The AI investment thesis

This is the defining market debate of the mid-2020s, and it is a debate, not a fact, which is exactly why it makes a good teaching case. This guide does not tell you whether to be bullish; it frames the question the way a professional does, lays out the strongest version of each side, and points at the evidence that will actually resolve it. It applies every tool in the school to a live, unsettled thesis.

The short version

The AI thesis is really several stacked bets: on compute (chips, power), on infrastructure (clouds, data centers), on models, and on applications, each with a different risk and a different competitive structure. The bull case is a genuine general-purpose-technology platform shift with enormous productivity upside; the bear case is a capex boom running far ahead of end-customer revenue, echoing the 2000 telecom build-out. The unresolved crux is MONETIZATION: whether AI generates enough durable end revenue, fast enough, to justify the infrastructure being built for it. Professionals size for a wide range of outcomes, separate the layers (some win even if the thesis disappoints), and watch capex-versus-revenue as the decisive gauge.

The debate, framed properly

The mistake at both poles is treating "AI" as one trade. "AI will change everything" and "AI is a bubble" can both be true at once: the internet changed everything AND the 2000 telecom/dot-com capex bubble vaporized enormous capital on the way. A professional decomposes the monolith into layers and asks, at each, three questions: how big is the value created, who captures it, and is the current price already assuming the answer? The reverse-DCF habit from the valuation track, applied layer by layer.

The value-chain stack

Layer	What it is	Competitive structure & risk
Semiconductors	The training/inference chips and their memory	Near-monopoly economics today; the margin is the whole bull/bear (see the NVIDIA case). Custom silicon is the attack.
Power & grid	Electricity, transmission, cooling for data centers	The physical bottleneck; slower-moving, arguably underpriced (see the Power & Hyperscalers guide)
Infrastructure / cloud	Hyperscaler data centers renting compute	Oligopoly with huge balance sheets; the ones SPENDING the capex, so the bear case's front line
Foundation models	The frontier LLMs themselves	Fast-moving, capital-hungry, unclear moat: are models a product or a commodity trending to zero margin?
Applications	Software embedding AI to sell an outcome	Where end revenue must ultimately appear; fragmented, and the layer the whole stack's payoff depends on

The layers do not rise or fall together. It is coherent to believe compute wins and models commoditize, or that applications capture the value while infrastructure over-builds. Separating them is the analysis; conflating them is the error that makes people either buy the whole basket at any price or short it entirely.

The bull case, stated at full strength

A general-purpose technology. Like electricity or the internet, AI plausibly raises productivity across the entire economy, not one sector. GPTs create value for decades and diffuse everywhere; the winners are worth more than any single-industry framing captures.
Real demand, real revenue, already. Unlike pure story-stock manias, the leading infrastructure and chip players posted enormous ACTUAL revenue and cash flow, not just narrative. Hyperscalers with real profits are funding the build from operating cash, not junk debt: a healthier base than 2000's telecoms.
Reflexive moats. Scale (data, compute, capital, talent, distribution) compounds, and the leaders may pull further ahead, justifying premium economics.
Early innings. If AI is genuinely transformative, even today's large numbers are small against the eventual market, and the reverse DCF's aggressive assumptions turn out conservative.

The bear case, stated at full strength

Capex is lapping revenue. Hundreds of billions in annual AI infrastructure spending must eventually be justified by end-customer revenue. If the applications layer cannot monetize fast enough, the build-out corrects the way telecom fiber did in 2001: the technology was real, the demand eventually came, and the capital that funded the early over-build was still destroyed.
Margins are a target, not a moat. Today's extraordinary chip and cloud margins are precisely what every well-funded competitor and every large customer's in-house program is attacking. Extraordinary margins invite the capital that ends them.
Models may commoditize. If open-weight models stay close to the frontier, the foundation layer's pricing power erodes toward the cost of compute, stranding the capital raised on model-monopoly assumptions.
The concentration feedback. The AI winners became a huge share of the index (see the S&P 500 case), so "the market" is now partly an AI bet, and a disappointment transmits far beyond the sector, mechanically.

