13 posts from 4 reports.
What's new: JPMorgan launched broad China AI coverage this week with a model of who earns what. Hardware takes 93% of industry operating profit in 2026 and 30% by 2030. Models and apps go from losses to 54%. Total profit rises from $27bn to $243bn.
Why it matters: hardware gets paid once, when capacity is built. Models and apps get paid every time it is used. Yet models and apps are about 14% of the listed value tied to China AI, roughly their 12% share of this year's revenue. JPMorgan has that revenue share at 49% by 2030.
Own Tencent and Zhipu for the shift of China's AI profits from hardware to apps and models. The proof arrives over the next 12-24 months in paid usage and gross margin.
SkepticMy problem is timing. About three quarters of the downstream profit in their model lands in 2030. You're paying today for a 2030 number.
BullSure, but look at what's priced. Models and apps are 14% of the value and heading to about half the revenue. The market is valuing this year's mix. You don't need 2030 to land for that gap to close. You need the mix to start moving.
SkepticWho keeps the money, though? Pricing power at the model layer is unproven. That's why JPM is only Neutral on MiniMax.
BullWhich is why I'd start with distribution. Tencent has the users, the context and the payments. JPM calls those app assets largely unpriced. That's 53% to their target. If you want a pure model name, it's Zhipu. Revenue compounding 173% a year through 2028, 170% upside. I'd size that one smaller.
SkepticWhat makes you wrong?
BullGross margin after inference. JPM has it getting to 40% by 2028. If paid API revenue and that margin stall over the next year or two, the 2030 story doesn't matter. I'm out.
What's new: JPMorgan's model has China's token consumption growing about 60x from 2026 to 2030. Tokens are the chunks of text AI models read, write and bill by. Better models take on harder work. Cheaper inference brings the easy work to more customers. Enterprises go from half of usage to 69%, which makes demand stickier.
Why it matters: that usage runs on capacity someone has to build and rent out. JPMorgan models $778bn of AI investment by Chinese hyperscalers over five years, with cloud revenue up 12.5x. Alibaba is the one name in all three layers: T-Head chips, the cloud and the Qwen models.
Own Alibaba as the integrated chip, cloud and model play. JPMorgan's $210 target is about 81% above $116.
SkepticEvery "usage goes 60x" story skips the price line. Token prices keep falling. Why does revenue grow?
BullBecause volume outruns price. JPM literally calls it a Jevons effect. Cheaper intelligence makes more work worth automating, so the total bill goes up while the unit price drops.
SkepticFine. Then why Alibaba and not the chip names? Hardware earns first.
BullIt earns first, then it fades. Hardware goes from 93% of the profit pool to 30%. Alibaba gets paid on the build and on the rental. JPM has its cloud revenue compounding 62% a year through FY29.
SkepticLook at their capex line, though. Growth goes 40, 20, 10, then zero by 2030. And the return math has Alibaba earning 26.6% on new capacity when the industry does 10%.
BullThat's the right place to push. That 26.6% assumes better pricing and faster utilization than peers. If the cloud segment doesn't show it in the next few prints, the premium is gone. So is my case.
What's new: JPMorgan's framework for Chinese software comes down to one question. What can a general model not copy? Content generation, analytics and setup screens get cheaper to replicate with every model upgrade. Live workflow data, user permissions and the ability to execute a transaction stay hard.
Why it matters: that line sets the ratings. Kingsoft Office is Overweight because its WPS suite owns the documents and the workflow. Meitu is Neutral because general models now copy many of its features. Yonyou is Underweight because its enterprise assets still aren't turning into growth, profit or cash.
Own the workflow owners, like Kingsoft Office (about 40% upside). Avoid feature sellers like Meitu and names that still can't turn their assets into cash, like Yonyou (Underweight, 39% downside).
SkepticI'd push back on the premise. Cheaper models are a cost tailwind for every app company. Why is that bearish for anyone?
BullBecause the tailwind hits your competitors too. Look at supply. New App Store releases were up 84% year on year in the first quarter. JPM won't pin all of that on AI coding tools, but the timing fits. When anyone can build your feature, pricing power goes first.
SkepticSo what do you watch? Seats?
