14 posts from 7 reports.
More new data centers now run their own generators on site. That is already 40% of projects under construction. A generator can power a site in one to two years. Power is no longer the main limit. The new limit is the build itself. Bernstein thinks the US has enough skilled construction workers for only about 35 GW of new capacity by 2030. Demand implied by chip supply is closer to 70 GW. And 70% of projects sit where just 30% of those workers live.
The fix is modules built in a factory. Factory work instead of on-site work cuts build time by 30 to 60%. Schneider says its lead times fell from 88 weeks to 34. These modules should grow 24% a year through 2030. That is ten points faster than the wider market. Their share of new builds rises from 40% to 57%. The work moves from small towns to big factory hubs where more workers live. The thesis breaks if the big cloud firms keep building custom sites brick by brick.
Own the electrical suppliers that ship finished data center modules: Schneider, Vertiv and Eaton.
Cutting a 30-month build in half creates about $0.9bn of value per GW. The supplier keeps roughly $0.2bn of that. The power-module market is worth about $1.9bn per GW. The IT-module market is worth about $1.8bn per GW. Once a customer buys a whole module, the supplier designs it and picks the parts. It can fill 80 to 90% of the box with its own kit.
That shifts about 3 points of share to the integrated players. Parts-only suppliers lose it. Schneider has raised its factory capacity 270% in two years. Vertiv is pushing IT pods above 6 MW, aimed at gigawatt sites. Eaton is buying its way in. Narrower players like Legrand, ABB and Siemens lose unless they integrate further. IT modules are the scarce piece. Power modules are already common. The risk is a cloud giant that never lets the supplier own the design.
The most integrated suppliers gain share as buyers switch from parts to whole modules. Schneider leads.
Chip stocks are up 67% this year, even after falling 19% from their June peak. Then Anthropic's Dario Amodei called for slowing the pace of AI. Altman, Musk and Nadella agreed, and the stocks wobbled. But Amodei wants to go from very fast to somewhat fast. He is not talking about stopping training. Chip demand is shifting toward running models, not just training them, as AI agents spread. There is already not enough computing power for the models that exist.
The real reason may be China. Amodei wants to keep the bans on selling advanced chips and equipment. He also wants to stop the theft of model files, after reports that Chinese labs trained on Claude. Bernstein thinks the labs' current revenue plans already include this kind of talk. It still likes Nvidia (target $400 versus $218), Broadcom ($575 versus $362), Applied Materials, Lam and KLA. The thesis breaks if real rules cap how much computing power training can use.
Stay with Nvidia, Broadcom and the equipment makers. Demand to run today's models already beats supply.
Indonesia mines about 60% of the world's nickel in 2026. That heads for 70% by 2030. This gives the government real control. Officials now say a comfortable price is $18,000 to $20,000 a tonne. If the price runs higher, they can cut production quotas. In April Indonesia also raised the official price of nickel ore. Some reference rates doubled. Benchmark ore jumped 98%. That raises smelter costs and holds the metal price up from below.
This supports the floor but does not open any upside. Bernstein still expects a modest surplus for the next decade. Indonesian mines keep ramping up. Cheaper LFP batteries take share from the nickel-heavy kind. Nickel demand grows 2.8% a year to 2030, below copper's 3.2%. Bernstein cuts its long-term price to $19,000. Keep the metal shortage trade in copper, not nickel.
Do not trade nickel like copper. The price is set by policy, and the market has a surplus.
Glencore, BHP, Anglo American, Vale and Boliden all mine some nickel. For each of them it barely moves the numbers. Nickel is about 8% of Vale's profits and 3% of Glencore's. Anglo is selling its nickel business to MMG. BHP's nickel mines in Western Australia are shut and idle. A review is due in February 2027.
Higher-return copper projects make a restart unlikely. Nickel spiked toward $50,000 in 2022. Stainless steel mills complained loudly. Indonesia is now capping the price to stop that happening again. So there is no shortage story here. There is also no clean stock to bet on one. Stick with metals that still have a real shortage.
Nickel is a tiny part of every big miner. Avoid it as a single-metal bet.
New AI agent launches revived an old fear. One all-purpose assistant might book rides and fill grocery carts for you. That would move discovery off these apps and cut their fees. Bernstein's basket of these stocks fell about 7% in early September. The same scare hit earlier this year. Those stocks dropped 20%, then recovered. AI chatbots still send under 1% of US marketplace traffic, even though millions use ChatGPT.
Their real strengths are not search. An agent cannot replace Uber's supply of drivers or its live routing. It cannot replace DoorDash's selection and delivery network. Subscriptions and perks keep users inside the apps. Bernstein is buying DoorDash (target $270 versus $197), Instacart ($60 versus $47) and Uber ($95 versus $71). The thesis breaks if agent checkouts start showing up in the order numbers.
Buy the drop in DoorDash, Instacart and Uber. The threat is still only theory.
