12 posts from 8 reports.
What's new: Goldman's strategists tested what the market prices into AI chip stocks. Micron, SK Hynix and Samsung trade at 4 times two-year forward earnings, half their 15-year average. Nvidia, Broadcom and AMD trade at 13 times against a long-run median of 18.
Why it matters: a low multiple on peak profits is a warning more than a bargain. Memory gross margins near 80% are more than double their history. Earnings would need to fall about 50% for the P/E to reach its average. Past memory cycles saw bigger falls, even losses. Fabless margins sit only modestly above average, so a 10% dip would normalise their P/E. GS sees hyperscaler capex growth slowing from nearly 100% this year to 54% in 2027.
Prefer fabless designers Nvidia and Broadcom (Buy at Goldman) over Micron (Neutral), SK Hynix and Samsung into the 2027 capex slowdown.
SkepticFour times earnings. Our 12 September digest had Christopher Wood buying SK Hynix and Samsung on a supercycle. How much worse can it get?
BullPlenty, on history. Margins are double normal, so profits can halve and you're only back at an average multiple. Memory has printed losses in past downturns. Nvidia and Broadcom only need a 10% dip.
SkepticFair. Then 2028 is the test. GS says memory pricing is too bearish if 2028 forecasts hold. If those estimates survive the 2027 slowdown, rotate back.
What's new: Goldman compares the market value each AI infrastructure group has added since 2023 with the EBITDA it has added. A high ratio means investors expect today's profits to last for years. The ratio is highest for optical networking, electrical grid and liquid cooling stocks. It is lowest for semiconductors and utilities, whose valuations already assume earnings fade.
Why it matters: GS analysts see hyperscaler capex reaching $1.2trn in 2027, above consensus. Growth still slows sharply. For most suppliers, slower capex growth means slower earnings growth. Shortages of power, labour or memory could also cap spending, which would hit supplier earnings first. Our read: the groups priced for the longest boom have the least cushion if the slowdown arrives on schedule.
Trim optical networking, grid and liquid cooling winners before AI capex growth slows in 2027. Shift that exposure to semiconductors, which already price fading earnings.
SkepticOur 14 September digest said optical demand is real and supply is the only limit. Scarcity is pricing power. Why fade the bottleneck?
BullBecause the scarcity is in the price. Optical, grid and cooling carry the highest ratio of added market value to added EBITDA since 2023. Semis and utilities sit at the bottom, trading as if the cycle already turned.
SkepticThen the kill criterion is capex revisions. Consensus 2027 capex rose $175bn in one earnings season. Another upgrade that size and the long-duration names keep working.
What's new: Goldman's agentic commerce deep dive splits ticketing in two. Primary sellers like Ticketmaster hold exclusive inventory through contracts with artists and venues. Resale sites like StubHub, SeatGeek and Vivid Seats list the same seats as each other and buy traffic through paid search. Amphitheaters sell 60-70% of their seats and theaters 65-75%. Ticketmaster already works with Meta's Muse.
Why it matters: an agent that knows a user's city, dates and taste can surface shows the user never searched for. That lifts sell-through without adding a single concert. On resale sites, the same agent compares identical seats on all-in price, which squeezes service fees. Agents also skip the checkout pages where parking, insurance and upgrades are sold.
Own Live Nation (Buy at Goldman, $210 target, 24% upside) as ticketing's main winner from AI agents. Prefer it over resale platforms such as StubHub.
SkepticGS rates StubHub Buy at $5.31 with a $16 target. Agents could cut its paid-search bill. Why pass on 200% upside?
BullCheap traffic lasts until the agent platforms start charging. GS expects OpenAI, Google and Meta to gain pricing power and charge placement fees. Live Nation owns the inventory every agent needs.
SkepticSo the tell for StubHub is its customer acquisition cost. If agent traffic stays cheaper than Google search for a few quarters, revisit it. Until then, own the primary.
