1Three big cloud providers' AI revenue keeps tracing back to the same two loss-making customers
The sharpest material of the day was three separately sourced estimates that land on the same two counterparties. Prof G Markets laid out three of them: Bloomberg puts roughly 70% of Microsoft's AI revenue with OpenAI, a company Microsoft has invested in; Barclays estimates 75% of Amazon's AI revenue comes from OpenAI and Anthropic; UBS estimates those same two account for 30% of Google's cloud revenue. OpenAI lost $21bn last year, and the show puts Anthropic's loss near $11bn. Felix, on Felix & Friends, got to one leg of the same point from the disclosure side — Microsoft said publicly for the first time that about 70% of its AI sales are OpenAI, which he reads as circular revenue rather than third-party demand. Prof G Markets closed on the structural version: the boom is dependent enough on one customer that, in the host's phrase, nothing can go wrong.
The credit market is where Felix says it shows first. His named canary is Oracle , whose five-year credit-default swaps he quotes near 215bp — above where he says they sat the week Lehman failed — on a business with revenue growing, cloud up 93% and a $600bn backlog, but negative free cash flow filled with new debt. Treat that comparison carefully: it is asserted on air without a cited source, in a video whose actionable payload is a seminar signup.
The tape moved on the same worry. Western Digital fell more than 16% on AI capex-spending concerns per TradingKey, and closed among Yahoo Finance's biggest decliners at −10.90%; SanDisk fell more than 10% in the same memory and storage selloff, and SK Hynix was down about 6% in Thursday pre-market. Micron fell with them, though its listing on the sell side today comes from a short book rather than from the tape. Separately, and with no reason recorded in any of the day's sources, the software complex had a much worse session than the index: HubSpot −21.17%, AppLovin −17.49%, with Figma and Datadog named alongside them as leading the decline.
Hold the counter-argument in view, because it uses the same three names for the opposite conclusion. Chan Yang, head of research at Variant Perception, told Excess Returns that the capital cycle reads the other way round: semiconductors remain relatively capital-scarce because they have not over-invested in themselves, and the over-investment sits at their customers, the cloud providers, who have not generated matching returns on it. His next leg is what he calls the Japanese paradox — costs fall, usage rises, and the profit pool broadens from the bottleneck hardware to the AI adopters, with the trusted enterprise incumbents , and as the complementary-asset winners rather than the model labs. So the concentration risk in one frame is the winners' list in the other. What would settle it is whether OpenAI and Anthropic keep spending, and nothing in the day's material puts a date on that.
2The AI short has no dataset behind it the way the 2008 mortgage short did
Michael Burry's Tuesday note, relayed on Prof G Markets, keeps shorts on , , , and the semiconductor index, and says it is "possible we are near a major top", with a "1987 type fall" as the shape. The positions come from the note, not from him on camera.
Steve Eisman, the other 2008 short, rebutted it in the same episode, and his objection is about evidence rather than valuation. In 2008 the pair had monthly securitisation data showing mortgage credit deteriorating in real time; there is no equivalent series for AI, and on Eisman's reading Burry "doesn't have a data point that you and I don't have." His own picture is benign where it can be measured: bank credit statistics reported in mid-July were about as clean as they come, Visa and Mastercard payment volumes are robust, M&A and the IPO calendar are fine. He calls the economy K-shaped, but says the weak leg shows up as Procter & Gamble posting no revenue growth, not as credit stress. His one stated trigger is a price war in the model layer: Chinese open models are cheaper and switching costs are nil, so if OpenAI and Anthropic get into trouble they stop spending, chip orders fall, and the trade unwinds. Until then he will not short it — and he notes he expected a 10-year yield above 4.5% to matter, it did not, and he now thinks it takes 5%.
He is not uniformly bullish. On Meta he is blunt about the quarter: 28% revenue growth against 55% expense growth, and $785m of free cash flow. Chip depreciation moved from roughly $4bn to $6bn in a single quarter, and that capex is only now rolling off the balance sheet and into the income statement, which he expects to weigh for years. The upshot is that the day's best-known bear case and its best-informed rebuttal came out of one 35-minute episode, and neither carries a date: one names a condition, the other names a shape. Nothing here is falsifiable this week.
3A 5.83% growth nowcast put a September rate hike back to a coin flip
Polymarket's September Fed contract — $16.2m of total volume and $2.37m of liquidity, the deepest on the board — prices no change at 48.5%, down four points on the day, a 25bp increase at 47.5%, up two, and any cut at about 2.8%. A day earlier the hold side led 52.5 to 45.5 after a soft private payrolls print. So the day's move is a flip back to a toss-up, and it runs against the weekly drift, which is dovish: hold up four points on the week, hike down five.
