1The AI trade's demand side is two loss-making firms — and some of its profit is paper
Two videos put numbers on the AI trade's weakest joints. On Prof G Markets, the bear case was sized: justifying current hyperscaler capex needs roughly $2.5trn of incremental AI revenue against about $150bn produced cumulatively so far — a fifteenfold gap, and more than all of big tech's present revenue combined. The demand behind that number is dangerously concentrated: OpenAI was 70% of Microsoft's AI sales last year, and both OpenAI and Anthropic are deeply loss-making private firms, with OpenAI estimated to have lost $21bn last year and Anthropic guiding to no free cash flow until 2028. The demand side of the entire trade is two companies that do not yet make money.
The profit side has its own soft spot, and this is where two independent sources line up. Equity Mates flagged that Alphabet's $112bn quarterly profit included roughly $98bn of investment gains on its Anthropic and SpaceX stakes — mark-ups on private holdings, not operating income. Felix, on Felix & Friends, reaches the same place through a Goldman decomposition showing headline S&P profit growth near 30% falls to about 13% once "other income" is stripped out — that other income being exactly these private-stake mark-ups held by Amazon, Alphabet and Microsoft. Two separate channels, one conclusion: a slice of the AI era's reported earnings is paper. is the clearest single example, and is the name Equity Mates flags as first to break if the trade turns, its credit-default-swap cost already at 2007 highs.
For balance: the Prof G presenter is openly, repeatedly bearish and says so, yet still counsels against selling wholesale — lower leverage, diversify, move some gains "an arm's length from the AI trade," not exit it. And his loudest single figure, that hyperscaler capex drives most of GDP growth, is asserted without a clean source and looks overstated. The $2.5trn revenue gap is the load-bearing claim and the one worth checking independently.
2The last bubble starved energy; this one is starving housing
The most distinctive thinking of the week came from an Excess Returns panel with David Rosenberg and Rich Bernstein. Bernstein's frame: bubbles are inflationary because they misallocate capital, and the sector a bubble starves is where the next decade's returns hide. The dot-com bubble starved energy of capital — which is precisely why energy led the following ten years. His claim is that this bubble is starving housing. The supporting number, from Rosenberg: roughly 50% of US business capex is now AI-related and growing around 18% in real terms, while capex outside AI is running negative year over year. Housing supply is not responding because the cost of capital is too high and homebuilder equities are performing badly, so no capital gets drawn in — the mechanism, not just the observation.
The panel's sequencing call is the part with a use. Chanyang, of Variant Perceptions, argues the hardware and bottleneck leg of the trade "is done," and the next leg is a Jevons-paradox rotation: as compute costs fall and usage rises, the profit pool broadens from the AI suppliers to the AI adopters, with the first signs likely in non-correlated sectors like healthcare, insurance and consumer staples. Rosenberg and Bernstein were left openly disagreeing on how the inflation transmits, which the hosts presented as unresolved rather than smoothed over — a mark in its favour.
3The cleanest dated catalyst of the week is a rare-earth deadline on 10 November
Ross Givens supplied the one hard, dated catalyst in the run. China suspended its October-2025 rare-earth export controls for twelve months as part of the trade truce, and that suspension expires on 10 November 2026 — about thirteen weeks out. The escalation ladder is already moving: the US added Chinese humanoid robots to a covered list in late July, Beijing threatened countermeasures, and China had already placed the two main US producers on its own export-control list in June. The bottleneck is processing, not geology — the US has the ore, but China holds around 90% of processing, and magnets are the pressure point.
His two names are (MP Materials) and (USA Rare Earth), both backed by unusually concrete government support: carries a Defense Department preferred stake, a ten-year price floor at roughly double the market at signing, and a minimum-EBITDA guarantee, with Q2 revenue up 89%; has a $1.6bn Commerce package and a magnet plant ramping toward a 1,200-tonne run-rate. The honest caveat is Givens' own — these names "trade purely on news, on rumours, on mentions by the president," and several are not yet profitable. (He named a third, smaller processor whose ticker could not be verified from the transcript; it is deliberately left out here rather than printed as a guess.)
What the sources recommended
The run carried three clean buy theses and two shorts, plus a growth-ETF sleeve that is a portfolio-construction idea more than a set of single-name calls.
The buys. The rare-earth pair and , above. A five-fund growth sleeve from Business with Brian — , , , and — chosen for low mutual overlap rather than individual performance, the genuinely useful point being an overlap audit (equal-weighting all five quietly makes Micron a bigger position than Nvidia). And from Prof G, the one mega-cap the bearish presenter is constructive on, at 22x earnings versus 31x pre-AI — "the anxiety is priced in."
The sells. (SpaceX) drew the same verdict from two independent channels: Prof G walked through capex of $18.5bn against $7.8bn of revenue, a burn that runs through the entire $86bn IPO raise in about sixteen months, and a lockup that just released 911m shares; Equity Mates reiterated a pre-IPO "overvalued" call as the float widens from 5% to 12%. , from Prof G, on an "Elon premium" the presenter thinks is dying — his argument that it should trade toward the top of the auto range implies substantial further downside.
The whole board gets marked to market on Wednesday
The Fed argument sits underneath all of this, and it resolves partly on Wednesday. Polymarket's September contract still prices a hold at 57.5% and a hike at 41.5% — the hold bid has leaked about five points over two sessions, but the market is pricing a live hike, not the cut a weak jobs print would normally imply. The reason is the split inside the July data: payrolls fell 23,000 while unemployment fell to 4.1%, and the market has taken the supply-side reading. Wednesday's July CPI is the binary — a 3.5% headline puts a September hike close to a coin-flip, while 3.3% or below with soft core should deflate the hike bid fast. The oil backdrop keeps a thumb on the inflation side: the market prices a truce that holds (a ceasefire near 94.5%) but a Strait of Hormuz that does not reopen (normalisation a coin-flip at 46.5%), so the crude risk premium survives the peace.