What's new: Meta's Muse launched payments on Stripe's Link, which charges the shopper's card and then issues a one-time virtual card to the merchant. Goldman's conclusion is that cards are now the default platform for agent payments. Visa's tokens, which replace card numbers with secure codes, cover 54% of its online transactions, up from 4% in 2019.
Why it matters: the networks charge a fixed fee per transaction. Agents buying single items across more merchants lift network revenue per dollar spent. Fraud also gets harder to catch, because genuine agents run from shared cloud servers at machine speed. That raises demand for the networks' paid security services. Who carries agent fraud losses is still undecided. Merchants lose 90% of disputes today.
Own Visa and Mastercard (Buy at Goldman, 21% and 25% upside) as the default rails for agent payments. Mastercard is the pick for fraud and identity services.
SkepticStripe runs the payments inside Muse. Why isn't the new processor the winner?
BullStripe charges one card and issues another. Both legs ride the networks. GS says a 5% pickup in cyber services adds 20 bps to Visa's net revenue growth and 80 bps to Mastercard's.
SkepticThat gap is why Mastercard is the pick. The kill criterion is liability. If banks and merchants refuse to carry agent fraud, they add friction and adoption stalls.
What's new: Goldman expects agents to run the whole purchase for toilet paper, floor cleaner and packaged food, reordering from past history. If a product is out of stock, the agent may pick another brand or a store brand. Agents also show far fewer products than a search page. Mondelez found Oreos in only 10% of chatbot cookie answers, likely because it blocked AI crawlers.
Why it matters: loyalty in low-involvement aisles is mostly habit. An agent optimizes for price, speed and availability, which favors store brands and scaled retailers like Walmart. Brands now rewrite product pages so machines can read them, a shift Goldman calls generative engine optimization. Prestige beauty stays a human choice, with the agent suggesting and the shopper deciding.
Avoid staples most exposed to store-brand substitution, such as Clorox and Kraft Heinz. Prefer high-involvement brands like Estee Lauder (Buy at Goldman).
SkepticMondelez is already fixing Oreo's visibility. Clorox can do the same. Isn't this a one-off search fix?
BullBeing readable gets you on the list. It doesn't make the agent pay up for bleach. The risk is substitution: item out of stock, agent buys the store brand. Brand premiums in commodity aisles shrink.
SkepticThen it's slow money. GS frames this as a three-to-five-year shift. The kill criterion is adoption. If automatic reordering stays niche through 2027, the private-label risk stays on paper.
What's new: Goldman sized what the six US hyperscalers must earn on $1.73trn of AI capex planned for 2026-27. A 15% return needs $1.42trn of revenue over 2028-30. For the three public clouds, the hurdle is $1.0trn against a combined backlog of $1.69trn, up 152% in a year. Microsoft needs about 40% of its $678bn backlog, Amazon 68% of $514bn and Alphabet 78% of $496bn.
Why it matters: the companies expect to book about half of their backlog as revenue within 24 months. Our read: the later half of Microsoft's backlog, about $339bn, still covers its $268bn hurdle. Alphabet's later half, about $248bn, falls short of $385bn. Alphabet's consensus two-year capex doubled this year to $507bn, the biggest jump of the five.
Prefer Microsoft among the cloud giants for visibility on AI returns, with Amazon second and Alphabet third. Alphabet's third-quarter backlog is the next test.
SkepticBacklog is the wrong yardstick for Alphabet. It shifts capacity between Search, YouTube, DeepMind and Cloud as it likes.
BullFair, and that's the point. Microsoft's hurdle sits in signed contracts. Alphabet's leans on products it still has to win. Amazon at least says its servers break even in about three years on a five-to-six-year life.
SkepticThen the kill criterion is new bookings. GS assumes zero backlog growth. If Alphabet keeps adding double digits a quarter, its 78% falls fast and the ranking flips.