Retail is spread across many sellers. So there is more room for price comparison than in rides or food. An agent can do that comparison well. eBay and Etsy mostly help people find products. That job is easy for a bot to take over. Wayfair is safer. It sells unique, visual products and runs its own delivery, which is hard to route around.
eBay also has unique stock. About 40% of its sales are used or refurbished goods. That gives it a reason to exist even if a bot handles the search. Instacart is the most exposed delivery name. It makes money from ads shown early in the shopping process. That ad money could move somewhere else. Bernstein still rates Wayfair a buy at $125. Watch whether AI-driven sales show up in results over the next year.
Favour unique inventory and delivery. Sites that only help you find things are the most exposed.
Second-quarter profits grew 32% without the one-off gains at Google and Amazon. That beat the 22% expected. Guidance from tech and industrial firms was firmer than usual. BofA raised its 2026 profit forecast from $345 to $365. But the growth is narrow. Nvidia, Google, Micron, Microsoft and Apple make a record 27% of the index's future earnings. Tech plus the Magnificent 7 is half of it.
Semiconductors alone should provide more than 60% of the index's profit growth in 2027. Their 2026 growth forecast jumped in a year from about 30% to about 100%. That is where the profits really are. The takeaway is to own the companies driving the numbers. Do not pay a high price for the whole index on a recovery that is not broad.
Own the AI profit engines, semiconductors above all, not the whole index at a high price.
BofA starts 2027 at $410, up 12% after the 33% jump in 2026. Some of the recent margin gain will not last. One-off gains sit in the 2026 number. Cloud firms' depreciation costs rise to 12% of sales in 2027, from 10%. So reported profit looks better than the actual cash. BofA sees net margin, excluding banks, at 16.3%. That is just below the consensus.
History shows strong profits often fail to lift the index once growth stays high but slows down. That is the 2027 setup. BofA's year-end target implies about 3% downside. Inflation is sticky. Cash in the system is tighter. There has been no 5% pullback in nearly six months. Energy and consumer stocks are expected to see profits fall in 2027. The thesis breaks if 2027 chip forecasts get raised again.
Do not chase the index here. Margins and cash flow are weaker than the headline profit.
Bernstein calls this a once-in-a-generation change, as big as the move from local servers to the cloud. AI agents cannot act safely on their own without trusted data, clear rules, permissions and an audit trail. Business software already holds all of that. So the winner is whichever platform becomes the trusted place to run automated work.
Oracle leads today. Its apps, database, servers and business software sit in one system. That makes safe automation simpler. The edge is strongest with customers who already run mostly Oracle. Pricing is starting to change too. Instead of charging per user, vendors will charge when the automated work actually gets done. Bernstein rates Oracle a buy at $325.
Hold Oracle for its all-in-one lead today.
SAP has the deepest grasp of business processes across finance, buying, factories and supply chains. If customers keep moving to its newest system, SAP becomes the safest place to let software run those processes on its own. The risk is delivery. Customers still have to upgrade their old systems first.
Microsoft owns the screen where work already happens: Copilot, Teams, Microsoft 365 and Dynamics. It does not need to replace the core ledger. It just needs to own where work gets started, approved and watched. Newer AI firms will grab painful edges like paying invoices, but not the core. Bernstein rates Microsoft ($660), SAP (€273) and Workday ($238) as buys.
Add SAP and Microsoft next to Oracle for the shift to automated work.
Top language models are built at scale in only two countries. Everyone else rents them. Physical robots are different. A robot has to weld and hold tight tolerances in heat and dust. A software link to a model cannot do that. In the first half of 2026, over 60% of humanoid robot shipments went to entertainment or data collection. Only 12.8% went into real manufacturing. Most of those sit in test setups.
The economics already work on paper. A robot pays for itself in 18 to 24 months. It does not quit. The motors that move it should get about 50% cheaper as production scales. Cheap surplus labour is not the counterargument it seems. Harsh factory floors lose workers every few months. As poorer countries get richer, cheap rural labour runs out. Robots stop being a rich-country toy. The place to invest is physical automation.
Look past another language-model bet toward robots and factory automation.
To train a robot you need video of humans doing physical tasks. India sells that footage to US robot firms for about $7 an hour. Some other countries sell it for even less. China does the opposite. It bans this kind of robot data from leaving the country and treats it as national infrastructure.
Bernstein compares the cheap sellers to countries that export raw ore. The buyer keeps the finished, valuable product. For investors, the lasting asset in physical AI is the private data and the hardware that collects it. It is not the model layer, which anyone can license. Owners of unique real-world data should be worth a premium.
Value moves to whoever owns real-world physical data. Treat it as a strategic asset.
Making the code and the images is getting cheap. That demo replaced about nine months of one person's work. An AI video firm cut the cost of a commercial from about 5 million rupees to 300,000. One client went from 1.7 million to 10 million AI videos in 18 months. The lasting assets were celebrity contracts and small models trained on private data. The output itself was not.
IT services face the same squeeze. A four-week expert-interview job shrank to a week and a half, with AI running 90% of the calls. Two-thirds of the cost of a consulting agent is now the senior judgment behind it. That is the opposite of coding tools, where raw computing cost dominates. Firms that still bill by the hour will adapt slowest. They lose as clients find cheaper options.
Favour firms whose value is private data and contracts, not everyday software output.
