Artificial intelligence is the most widely recognized investment trend I’ve seen in 27 years of managing money, which is why recognizing it is worth nothing now. I’d place it in Stage 3 of the lifecycle I use for emerging trends: scaling and competitive expansion, past discovery and early adoption, not yet commoditization. The question now is narrower: who captures the incremental cash flows as it scales, relative to what the market has already priced in. I’ll say where the evidence cuts against it.
The Spending Is Real and Still Rising
In the September quarter, Microsoft spent $45.2 billion on capital expenditures including finance leases and guided to more than $50 billion for December. Alphabet spent $51.4 billion, held its 2026 guidance at $195 to $205 billion and said 2027 will rise significantly. Meta spent $34.9 billion against $135 to $145 billion for the year, and Amazon spent $75.9 billion on its way to roughly $220 billion with meaningfully more expected in 2027.
The revenue AI produces for the buyers is also real, just smaller. Azure grew 48% to $33.2 billion in the quarter, Google Cloud grew 103% to $30.7 billion, AWS grew 39% to $46 billion, and Oracle’s cloud infrastructure grew 121% to $7.4 billion in its quarter ended in August. Those are extraordinary growth rates at that size, bought with cash the buyers no longer have to spare: Alphabet’s free cash flow was negative $9.1 billion for the quarter, and Amazon’s cash capex over the first nine months of the year, $172.2 billion, exceeded its operating cash flow of $113.8 billion.
Who Is Getting Paid Now
The cash is landing upstream. Nvidia’s data center revenue reached $89 billion in its July quarter, up 117%, at a 75% gross margin. The surprise of 2026 has been memory. SK hynix reported a 79.7% operating margin on record revenue, with DRAM prices up in the high-20s percent in a single quarter, so a memory maker’s operating margin now exceeds Nvidia’s gross margin.
The neoclouds show the other side. CoreWeave had $129 billion of backlog as of August, yet Goldman Sachs Research estimated in October that serving it takes roughly $60 billion of new capex, most of it debt-financed, and the stock was down 51% over the prior year when the note was written. Demand for compute and economics for the company reselling it are two different things.
Usage Is Compounding While the Price Collapses
Adoption is the strongest part of the case. Google processed over 3.2 quadrillion tokens a month in May, up from 9.7 trillion two years earlier. Contracted backlog at Amazon ($578 billion), Oracle ($664 billion) and Google Cloud ($561 billion) totals $1.8 trillion, and OpenAI’s annualized revenue reached about $50 billion at the end of September.
The price of the product is deflating almost as fast as usage grows, because Google’s cheapest API model runs at $0.10 per million input tokens and OpenAI cut the output price on its frontier model by a third in October. Volume is overwhelming price compression for now, but it tells you where Stage 4 begins: when the token is the commodity and the model vendors compete on price alone.
The Constraints Moved to Power and Capital
Chips are no longer the binding constraint. Bank of America estimates data centers add roughly 125 gigawatts of U.S. load through 2030, more than 100 gigawatts beyond planned utility additions, so developers have announced about 101 gigawatts of on-site gas generation, which is why gas turbines, transformers and utilities with spare interconnection are capturing cash flow a software analyst never modeled.
Capital is the second constraint, and it’s where the risk concentrates. Goldman Sachs counted nearly $500 billion of AI-related debt issuance by August, with the hyperscalers only 40% of it, and estimates about a third of 2026 capex is debt-financed. Alphabet announced an $80 billion equity raise in June, and Nvidia’s 10-Q discloses guarantees capped at $105 billion on leases for data centers that will host its own chips for OpenAI, which is vendor financing by another name. With the 10-year Treasury above 5%, Goldman’s strategists noted on October 16 that high-yield spreads were widening fastest for AI-related issuers, and when a buildout depends on refinancing, the cost of capital becomes the thesis.
The Second-Order Evidence Cuts Both Ways
I expected software and services incumbents to lose economics, and the market agreed in February, taking the S&P 500 software and services index down 26% into April. Then the companies reported, expected 2026 earnings growth for the group rose to 20.6% from 13.8%, and the index made new 2026 highs in October. The erosion may still come, but it hasn’t shown up in the numbers yet. Where the second-order effect has shown up is in the memory squeeze, because DRAM prices up that much mean fewer PCs and phones get built. The indirect winners so far are the owners of power and land, and the lenders writing the private credit.
What the Market Is Confirming, and How We Stay
The Philadelphia Semiconductor Index peaked on June 22 after rising 311% from its April 2025 low, and closed November 4 down 29.8% from that peak and 5.1% below its 200-day average. It’s still up 45% for the year, so the longer trend is intact, but the intermediate trend has changed state. Leadership rotated inside the cohort: the neocloud group fell 44% from June through mid-October while memory led, then memory broke in November, and on November 4 Microsoft closed at a record while the chip index fell 3%. Capital is moving from the companies selling the hardware to the platforms that own the backlog, before capex has peaked.
That’s consistent with Stage 3. Stage 4 would be marked by token pricing dominating the conversation, custom accelerators selling outside their own clouds, memory prices normalizing, and capex growth falling below revenue growth. Goldman already estimates AI capex is about half of this year’s S&P 500 earnings growth and that the impulse fades through 2027, so the market is repricing now rather than waiting.
The market has to confirm the thesis before capital goes in, which means the exposure is in an uptrend by our definition and not just in the news. The downside is defined before entry with a predefined exit, the risk is sized from that exit to the risk budget, and we stay with the position while the expected payoff remains asymmetric. When the trend changes state, as the semiconductor cohort did this fall, the exit does its work and we rotate to where the cash flows are captured next. Everyone sees AI. The edge is in who gets paid, and in being gone when the market says they won’t.