What's new: Goldman's chip analysts built AI capex from supplier guidance. They get $1.3trn in 2027 and $2.0trn in 2028, about 57 new gigawatts that year. Broadcom guides to 20 gigawatts for its customers in fiscal 2028. GS has Nvidia at 22. GS models MediaTek's data-centre chip sales rising from $2.2bn this year to $52.5bn in 2028.
Why it matters: in GS's build, a gigawatt of Nvidia systems costs about $50bn of capex and a custom-chip gigawatt about $25bn. The blended cost falls from $42bn to $35bn by 2028. Cheaper gigawatts lower the hurdle: at $34bn each, a 15% return needs $9.4bn of yearly revenue per gigawatt instead of $11.6bn. Nvidia still gains, with content per gigawatt rising from about $25bn on Blackwell to $40bn on Rubin.
Own Broadcom (Buy at Goldman) for the fastest gigawatt growth and put MediaTek on the research list. Keep Nvidia as the core holding.
SkepticSupplier guidance runs hot when capacity is short. GS says so itself. Why trust 20 gigawatts from Broadcom?
BullHaircut it by half and Broadcom still doubles its gigawatts in two years. Every hyperscaler wants a cheaper gigawatt. My read: MediaTek at $52.5bn says custom silicon is spreading beyond Broadcom.
SkepticThen the kill criterion is land and power. GS flags shells and power from 2028 as the main supply risk. If gigawatts slip, the custom-chip ramp slips with them.
What's new: at Connect, Meta gave Muse computer use and new shopping and payment links, plus new glasses and a pocket device. Goldman kept its Buy and its $725 target, 2.6% below the $744 price. GS models capex doubling to $138bn this year and reaching $214bn in 2027. Free cash flow turns to minus $26bn in 2027, with no buybacks.
Why it matters: in GS's return framework, servers are written off over five years, so this spending weighs on profits well past 2027. Gross margin falls from 82% in 2025 to under 70% by 2027. Goldman's return framework says Meta's $309bn of 2026-27 AI capex needs about $84bn of new revenue a year in 2028-30 to earn 15%. Meta has no cloud backlog to point to, unlike the three public clouds.
Trim Meta above $725, Goldman's own target, despite its Buy rating. The third-quarter report is the next test.
SkepticGS is Buy-rated. It calls Connect more proof that Meta is turning compute into products people use.
BullThe product works. The model doesn't pay for it yet. Twenty-two on the top line in 2027, minus one on EPS. The stock sits above target. That's a Buy with no upside.
SkepticThen the add-back signal is usage. GS's milestones are Muse retention and more commerce and payment partners. Show those at the third-quarter print with capex on plan and you buy back.
What's new: DB argues that growth alone explains this year's markets. Yields rose almost entirely because the neutral rate rose. Stocks sit near records and credit spreads near their tightest. Global data surprises are on track for their largest positive run on record. The driver is a global capex cycle, powered by AI and by government-funded building in defence, energy and supply chains.
Why it matters: DB finds only two big capex booms since postwar rebuilding, the late 1980s and the mid-2000s. The average top-five boom lasted three years. DB's conservative case has this one running five years and adding a cumulative 3% to GDP, with 6% in its upper case. Goldman adds that hyperscaler capex in 2027 would be a bigger share of GDP than any tech cycle since the railroads.
Do not chase the dollar's recent jump. The dollar drifted weaker through every big capex boom in DB's table.
SkepticThe Fed just hiked and the ten-year touched 5.22%, the highest since 2007. Rate gaps say own dollars.
BullDB's point is that those yields reflect growth, through a higher neutral rate. Its table shows the dollar drifting weaker through every boom on the list. My read: this boom is global, so the growth premium isn't only American.
SkepticThen the kill criterion is the capex path. If hyperscaler capex growth undershoots GS's 54% for 2027, the growth story fades and the dollar trade flips.