The mechanism was growth data rather than labour weakness. Initial jobless claims for the week ended 1 August printed 199,000 against 198,000 the week before, a four-week path of 217k, 209k, 189k, 198k and 199k — layoffs are not happening. Second-quarter nonfarm productivity rose 0.35% on the quarter, roughly 1.4% annualised, up from about 0.8% in the first quarter, while unit labour costs rose 0.33%, essentially the same pace as the quarter before. Productivity accelerating with unit labour costs flat is disinflationary from the wage side. And the Atlanta Fed's third-quarter nowcast came in at 5.83% against 1.54% for the second quarter. An economy running that hot gives a chair focused on inflation credibility room to tighten without breaking much, which is what pulled the hike leg back up.
Positioning agrees. The two-year Treasury at 4.20% sits 57bp above effective fed funds at 3.63% — the front end is priced for hikes, not cuts. The largest contract on the board, on the number of 2026 cuts, prices zero cuts at 87.6% on $46.9m of volume. The separate "rate hike in 2026" market sits at 63.5%, with October priced above September at 56.5% versus 47.5%: the hike is being pushed back, not priced out. Meanwhile the 10-year breakeven is flat at 2.22%, so the hawkishness is about credibility rather than expectations.
Against all of that, Chan Yang's call on Excess Returns is that the Fed does not hike this year — and he concedes that on a normal reaction function it should. His reasoning is political: the chair is spending political capital on structural reform, unwinding the excess-reserve regime, rolling back Basel III and shifting credit creation toward regional banks, and will not burn it on 25bp against a president who does not want it. The weakness in that argument is that it rests on an inferred private understanding, asserted rather than evidenced, and it arrives in an episode that ends with a pitch for his own fund.
The two thin contracts resolving tomorrow point in opposite directions, which is the most interesting thing on the board. On July jobs added, the 100–150k bucket lost nearly ten points to 19.9% while 0–50k gained four to 26% and the below-zero bucket ticked up — the whole distribution shifted down. On the July unemployment rate, the move went the other way: 4.1% gained five points to 27%. Weak hiring with a falling unemployment rate implies a shrinking labour force, not a strong one. Both contracts carry only $27k and $41k of volume, so treat them as colour rather than signal, but the combination is the one print that would let the hawks and the doves each claim tomorrow.
4The market believes a Hormuz deal eventually, just not one that moves tankers this month
The ceasefire contract faded hard on the day after a huge week: a US–Iran effective ceasefire by 14 August prices 81.5%, down 4.5 points, and by 31 August 85.5%, also down 4.5 on the day but still up 31 points on the week, on $13.4m of volume and $493k in the last 24 hours. Near-term throughput deteriorated alongside it. The contract on average ships transiting the Strait of Hormuz at end-August moved to 43.5% for the 0–20 bucket, up six points, with 20–40 down 4.5 to 32%; the weekly contract for the week of 3 August moved the same way, the 50–74 bucket down eight points to 28%.
Long-dated normalisation did not follow. Hormuz traffic returning to normal by 31 December held at 62.5%, up a point on the day and eleven on the week. That is the day's cleanest contradiction and it resolves: the pricing is not doubting that a deal eventually happens, it is doubting that a deal moves tankers this month. Reuters supplied the mechanism in the afternoon — shipping sources describe the proposed passage arrangement as "not feasible for the shipping industry", Iran warned Gulf states to tell the president to desist, and Houthi forces attacked a Saudi tanker.
The operational detail is worse than the headlines. Ryan Peterson of Flexport, on Prof G Markets, says container shipping through the strait is effectively closed at one to two transits a day against a peak near 100, that tankers are down but running higher than boxes, and that war-risk insurance now costs 10–15% of a ship's value per voyage. The Red Sea has been shut since December 2023, so almost everything already routes around Africa. The tape reflected the reversal: the Dow closed at a record the day before on Hormuz optimism and gave back 0.83%. On this evidence, crude strength is a supply-risk move rather than a demand signal.
5A 40-year link between gold and real rates has stopped working
Felix, on Felix & Friends, counts eight institutional houses turning bullish on gold inside about two weeks and names seven of them: State Street with a 4,700–5,500 base case and a 6,250 bull case; Deutsche Bank at 4,600 by year-end, describing gold as in an explosive phase since August 2024; Goldman Sachs advising clients to "start nibbling" with a floor near 4,000; UBS at 5,000 by next March; BNP Paribas rebuilding longs and flagging squeeze risk; ANZ shifting to accumulation; and Jefferies, whose framing is the load-bearing one — a "structural decoupling" in which the 40-year rule that gold falls when real rates rise has broken. Real rates have risen; gold has not fallen.
The supporting legs are the sovereign bid and supply. Central banks bought a record of roughly $45bn of gold in a quarter, and China imported in a single month the equivalent of what the ten largest producing countries mine combined, while a new mine takes 10 to 15 years to bring online, so supply cannot answer a demand shock. His stated cause is the AI debt stack from the first section — roughly three quarters of a trillion dollars a year of cloud capex, increasingly debt-funded and partly off balance sheet, with total AI-related obligations he puts near $3tn.