BullSeats are the trap. JPM's point is that seats can shrink while agents do more of the work. I'd watch gross profit after AI costs, and whether pricing moves to usage or transactions.
SkepticYonyou has all the enterprise records you'd want. Why underweight it?
BullOwning the data only matters if you get paid for it. On JPM's numbers that's 6% revenue growth and Rmb470mn of profit by 2028. Until the cash shows up, I'd rather own Kingsoft.
What's new: the 10-year jumped 13 basis points to 5.09% yesterday. S&P Global's September survey showed the strongest improvement in US business activity since early 2015. Fed governor Michael Barr said further policy adjustments are likely to be needed. Futures now price a 66% chance of a hike in late October, days before the midterms. Brent rose 3.4% to $102.58.
Why it matters: this is no longer only a US story. The OECD says the average G7 10-year yield has hit 4% this year. Governments locked in cheap debt years ago, so the real cost arrives as that debt rolls over.
Keep bonds short-dated and favour companies with little floating-rate debt. If the Fed hikes in October with yields still rising, cut expensive growth first.
SkepticFive-oh-nine on the tens, and the Nasdaq 100 closed at a record on Tuesday. Somebody's wrong.
BullMaybe nobody. Earnings are doing the work. Max Kettner's point in the FT is that US earnings run about 50% above the pre-Covid trend, even outside tech. And only about 10% of S&P 500 company debt is floating, so companies feel higher rates slowly.
SkepticCompanies, fine. Governments don't get that luxury. Who's on the other side of this?
BullLong-duration holders. Every year more cheap government debt rolls into 5% money. I'd keep bonds short and let someone else own the long end.
SkepticAnd on equities, where do you blink?
BullMike Wilson's line was 4.50% on the ten-year. Above that, multiples tend to compress, and we're about 60 basis points through it. If they hike in October and the long end keeps backing up, I'm cutting the expensive growth names first.
What's new: the Trump administration is looking at curbing diesel exports to bring down US pump prices. The US refines far more diesel than it burns and ships out a net 1.5mn barrels a day. It is Europe's biggest supplier, at 506,000 barrels a day in August. European diesel prices jumped on the headline. FGE NexantECA warned global diesel could reach $350 a barrel if the flow stopped.
Why it matters: a refinery can't make only diesel. Each barrel comes out as a fixed mix of petrol, jet fuel and diesel. S&P Global estimates a full ban could eventually force US refiners to cut output by about 2mn barrels a day, around 12%. Petrol could rise 25 cents a gallon.
Own refiners with spare capacity outside the US. If a real ban comes, it turns into a pair: non-US refiners against US refiners forced to cut output.
SkepticI don't see a full ban. Chris Wright has more or less said it doesn't work, and Burgum is with him.
BullAgreed. So is Lex. The realistic version is TD Cowen's: a limited quota, maybe pulling exports back toward prewar levels. Even that takes barrels out of a market that's already tight.
SkepticTight how? Crude is moving fine.
BullThe pinch is in refining. Bernstein reckons drones have hit about 90% of Russia's refining capacity. Moscow has banned fuel exports until the end of January. Europe could lose Russian and some US product at the same time, and it's bidding against Latin America for what's left.
SkepticSo what's the trade?
BullRefiners with spare capacity outside the US. If Washington actually pulls the trigger, my read is it becomes a pair: US refiners cutting output on one side, everyone else's margins on the other. The one headline that kills it is a Russia-Ukraine ceasefire. Russian refineries come back and the margin closes fast.
What's new: Sasac, the regulator that oversees China's state-owned companies, is surveying how dependent they are on Broadcom's data centre switches, according to the FT. One person put Broadcom's penetration among them as high as 90%. Sasac is also looking at whether Broadcom used its market lead to bundle other products or push large minimum orders of switch chips. The first findings could lead to informal guidance to cut usage.
Why it matters: switches move data between servers, which matters most in AI training. Many state data centres that run Nvidia chips still route their traffic through Broadcom.
Own the Chinese switch makers H3C and Ruijie Networks as state buyers move to local suppliers. Treat Broadcom's Chinese state sales as a shrinking base.
