For the next 36 months, the market’s main trade is divergence: own the 20 to 30 companies with AI infrastructure, distribution, data, or workflow franchises, and avoid the 400 to 500 public companies whose economics get competed away by those same forces. The AI super-cycle is creating a chasm. By 2027, I expect $300 billion to $400 billion of annual AI infrastructure spend across chips, data centers, networking, power, and model training. The winners will compound revenue at 20% to 40% for multiple years. The laggards will argue valuation while their margins fall 200 to 500 basis points.
That is the trade.
Start with the money. In fiscal 2024, Nvidia’s data center revenue was roughly $47.5 billion. By early fiscal 2025, the quarterly data center run rate was already above $90 billion annualized. That is what a super-cycle looks like in year two. TSMC controls roughly 60% of global foundry revenue and more than 90% of the most advanced logic manufacturing. ASML ships the machines that make the most advanced chips, with EUV tools priced around $180 million to $200 million each. Broadcom has a custom silicon and networking franchise tied to the hyperscalers. These are choke points. Choke points earn rents.
Follow the capex. Microsoft spent more than $50 billion in capital expenditures in fiscal 2024. Alphabet’s 2024 capex pace moved toward the $50 billion range. Meta guided 2024 capex to roughly $35 billion to $40 billion. Amazon’s cloud and AI buildout pushed total capex toward the high tens of billions. Put the big four together and you are staring at roughly $180 billion to $220 billion of annual capital intensity in 2024, with a path to $275 billion to $325 billion by 2027 if AI demand keeps scaling. That spending does not spread evenly. It concentrates in chips, advanced packaging, optical networking, memory, cooling, power, and the cloud platforms that can fill the machines.
Markets reward concentration. From January 2023 through mid-2024, the AI franchise group drove the majority of S&P 500 market cap gains. Seven companies accounted for well over half of the index’s increase during that period. That number should make every portfolio manager uncomfortable. The old playbook said buy the index and go fishing. The new playbook says the index itself is becoming a barbell, with five to ten companies funding the returns and hundreds of companies becoming passengers.
AI-native companies share four traits. First, they own scarce infrastructure or distribution. Second, they sit inside daily workflows. Third, they have proprietary data or user intent. Fourth, they can turn AI into either revenue growth above 15% or margin expansion above 300 basis points by 2027. If a company cannot pass at least two of those four tests over the next 24 months, I want it in the avoid pile.
This is where the split gets brutal.
A software company that charges $100 per seat and adds a credible AI agent can take price to $120 or $150 by 2026 if it saves a customer 5 to 10 hours per employee per month. A software company with no data advantage will see open-source models and copilots compress that same seat price by 10% to 30% over the same window. The first company gets multiple expansion. The second company gets a value investor conference.
Call centers are the cleanest example. The U.S. has roughly 2.8 million customer service representatives. Fully loaded cost can run $45,000 to $65,000 per employee per year. If AI agents handle even 20% of tier-one volume by 2027, that is $25 billion to $35 billion of labor value shifting somewhere. It will not shift to every software vendor. It will accrue to the platforms that own the workflow, the data, and the channel.
Search, advertising, cloud, payments, security, and enterprise systems have similar math. If AI increases conversion in digital ads by 5% by 2026, the largest platforms can capture billions in incremental operating income. If AI reduces fraud losses in payments by 10 basis points across trillions of volume by 2027, the right networks and processors gain real dollars. If AI cuts security analyst workload by 30% in a market short hundreds of thousands of workers, the winners take budget from weaker point solutions.
The losers are easier to spot than the market wants to admit. Labor-heavy service companies with 25% to 40% of revenue tied to headcount face margin pressure by 2027. Generic application software with sub-10% growth and no proprietary data faces pricing pressure by 2026. Consumer internet companies without owned distribution will pay more for traffic as AI changes discovery. Retailers with thin 3% to 5% operating margins and no automation plan have little room for error. Banks and insurers with 1970s core systems will spend billions just to keep pace.
Valuation matters, but franchise comes first. I would rather pay 28 times 2026 free cash flow for a company compounding cash flow at 25% than pay 12 times for a company whose free cash flow falls 5% a year through 2028. Cheap can get cheaper for 12 quarters. Quality with a widening moat can look expensive for 12 quarters and still triple over five years.
The portfolio test is simple. For every public company you own, write down three numbers by the end of this week: AI-driven revenue as a percent of 2027 sales, AI-driven margin change by 2027, and capital required to stay relevant through 2028. If the first number is below 10%, the second is below 300 basis points, and the third is rising faster than revenue, sell it or size it like a mistake.
The divergence trade demands a decision. Own the franchises. Fund them by selling the pretenders. Do it before the next 24 months make the gap obvious to everyone.