1Capex is compounding into memory prices, and in one session about $500bn moved from the AI spenders to the memory makers
The most concrete number of the day came from Animal Spirits 471, on The Compound. In a single session the memory complex — Micron , SanDisk , Western Digital , SK Hynix and Samsung — added roughly $500bn of market value, while , , , and lost roughly the same amount between them. Those are not two unrelated moves. They are one mechanism seen from both ends: memory has become an input cost for the companies building the data centres, so the price that pays one side is the price that squeezes the other.
The cost side is already visible in a share price. fell 6% on device price rises the show traced back to expensive memory. The capex side is visible in a forecast rather than a print: hyperscaler free cash flow — , , and among them — is modelled on the show as falling from roughly $700bn to about zero over twelve months as capital spending compounds. is up about 1,400% since January 2025 and now carries a $1trn market capitalisation. The hosts decline to call any of that a bubble: their stated bar is a 50% drawdown in the Nasdaq 100, and they describe at 24 times forward earnings as "cheap-ish".
Ross Givens reads the same complex the other way round. , and have doubled or tripled off their April lows, which to him makes them the extended side of the market rather than the paid one, and he calls leadership out of semiconductors and into biotech, with the NBI biotech index breaking out of a four-to-five-month base. So one show has memory as the rational trade and another has it as the crowded one, on the same week's tape. Neither can talk the other out of it: the split resolves only if memory pricing keeps turning up in device makers' costs, or stops.
2Margin debt is up 54% in a year, and the risk it describes shows up in single names, not the index
FINRA margin debt is up 54% year on year, per the same Animal Spirits episode — the fourth time it has risen more than 50% in a year since 1997, the previous instances being 2000 and around 2021. The hosts are explicit that this is not a call for a generational top. Their worry is narrower and better evidenced: air pockets in individual names. fell 27% in three days and moved more than 5% a day for seven consecutive sessions, and it did that with the memory story intact.
The series also understates the leverage it measures. It does not capture swaps, single-stock and 2x/3x leveraged ETFs, or prime-broker lending, all of which have grown alongside it. Retail leverage manias in Taiwan and South Korea were flagged on the same show as the sharper version of the same behaviour.
The counterweight is breadth, and it is broadening rather than narrowing. The Russell 2000 is up 22.3% year to date against the S&P 500's 7.5%, 67% of the smaller index is higher on the year, and about 900 of its members are up more than 20%, with transports, industrials, small caps and regional banks leading. Michael Batnick disclosed buying and for himself the previous day, on a stated five-year hold. Put the two halves together and you get a market where a levered single position can lose a quarter of its value inside three sessions while the index barely moves — which is a different risk from the one an index drawdown chart would ever show you.
3The July 4 gold revaluation rumour fails on mechanics, and gold is sitting at a seven-month low anyway
The rumour ran like this: the US revalues the 8,000 tonnes of gold it still carries at a book value of $42 an ounce up to the market price, does it on July 4, and gold goes to $10,000. Felix, on Felix & Friends, takes it apart — and he is a long-term gold bull doing it, which is what makes the debunk worth reading. There is a real kernel: the $42 book value is genuine, and Judy Shelton's Treasury-Trust-Bond proposal is a real idea in circulation. But a revaluation would need the Treasury, the Fed and legislation to move together, it could backfire, and he finds no pre-positioning in the flows ahead of the date. His verdict is that it is not happening as sold.
His structural case is separate and survives the debunk. Central banks are buying roughly 1,000 tonnes a year, about twice the normal rate, with much of it undisclosed on Goldman's read; US federal debt is $40trn; and de-dollarization has been running since Russia's reserves were frozen. He calls the current drawdown of about 28% healthy and says corrections of 30-50% should be expected in an asset like this. He is not in the miners at the moment, on the view that money is flowing elsewhere.
Animal Spirits attacks the same thesis from the driver end. Its debasement segment argues gold's rise is more a price effect than central-bank accumulation, and that the dollar is not going anywhere while it sits on 89% of foreign-exchange volume. Meanwhile the metal printed a seven-month low into firm yields and building rate-hike bets. The rumour is settled; the driver is not, and the driver is what decides whether a seven-month low is a dip inside a structural bull or the front of something worse.
