actually reopens.
The administration's own signal points the same way without resolving it: the president reposted a headline calling a deal imminent as Iran talks restart Monday, and Reuters reported an all-day US–Iran negotiation on Tuesday.
What the sources recommended
Seven single-name calls came out of three videos, and six of the seven are one person's list built from one framework — read them as a single thesis, not seven independent opinions.
The buy. (Veeva Systems), from Ross Givens, on a months-to-years horizon: the regulated audit trail as the moat, a licence costing 1–2% of the annual cost of the rep who uses it, and AI agents given away free through 2030. It is a thesis pitch rather than a trade ticket — no entry price, no stop, no size — and the same video pitches a $5-a-year paid membership twice, once mid-thesis, which is worth knowing when weighing how it is framed.
The sells. Five of them are the same test applied five times: (CoreWeave) on term loans at 11–15% funding a cash-burning business, (Nebius) as the same levered-neocloud bucket renting a commodity anyone else can rent, (Micron) and (SK Hynix) on the argument that memory has never once earned its cost of capital through a full cycle, and (Bloom Energy) as capital-intensive build-side power capex. The SK Hynix line is a name-level call rather than a ticker-level one — its US listing could not be confirmed from the transcript. The sixth is separate and comes from Nicholas Owens of Morningstar: on a $62 fair value against a price already cut in half, and 38% of revenue resting on two customers.
The Iran and Fed questions have resolution dates; the AI one does not
July CPI resolves on 12 August with its contract drifting toward 3.4%, and it is the input the 16 September Fed decision keys off. The first ceasefire deadline is 14 August at 86%, with talks restarting Monday, while the shipping question does not resolve until 31 December — which is where the gap between a deal and moving tankers finally gets settled. Apple's preliminary-injunction hearing against OpenAI is 1 October.
The handoff has no such date. Until the application-layer numbers keep compounding — the Copilot seats, the Agentforce revenue — or a build-side balance sheet actually cracks, the same 29% semiconductor drawdown reads as a rotation or as the first leg of a bust depending on which frame the reader brought to it. No single print resolves that one.