SkepticIt's a survey. Informal guidance, no ban, and it lands just as Xi arrives in Washington. I wouldn't read much into it.
BullI'd read a lot into it. Informal guidance is how Beijing moves state procurement. Nobody needs a ban when the approved vendor list does the work. From 90%, there's one direction.
SkepticThe substitutes aren't free, though. Industry people told the FT that Huawei's switches add to power bills compared with Broadcom's.
BullTrue. I think it slows the switch rather than stops it. The winners are H3C and Ruijie, with Huawei making real strides in high-end switches. The FT says the push could be worth billions of dollars.
SkepticAnd the bigger read?
BullMy read: the whole China debate is about Nvidia, and nobody's pricing the network. Switches are the piece Beijing can swap first. The trade only stalls if Broadcom stays on the procurement lists.
What's new: Nscale, the Nvidia-backed AI cloud out of London, filed for a US IPO last week. It competes with CoreWeave and Nebius, renting out Nvidia computing power. The FT found that 73% of its $33mn revenue last year, call it $24mn, came from Spring (SG), a Singapore entity that people close to the company confirmed is a ByteDance subsidiary. The 192-page filing never names ByteDance. The link sits in a loan agreement in the exhibits.
Why it matters: ByteDance used Nscale's facility in Norway to reach Nvidia chips it can't buy in China, through a gap in US export rules. That's legal. Since then Nscale has signed multibillion-dollar contracts with Microsoft and Anthropic.
Treat part of AI cloud demand as Chinese demand routed offshore. That supports Nvidia volumes now and is a policy risk for any AI cloud renting to Chinese tenants.
SkepticThe FT has the IPO pitch at up to $35bn. On $33mn of revenue that's roughly a thousand times sales, and the biggest customer is in a loan exhibit.
BullFair, and I'd want the disclosure answered on the roadshow. But nobody's paying for last year. The deal is sold on the Microsoft and Anthropic contracts, which are worth many times Nscale's whole 2025 revenue.
SkepticSo ByteDance doesn't matter?
BullFor Nscale, less and less. For the sector, a lot. My read: some Chinese demand for Nvidia chips never shows up in export data. It rents capacity abroad, the way ByteDance did in Norway. That supports Nvidia volumes while the gap stays open.
SkepticWhat happens when Washington closes it?
BullChinese tenant revenue goes to zero overnight, at Nscale and at any other cloud that has it. For Nscale the kill criterion is timing. If the rule changes before the new contracts ramp, there's nothing left to bridge to.
What's new: Muse, Meta's personal AI agent, launched two weeks ago and is already the most downloaded free app in the US. Its daily US downloads beat every other Meta app. It does tasks rather than just answering questions: groceries, restaurant and travel bookings, price tracking. The free tier gives 100M tokens a week, against $20 a month for Gemini Spark and over $30 for Grok Bot. Meta has added Shop Pay checkout and opened Muse to outside developers.
Why it matters: the rally took Meta to a $1.9tn market value. JPMorgan only moved back to Overweight on September 10.
Own Meta. JPMorgan is Overweight with an $820 target against $741 and sees Muse opening a commission business beyond ads.
SkepticTwenty-one percent in two weeks on a free app. JPM doesn't see revenue beyond heavy-user subscriptions until at least 2027. What exactly got priced?
BullA second business model. Meta wants a take rate on what Muse completes: checkouts, bookings, leads for businesses. It's the ad model applied to transactions. At 21 times JPM's 2028 earnings, that's not a heroic multiple.
SkepticAmazon has already blocked it. The big platforms won't hand over their customers.
BullEarly on, no. JPM's view is they come around once agents prove they bring extra traffic and sales. Shopify merchants are already in through Shop Pay. That's the first crack.
SkepticSo what's your sell signal?
BullEngagement after the giveaways fade. Right now they're pushing Muse with Instagram and Facebook ads, a national TV campaign and a billion free tokens for referrals. If usage rolls over when that stops, and AI spending keeps hitting earnings, 21 times stops looking cheap.