What's new: BofA counts about $630bn of AI-related debt issued this year. Only about $4.5bn sits in CLOs, the funds that buy leveraged loans. Most CLO deals until 2025 excluded project finance. Most 2026 deals allow it. BofA sees room for at least $100bn of data-centre and GPU loans. GPU loans pay about 540 bps over SOFR and data-centre loans about 280 bps.
Why it matters: GPU loans amortize, so the debt is repaid before the chips age out. The buyer of the compute is often a BB or investment-grade name such as Meta or Nvidia on take-or-pay terms. That is how a B+ loan trades above par. On the other side of the AI trade, 12% of software loans trade below 80. About 20% mature by 2028.
In leveraged loans, own data-centre and GPU financing paper plus loans to power and cooling suppliers. Underweight loans to general business-software makers, which average 90.4.
SkepticOracle just flagged force majeure and may defer rent on a data centre. Why lend against single projects now?
BullThat's construction and site risk. BofA says CLO managers avoid it. The CLO paper is mostly GPU contracts that amortize, with a strong offtaker. Five-forty over SOFR for that is wide.
SkepticThen cap it. BofA suggests limits around 10% per industry. The kill criterion is the offtaker. These loans are capped by its rating, so a downgrade reprices the whole chain.
What's new: imports supply 35% of US oil-country tubes, the steel pipe used to drill and line wells. All imports have paid a 50% tariff since June 2025. Antidumping cases now target a fifth of them, from Taiwan, Austria and the UAE. Seamless prices reached $2,787 a tonne in August. Tube prices overall are the highest in about two and a half years.
Why it matters: seamless margins are up about $280 a tonne this year, since prices rose $306 while scrap steel rose $25. Welded margins slipped as coil costs rose. Vallourec makes only seamless pipe, while Tenaris is about 20% welded. Offshore adds volume: BofA's tracker has committed offshore capex up 61% in 2027. The same cycle drives Saipem, BofA's top oil-services pick.
Own Vallourec, BofA's top tube pick after a double upgrade to Buy, ahead of Tenaris. Catalysts are US antidumping rulings in November and Vallourec's November 20 results.
SkepticTube stocks trade as oil beta. Brent at $105 won't last. The tube new-order index just fell seven points to 43.
BullThe drivers aren't oil. Tariffs cut supply. Offshore orders are booked into 2027-28. Vallourec booked a year of offshore work in six months. BofA has its EBITDA up 35% next year, 10% above consensus.
SkepticFair. Then the kill criterion is trade policy. If the November rulings go against the petitioners or the 50% tariff is cut, US pricing loses its floor.
What's new: Deutsche Bank's November 2025 primer counts $331bn of tokenized assets, up from about $4bn in 2019. Stablecoins are 90% of that. Real-world assets such as Treasuries and private credit total about $33bn. Tokenized money-market funds held about $2bn in 2024 against a $7trn industry. DB's conservative forecast is $1.5-2trn of tokenized real-world assets by 2030.
Why it matters: DB assumes over 80% of that stays in Treasuries, money-market funds and credit, because simple bonds are easy to put on a blockchain. Stocks and property stay small until the late 2020s. Liquidity is the gap. About $15bn of bonds have been tokenized, yet only 47% of those projects have traded over $1mn. DB expects custodians and other intermediaries to survive the shift.
Keep tokenization exposure in the incumbents issuing tokenized cash and Treasury funds, such as BlackRock and Franklin Templeton. DB does not expect deep trading before 2029-2032.
SkepticThis primer is ten months old. Our 14 September digest already said to own the rails through BlackRock and JPMorgan. What's new?
BullThe timeline. DB puts real market depth at 2029 to 2032, once settlement networks connect. Until then the product that works is tokenized cash. BlackRock's BUIDL charges 0.20-0.50% and settles faster.
SkepticThen size it as a slow compounder. The kill criterion is regulation. DB calls it the biggest barrier, and US rules are still fragmented.
