How do companies actually make money from AI?

Four models dominate, and only two of them are clearly profitable today: selling the picks and shovels, selling access to models, selling AI-enhanced software, and selling outcomes. The infrastructure layer is the one reliably making money. **1. Infrastructure — the picks and shovels.** Chips, data centres, cloud compute. This is where the profits demonstrably are: everyone building AI must buy compute regardless of whether their own product succeeds. The classic gold-rush pattern, and it explains why chip and cloud companies captured so much of the value so far. **2. Model providers — selling access.** Subscriptions and per-token API pricing. Real revenue at meaningful scale, but with brutal economics: enormous training costs, ongoing inference costs, aggressive price competition, and capable open-weight models applying constant downward pressure. Revenue growth has been steep; profitability is a different question and several major players are heavily loss-making by design. **3. Applications — AI inside software.** The largest opportunity by count. Existing software adds AI features and raises prices or retains customers better; new products solve a specific workflow. The winners here tend to be ones with something a model provider can't replicate: proprietary data, deep workflow integration, distribution, or regulatory position. The losers are thin wrappers around an API, which get commoditised the moment the underlying model adds the feature natively — a pattern that has already repeated several times. **4. Outcome-based** — charging for work done rather than software. Support tickets resolved, documents processed, code reviewed. This is where the model shifts from replacing software budgets to replacing labour budgets, which is a much larger pool. Early, but it's where a lot of the ambition is pointed. The economics worth understanding if you're building or investing: - **Inference costs are real and recurring**, unlike traditional software where marginal cost is near zero. An AI feature that gets heavy use costs you money per use, which changes pricing and unit economics fundamentally. - **Costs per unit of capability have fallen dramatically** and continue to. This is good for application builders and hard for anyone whose advantage was affordability. - **The defensibility question is unresolved.** If the model is a commodity, value accrues to whoever owns distribution, data or the workflow — which historically favours incumbents. The honest summary: enormous capital deployed, genuine revenue at the infrastructure layer, promising but unproven economics at the application layer, and a wide gap between activity and profit that hasn't closed yet. Whether the current investment is justified depends on capability improvements that haven't happened yet — which is a reasonable bet and not a certainty.