Both cases are strong, which is the point

Notice that a serious version of each side is compelling: that is the signature of a genuine, unresolved thesis, and the reason confident certainty in EITHER direction is a warning sign about the person, not a conclusion about the market. The professional stance is not to pick a side and defend it, but to identify the specific, observable evidence that would move the probability, and to size so that being wrong is survivable in either direction.

The crux: the monetization question

Strip away the noise and one empirical question decides the thesis: does AI generate enough durable end-customer revenue, fast enough, to justify the infrastructure built for it? Everything else is derivative. The infrastructure spend is a bet ON that revenue appearing; the chip margins depend on that spend continuing; the index level depends on those margins. It is one question wearing four costumes, and it is genuinely open: enterprise AI revenue was real and growing but small against the capex when this guide was written, and the gap either closes (bull) or the capex slows to meet it (bear). An honest analyst holds this as a probability distribution updated each quarter, not a conviction.

What professionals actually watch

The capex-to-revenue ratio, aggregated. Hyperscaler AI capex guidance against disclosed AI revenue: the single most important series, and the one the bear case lives or dies on. Rising capex with lagging revenue is the 2001 setup; converging lines are the vindication.
Unit economics of inference. Falling cost per token is bullish for applications (more use cases become profitable) and bearish for model pricing power: the same data point cuts differently by layer.
Custom-silicon progress in hyperscaler disclosures: the direct measure of the chip-margin bear case.
Power availability. Increasingly the binding real-world constraint; grid interconnection queues and power-purchase deals are becoming the true near-term ceiling on the whole build.
Enterprise adoption depth. Not pilots but production deployments with measured ROI: the evidence that the applications layer is closing the monetization gap.

How professionals position a thesis this uncertain

Separate the layers. Some (power, arguably) win across a wide range of AI outcomes; others (a specific model company) need the full bull case. Own the ones whose payoff does not require the most optimistic assumption.
Size for the distribution, not the mode. A thesis with genuinely bimodal outcomes (platform shift vs capex bust) is sized so that the bust is survivable and the boom is meaningful: the asymmetry test from the valuation track.
Respect the concentration you already own. An index holder is already long AI; the factor-regression habit reveals how much, before adding more.
Write the falsifier now. "I exit if the capex/revenue gap widens for three more quarters" is analysis; "I'll know a top when I see it" is not. The thesis-doc discipline matters most exactly where conviction runs highest.

Glossary for this guide

Reverse DCF: Running the machine backwards: instead of estimating value from assumptions, solve for the growth and margins the current price already implies, then judge whether those are beatable. Often more honest than the forward version, because it removes your favorite input.
Total addressable market: The revenue available if a company served every possible customer. Useful as a ceiling and abused as a pitch; the discipline is asking what fraction is realistically serviceable and at what cost.
Moat: A structural barrier that stops a company's high returns from being competed away: network effects, switching costs, brands and patents, cost advantage, or efficient scale. The durable question is never whether a business is profitable but what protects the profits.
Gross margin: Gross profit divided by revenue: the share of each sales dollar left after the direct cost of what was sold. The first test of whether the product itself, before any overhead, makes money.
Capex: Cash spent on long-lived assets: plants, equipment, software. Maintenance capex keeps the current business running; growth capex expands it. Both reduce free cash flow now in exchange for cash flows later.
Crowding: Too much capital in one trade. Crowded strategies unwind together, briefly correlating at the worst moment; measuring how much of a story is already positioned is half of risk management.
Falsifier: The observable fact, chosen in advance, that would prove the thesis wrong and trigger exit. Deciding it while calm is cheap; deciding it mid-drawdown is expensive, which is why professionals write it down first.
Expected value: The probability-weighted average of the outcomes: bull, base and bear prices times their odds. The number that decides whether a position is attractive, and the reason a likely-wrong idea with a huge payoff can beat a likely-right one with none.
Factor regression: Regressing a portfolio's returns on the standard factors to see what it actually is. The X-ray that turns many celebrated records into market beta plus a static tilt available in a cheap ETF.
Catalyst: The event expected to force the market to reprice toward your view: earnings, a spin-off, a regulatory decision, a contract. Cheap without a catalyst can stay cheap for years; the catalyst is the thesis's clock.