4Code is not the moat once code is cheap — the durable asset is proprietary data and a closed loop
Ritavan, author of The System Gambit, on Excess Returns, argues for throwing away the moat checklist and asking one question instead: is this business a self-improving, path-dependent, hard-to-copy loop that has crossed into a new paradigm, usually by sacrificing a short-term metric to get there? His case studies are illustrative rather than calls, and no entry or target is attached to any of them. is the strongest version — warehouses, then a data-driven storefront, then FBA as a platform, then Prime. executed a late gambit of its own on physical-shopper data that Amazon cannot replicate. 's moat is institutional: engineers embedded in customers and guaranteed fab output. is the negative example, agile without a causal model of its own business. re-rated on an AI-lab data-licensing deal, and gets cited for the forward-deployed-engineer model.
The AI-era half of the argument leans on Sparkline's work: a K-shaped economy in which AI has accounted for close to all of the market's returns over the trailing period, producing shoot-first dispersion in which many of the names sold off go on to recover. Once code is cheap, code is not the moat — the value sits in signal quality, in closing the loop, and in complementary assets and proprietary data. The consequence he draws is uncomfortable rather than actionable: the software names falling on the day are a mix of future zeros and generational buys, and an AI adoption announcement does not separate them, because on this framing adoption on its own carries no information at all.
What the sources recommended
Ten buys and one named sell, and nine of the ten buys came out of two videos totalling eighteen minutes between them. No ticker appeared in more than one channel — each source ran a separate thesis.
The edge-AI seven. Business with Brian's twelve-minute video argues that on-device inference is the next leg after data-centre AI, with edge chips forecast to grow more than 30% a year, and treats the pullback as the entry: (Calix, ISP-gateway software, gross margin around 57%, recurring revenue), (Ambarella, pure-play edge vision, roughly 80% of revenue from edge AI, a ten-year Hanwha deal), (Broadcom, custom ASICs at about 60% share plus a Wi-Fi 8 neural chip), (the Jetson-Thor robot brain and the Isaac and Groot stack), (Mobileye, its IQ silicon in 230m vehicles, with the ownership overhang named as the risk in the same breath), (Cohu, chip-test handlers, orders up 57% year on year) and (Renesas, automotive microcontrollers on 3nm — flagged in the video itself as US over-the-counter and hard to buy). All seven are weeks-to-months calls from one presenter in one sitting, in a video carrying a disclosed sponsor and a subscription pitch, so the confidence attached to them is his alone and nothing here corroborates it.
The two biotech breakouts. Ross Givens takes (Immunovant) on a breakout around $39.40, stop at the base low for about 5-5.5% of risk, with a 25-30% run as the stated objective; and (TG Therapeutics), up 55% in June and coiling for three days, stop near $78. Both are chart calls with no fundamental catalyst attached, which is how he presented them. He name-checked , and as already-running biotechs without calling any of them.
The news-sourced pair. The CNBC Investing Club's "sell Goldman, buy Capital One" supplies both and , and a second exit was described only as a troubled stock after a muted quarter, with no ticker given. Both sells came off the wire rather than out of the five videos. One bearish observation stopped short of becoming a call: Animal Spirits called Domino's horrendous — a resigned CFO, share lost to chicken chains and DoorDash — but framed it as commentary rather than an exit.
Hormuz, not the earnings calendar, is what can move a holiday-shortened tape
US–Iran technical talks are the live swing factor, and the specific outcomes that matter are control of the Strait of Hormuz and whether shipping restarts; oil goes directional on the answer, with EIA crude inventories landing into the same window. World stocks paused as attention turned to Kevin Warsh, per Reuters, with rate-hike bets firming — the same force that has gold at its seven-month low. The earnings slate is thin and unrelated to any of the above: FactSet , MSC Industrial and UniFirst before the open, National Beverage , Greenbrier and Franklin Covey after it, and Lindsay tomorrow morning.
The calendar compresses all of this. US markets are closed on Friday 3 July for the July 4 holiday and Thursday's session may be shortened, so two thin sessions carry the positioning into a long weekend — on the back of a first half that Reuters wraps as the best quarter since 2020 despite the Iran war, and June global PMIs that printed strong through war-driven cost pressure. A transfer as large as the one in the first section either continues through that thin tape or unwinds in it, and thin tape is the worst possible conditions in which to tell those two apart.
