14 posts from 10 reports.
The five are Oracle, Microsoft, Amazon, Meta and Alphabet. Their capex was $97bn in 2020 and $416bn in 2025. The 2026 figures mix guidance and Street estimates.
The AI build now runs on outside money: over $220bn of hyperscaler bonds this year, plus leases and joint ventures kept off the books.
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR set up the $500bn chip facility in August. NVIDIA guarantees 25% of the chips' residual value. Google and Broadcom struck a similar deal for TPUs.
Part of NVIDIA's sales now leans on its own balance sheet. If used GPUs resell below the guaranteed value, NVIDIA covers the gap.
UBT251 acts on three hormone receptors (GLP-1, GIP and glucagon). The 19.7% came from a Chinese Phase 2 and the 18% from Lilly's global Phase 2, so treat the gap as indicative.
Novo now owns a drug that matched Lilly's retatrutide in an early trial, bought from China. JPM says the market will not value it until global data arrive in 1Q27.
The $3.7trn assumes a 10% return. The build equals 3.6% of US GDP a year, while the railroad boom ran at 2.2%. If chips last three years instead of six, the revenue need rises to $6trn.
That takes about 80% revenue growth every year until 2032. The math holds only if GPU rental prices stay near today's levels while 183GW comes online.
The tally is the Wall Street Journal's, cited in the Brookings paper: $904bn of future leases and $1.52trn of purchase commitments, mostly chips. Hyperion's $27bn of debt pays at least 100 bps more than Meta's own bonds would.
Reported leverage understates what Big Tech has promised to pay. Meta's $30bn Hyperion campus shows the template: about 90% debt, none of it on Meta's books.
Europe's dividend yield premium over the US also sits near its highs. About half of European sectors return more than 4% a year in total.
European companies now hand back more cash through buybacks than US ones, relative to their size. It is part of Morgan Stanley's case that Europe's valuation discount to the US keeps narrowing.
Morgan Stanley expects data-centre copper demand to rise from 720kt in 2026 to 1.15Mt in 2028. It sees deficits in 2027 and 2028. Newer data-centre forecasts could lift that 2028 figure by about 70%.
AI has become a real copper buyer while mines add nothing. About 75% of data-centre copper goes into power equipment, where it is hard to replace.
The reserve holds 284.6mn barrels. Draws of about 30mn every four weeks reach the roughly 70mn operating floor by March 2027. Diesel trades about $100 a barrel above Brent.
Washington's buffer against the Hormuz shock is running out. Our read: fuel-export curbs are the next lever. Morgan Stanley ranks Valero and Delek most exposed to them.
What's new: Morgan Stanley's defence coverage booked €87bn of orders in 1H26. Yet in the week to 24 September, its prime brokerage saw funds press shorts in European aerospace and defence. German defence orders are running below last year and below target.
Why it matters: orders are seasonal. Morgan Stanley has seen the sector rally from about mid-November every year since 2022 and expects a marked year-end pick-up. Consensus has defence earnings growing faster than any sector but semis to 2028. Upgrades now run slightly ahead of the European average.
Buy BAE Systems, Rheinmetall, Leonardo, Saab and Indra, Morgan Stanley's preferred names, before mid-November. Year-end German orders are the catalyst.
SkepticGerman orders are behind target. Morgan Stanley itself says fiscal worries and politics are holding back a re-rating.
BullThat's why the shorts are early. Year-end orders lifted the sector four years running. France adds 13% to its defence budget this year and 11% next.
SkepticThen Germany is the kill criterion. If orders don't pick up by December, the seasonal case fails.
What's new: FTSE and STOXX moved Greece to developed status on 21 September. MSCI, with the most money tracking it, follows on 31 May 2027. Morgan Stanley estimates $0.9bn of net passive buying, led by $327mn into NBG. The three dropouts face 10-20 days of passive selling.
Why it matters: Greek banks trade at a 10% P/E discount to European banks. They have only matched them for a year, after four years of outperformance. Financials make up about 71% of MSCI Greece, so the upgrade is mostly a bank trade.
Own NBG, Eurobank, Piraeus and Alpha Bank (a Morgan Stanley top pick) into the 31 May 2027 switch. Avoid Jumbo, GEK Terna and Allwyn, which miss the developed index.
SkepticPassive is the small number. Morgan Stanley's table also has EM active funds selling up to $22bn against $15bn of developed buying.