What's new: Brent was forecast to fall below $60 this year. The Gulf war sent it into triple digits instead. Futures sit near $105 after Houthi gains in Yemen and a drone strike that shut Saudi Arabia's East-West pipeline. The seven biggest Western majors plus Aramco earned $91bn in the second quarter, twice as much as a year earlier.
Why it matters: S&P Global notes that most oil companies still trade at the same price-to-cash-flow multiples as before the war. The five largest majors cut net debt by $36bn in the quarter, nearly 20%. All of them except BP have held or raised dividends and buybacks.
Own ExxonMobil, Chevron, Shell and TotalEnergies at pre-war multiples. Payouts are holding up. A deal cycle is starting.
SkepticOil and gas shares are up 40% this year against 12% for the market. I'd say the war is in the price.
BullOnly the oil price is. The multiple hasn't moved, so you're paying pre-war money for double the profits. Before the war they had announced an 11% payout cut.
SkepticBecause nobody trusts them with the cash. After 2022 much of the extra capex was just drilling and service-cost inflation.
BullFair. But the reinvestment case is real. Depletion removes a Saudi Arabia's worth of supply every two years. Oil and gas output could fall 31m barrels a day by 2040. Wood Mackenzie shows the majors' acreage awards this year at the highest since at least 2006.
SkepticSo where's the second-order trade? And what kills it?
BullDeals and drilling. Seven $1bn-plus deals in two months, with private equity, traders and Japanese buyers hunting too. My read: once budgets follow the acreage, drillers and service firms get paid. The kill switch is peace. Prices crashed in June on a tentative US-Iran deal. More tankers are already slipping through Hormuz under US escort.
What's new: The Economist re-ran its refinancing test. It measures how much budget tightening each G7 government would need to keep debt flat as a share of GDP if it rolled all its debt at today's five-year yields. France needs 3.9% of GDP, up from 3.1% a year ago. Italy needs 0.6%, even though its net debt is 129% of GDP against France's 110%. America's number more than doubled to 4.7%.
Why it matters: the gap comes from the budget before interest payments. Italy already runs a primary surplus and needs only a little more. France runs a primary deficit of nearly 3% of GDP. Its government revenue already tops 50% of GDP, which leaves little room to raise taxes.
Prefer Italian and Greek government bonds to French ones into the April 2027 presidential election.
SkepticFrench spreads are already near euro-crisis levels. Isn't that in the price?
BullSome of it. But the hole keeps getting deeper. France's fix went from 3.1% to 3.9% of GDP in a year. Britain at least wrote its adjustment into budget plans. France has no plan. Nobody will offer one before the election.
SkepticItaly at 129% of GDP as the safe leg? That's a stretch.
BullAt these yields the budget before interest matters more than the size of the debt. Italy needs 0.6%. Greece issued bonds below French yields this summer. The market is already re-sorting Europe that way.
SkepticFrench polls this far out are often wrong. Balladur and Juppé both led and then crashed out.
BullTrue. But Le Pen has polled under 30% in the first round only once in three years. Five new polls put her at 32-36%. She wins every run-off tested. On the left, Mélenchon's answer to the debt is "Burn it". I close the trade if the centre unites and run-off polls turn.
What's new: chip exports have lifted GDP growth to about 4% in Korea and 12% in Taiwan. In Taiwan the money is reaching households. Wages are growing 3% a year against a pre-pandemic trend of 1.9%. Welfare spending is set to rise 40% in 2027. President Lai Ching-te has also proposed a $300 "AI dividend" for every citizen, costing $7bn.
Why it matters: Korea's boom runs on prices. Export prices are up 57% in a year, while real fixed investment has risen just 3% in three years. Tax revenue is projected to jump 50% in 2027. Capital Economics expects about 70% of the windfall to be saved, mostly to cut the deficit.
Own Korean government bonds as Seoul saves its chip taxes. For consumer upside from the AI boom, prefer Taiwanese domestic-demand stocks to Korean ones.
SkepticThe KOSPI has more than doubled since 2023. Korean consumer names look like the obvious catch-up trade.
BullThe Bank of Korea disagrees. Each dollar of stock gains adds about one cent of spending, against roughly five in America. Wage growth is still below the old 3.5% trend. The gains sit with chip workers, under 1% of the workforce, and the memory duo's shareholders.