This one carries the heaviest discount of the day's material. The bank targets are shown on screen but not cited to dated notes, the Lehman comparison is asserted, and the presentation is wrapped in a paid-seminar funnel pitched at least five times, with no ticker, vehicle, entry level or stop anywhere in it — the entire actionable payload is a free weekend seminar. On the other side of the ledger, he discloses taking no sponsors and declining a gold-miner board seat. The claim itself is checkable regardless of the wrapper: if real rates rise again and gold falls with them, the decoupling is over.
6Luxury's growth moved from goods to experiences, and the goods side lost about 70 million customers
The final episode of Equity Mates' ten-part decade series makes the split quantitative. Luxury spend is roughly $1.6tn and forecast at $2.5–3.1tn over the decade, but since 2023 experiences have been the only source of growth: in 2025 experiences grew 3% while personal goods fell 2%, with cruises up 10% and fine dining up 8%. The customer base for personal luxury goods shrank from roughly 400m people to 330m. What is left is concentrated — Bain puts customers spending over €20k a year at 46% of the personal-goods market, up from 30% in 2019, and BCG has 0.1% of luxury consumers accounting for 37% of spend. The ultra-wealthy are price-insensitive; the aspirational buyer who drove the last decade is macro-sensitive, and has gone.
LVMH is the case in one company: 16 consecutive quarters of decelerating sales growth, from +44% year on year in December 2021 to −5% in December 2025, 2025 revenue of €80.8bn down 5% and profit of €10.9bn down 13%, the shares off 25% year to date and more than 50% from the high. Geography splits the same way — China fell as much as 20% in 2024, the steepest since 2011, while the US is the bright spot on AI-driven wealth creation, with Hermès up 17%, Prada up 15% and Cartier up 18% there.
Two things to weigh against it. The hosts state the risks themselves: Veblen pricing power has broken, with Hermès handbags up 50–70% since 2019 while category margins shrank, and a wealth-tax backlash is building, with California weighing a 5% one-off levy and the "tax wealth not work" argument gaining ground in the UK. And the packaged routes into the theme have a poor record of their own: three luxury ETFs have been liquidated in 18 months — Roundhill's LUXX closed in 2024, Tema's LUX in August 2025, KraneShares' KLXY in March 2026. Whatever the thesis is worth, the vehicles for it keep failing to gather assets.
What the sources recommended
Nineteen named calls came out of five videos, and the raw count overstates both sides: nine of the ten buys are one thematic survey's watchlist, and four of the nine sells are a short book relayed from a note that the same episode's other guest spends the segment rebutting.
The buys. is Chan Yang's own product — the Variant Perception Cycle Aware US Equity ETF, which he manages, pitched at the end of the Excess Returns episode as a satellite sizing of $5 against $95 in , not as a single-name conviction call. It listed in March 2026, holds about $35m across 140 positions, and is currently overweight energy and semiconductors after being heavily overweight tech in March and underweight tech by May. The other nine are the Equity Mates luxury names, and they arrive with their own disclaimer attached: one host says outright he owns no luxury companies and the other has Viking on a watchlist only, so these are the named ways to express the theme rather than endorsements. On the experiences side, , and — the two Madison Square Garden names reportedly up more than 100% in twelve months — plus on $25bn of revenue through Ticketmaster and VIP packages, up 40% year to date in a cruise category growing 10%, and Ten Lifestyle Group, a London-listed concierge business for the ultra-wealthy, which is a name-level call rather than a ticker-level one — no symbol for it resolves on the market-data sources checked. On the goods side, the stated survivors: on €16bn of revenue up 9% at a 41% recurring operating margin, on 13,000 cars, revenue up 7% and an order book through end-2027, and up 5% year to date on jewellery holding where leather goods have not.
The sells. is the day's highest-conviction single call: Ross Givens notes the first lockup tranche released at 9:30 ET, 911m shares against the 555m issued in the June IPO, so tradable supply nearly tripled overnight, with further tranches this month, another after October earnings and full expiry on 8 December. His level is $58 by year-end against about $110 now, and he is explicit that he likes the business — he calls it the most important company in the world — so this is a supply argument, not a thesis one. It is also the second consecutive day of SpaceX bearishness from a different channel, after Morningstar's $62 fair value carried in yesterday's digest. Next, , , and are Burry's relayed short book and should be read as one position rather than four, with the semiconductor-index leg carrying no ticker. is Eisman's, on the expense growth and depreciation above. And , (down 16% year to date) and (down 23%) are the goods side of the luxury split.
The labour-market question gets one print tomorrow
July payrolls and the July unemployment rate both release tomorrow, and together they are the direct test of the rate section above: a soft jobs number arriving with a lower unemployment rate is the shrinking-labour-force reading, and it is the single outcome that leaves both the hike case and the hold case standing. July CPI follows on 12 August, where the modal path prices 3.4% at 44.5% — and note that the separate contract on 2026 inflation has already resolved yes above 4%, which corroborates the Excess Returns claim that headline is running north of 4% while core is not. Further out, reports around 26 August, the first hard read on whether the order book behind the circularity argument is still growing, and the next SpaceX lockup tranches land later this month, so the supply argument gets tested well before it gets settled.