BullThat assumes every EM manager goes to zero. The May date lets them hold while developed ownership builds.
SkepticThen sell into the event. Exit by May, earlier if the 10% discount closes first.
What's new: Data Center Watch counts $130bn of projects cancelled or delayed in the first quarter alone. States are turning opposition into rules. Pennsylvania now makes new sites bring their own capacity. Massachusetts makes them supply their own clean power or pay into a ratepayer fund. In Europe, Ireland demands new renewables. Spain's draft adds hourly matching.
Why it matters: Morgan Stanley puts the US power shortfall at 57GW over 2026-28, still 33GW after on-site fixes. Grid connections take five to seven years in some regions. Europe runs two to three years behind the US on this trade, with data-centre demand rising from 78TWh to 195TWh by 2030.
Own the suppliers of on-site power, Bloom Energy, GE Vernova and Cummins, plus Europe's grid-ready generators, RWE, EDP, Fortum and PPC. Access to power now sets the pace of the build.
SkepticOur 14 September digest already covered the backlash. Isn't it priced?
BullMy read: the market priced it as a risk to builders. The new rules hand the money to whoever owns the power.
SkepticThen the kill criterion is Washington. Morgan Stanley sees no large-scale threat without federal action. A federal pause would hit power suppliers too.
What's new: China-sourced licensing deals rose from about $53bn in 2024 to $139bn in 2025. The first half of 2026 passed $100bn, up 46%. Chinese companies started 30% of global clinical trials in 2024, up from 1% in 2009.
Why it matters: Big Pharma buys de-risked drugs cheaply as its patents expire. The risk is whether Chinese data hold up in global trials. Obesity data travel well. Cancer data travel less well: ivonescimab beat a standard drug in China but was barely better in patients over 65. The typical US lung-cancer patient is about 70.
Own Kelun Biotech, whose sac-TMT just won a global Phase 3 for Merck, and Hansoh, which licensed two cancer drugs to GSK. Keep Akeso and Summit small until ivonescimab's global data in 1H27.
SkepticThe House just voted to bar Chinese trial data from US drug filings.
BullJPM doesn't see that becoming law. And sac-TMT already won a global trial for Merck, which runs 17 Phase 3s on it.
SkepticThen the kill criterion is ivonescimab's non-squamous readout in 1H27. A miss cuts the value of every China-sourced cancer deal.
What's new: Morgan Stanley counts a record share of S&P 500 stocks with a negative three-month beta. The index beat the Russell 2000 by 11% over that span. Meta is up 36% this month, its best month in 13 years. Jay Pelosky finds the average stock's link to the 10-year yield at its most negative since 1960.
Why it matters: higher yields split the market. Morgan Stanley ran the Fed's model on a lasting 100 bps rise in the extra yield on long bonds. GDP ends 1.5-2pp lower after two years. AI capex, near $1.5trn in 2027, barely reacts to rates. The average stock does.
Keep equity risk in AI capex beneficiaries and large-cap quality. Treat small caps as a rates-and-oil trade. Add them when a US-Iran deal pulls oil and yields lower.
SkepticThe US equity risk premium sits at the 3rd percentile of 25 years. The index is the expensive part.
BullAnd it's earning it. Morgan Stanley's 8,300 target for mid-2027 needs almost no multiple expansion.
SkepticThen the kill criterion is a yield spike without growth. Wilson sizes that at a 5-10% pullback.
What's new: Morgan Stanley's Terence Tsui downgraded Ericsson and cut his target to $9 from $11. He questions the rally's link to Nokia, since Ericsson has no direct AI or data-centre exposure.
Why it matters: Nokia's re-rating rests on AI networking orders. Our 14 September digest had its second-quarter order book at a record €2.8bn and network infrastructure growing 12% in the first half. Nokia also sits among the top picks in Morgan Stanley's European screen.
Pair long Nokia against short Ericsson. Morgan Stanley rates Nokia Overweight with a €14 target and just cut Ericsson to Underweight with a $9 target.
SkepticNokia already re-rated. You're buying the winner after the move.
BullThe laggard is the one priced for orders it doesn't have. Morgan Stanley cut Ericsson's target 18%.
SkepticThen the kill criterion is Ericsson booking data-centre networking orders. Until then, the spread should widen.