SkepticTaiwan is mostly one company, though. Why would its boom spread?
BullBecause TSMC builds. Its capex went from $15bn in 2019 to $41bn last year. Taiwan's real investment is up nearly 40% since late 2023. That boom runs on volume and hiring, so pay rises across the workforce.
SkepticFine. Then why Korean bonds rather than Korean stocks?
BullBecause the state keeps the cash. Seoul expects chip taxes to all but erase next year's deficit, while G7 governments need fixes of up to 4.7% of GDP. Christopher Wood made the same call on Korean bonds in our 12 September digest. The kill switch is memory prices. If they roll over, so does the windfall.
What's new: Dario Amodei called on September 12th for the industry to "slow the pace" of AI, just as Anthropic neared an IPO reportedly aimed at about $2trn. That is roughly the ten biggest previous tech listings combined. The call came after an Anthropic researcher resigned, warning AI could "kill us all" by the end of the decade. Sam Altman said OpenAI no longer plans to list this year.
Why it matters: The Economist argues a pause on Amodei's terms would not seriously hurt Anthropic. It would cut spending on training new models. It could also slow the open-source rivals that win business with cheaper models. The bigger risk is political. Eight in ten Americans back AI regulation even if it slows innovation.
Buy Anthropic's IPO if the safety scare knocks the price down. Hedge AI clouds and data-centre builders, which carry the political risk of a pause.
SkepticYou want to buy a $2trn IPO from a company that says its product might end humanity?
BullI want to buy it cheaper. There's speculation underwriters could trim the valuation or delay. Columbia's John Coffee says amending an S-1 is common and quick. Once the SEC signs off on the disclosures, securities-fraud liability is limited. A headline discount is my entry.
SkepticProduct liability is a different animal. David Sacks called it the "mother of all product-liability lawsuits".
BullThat's the real tail. I'd size for it. But run the economics of a pause. Training spend drops. The cheap open-source rivals have to slow down too. That helps the leaders' margins.
SkepticThen who loses if Washington acts? And when?
BullWhoever is building new capacity. The Economist flags pressure to freeze data-centre construction. Brad Carson expects every 2028 Democratic candidate to back a pause. My read: AI clouds carry that risk more than the labs. My kill switch is the Senate bill letting Washington seek court bans on model releases.
What's new: Aliko Dangote launched Africa's biggest IPO on September 14th. Dangote Petroleum Refinery aims to raise at least $1.6bn at a valuation of nearly $50bn. Shares start trading on the Nigerian Exchange in November. The minimum order is ten shares, pitched at 10m retail investors. Nigeria's benchmark is up 71% in dollar terms this year.
Why it matters: FTSE Russell cut Nigeria in 2023 from "frontier" to a virtually uninvestable "unclassified" status. It is now reversing that. Nigeria has removed lingering uncertainty about its currency and forced banks and pension funds to hold more capital. The earnings are real. MTN Nigeria swung from a 2024 loss to a record profit above $700m in 2025. Africa is still under 1% of global market value.
Own listed Nigerian leaders such as MTN Nigeria and Dangote's cement arm. The catalysts are FTSE Russell restoring Nigeria's index status and the Dangote refinery's November listing.
SkepticA 71% rally built on TikTok tips and $50 trades. Malawi ran the same play and is down 20% from its November peak.
BullFair on the froth. Retail got here first. The real liquidity comes from institutions. FTSE Russell restoring Nigeria's status opens that door. Citi's David Cowan expects the upswing to last a few years.
SkepticThe refinery floats $1.6bn on a $50bn valuation. That's about 3% of the company. A lot of value on a thin float.
BullThat's why I'd start with profitable names already trading. MTN Nigeria made over $700m last year. Dangote's cement arm made nearly as much. The IPO pulls new money onto the exchange. Bamboo users already traded more Nigerian stocks than American ones in the first quarter.
SkepticAny second-order angle?
BullMy read: the deal also lists Africa's largest refinery just as today's FT has Europe scrambling for diesel. The kill switch is FTSE Russell. If the reclassification slips, the institutional buyer never shows up.
