AI is proving to be a transformational technology. If nothing else, the AI labs’ record revenue growth shows how much value is being created. But massive investment and fierce competition between companies and nations have broken the old rules of investing1. Retail investors can no longer just lean on The Intelligent Investor and other classics, which were not written for times without precedent. Yet they can’t sit this out: AI companies take up a growing share of the index funds that hold their retirement savings.
To help retail investors make better decisions, I built a small set of metrics that summarise the AI buildout and show both its risks and its opportunities. They can be tracked here and are updated regularly; the points below explain what each one shows.
1. AI labs are growing fast, and their business runs on healthy gross margins2. The graph below reports annualized revenue for Anthropic and OpenAI, calculated as Revenue for a given period × Number of periods in a year3. This is the metric that justifies the huge sums being spent on datacenters. More revenue means more people using the labs’ models, and more usage needs more datacenters to run them.
2. To sustain that unprecedented growth, datacenters need to be built as fast as possible. Because of their datacenter building experience, huge accumulated war chests, and highly profitable business models, hyperscalers are footing the bill. You can track the capital expenditures of hyperscalers in the graph below.
3. Datacenters take years to build, so hyperscalers must invest ahead of the demand they expect from AI labs. The opportunity and the risk both depend on how accurate those forecasts are: build too little and growth stalls, build too much and the spending outruns revenue. Three checks help:
- Does capex (Figure 2) grow in step with AI lab revenue (Figure 1)?
- Do commitments (Figure 3) stay in proportion to contracted revenue?
- Can each company pay for its spending from operating cash flow (Figures 2 and 4)?
4. Hyperscalers also finance datacenters indirectly. By signing long-term leases and other off-balance-sheet commitments, they give partners such as neoclouds and datacenter developers the long-term contracts they need to borrow money and build. The graph below tracks these commitments alongside contracted revenue: future revenue the hyperscalers have already signed with AI labs, Fortune 500 companies and other customers.
Comparing the two shows how much of the spending is already backed by paying customers. The commitments pay for the capacity needed to deliver the contracted revenue, and may also fund new lines of business4.
5. As long as AI lab revenue keeps growing fast (Figure 1), the different AI buildout supply chain players would see return on their investments:
- Hyperscalers turn contracted revenue into actual revenue as they deliver the compute the labs signed up for.
- Neoclouds and datacenter developers get paid as hyperscalers start using the capacity they committed to.
- Chip and memory makers receive larger orders as everyone adds capacity.
If the revenue growth of AI labs unexpectedly slows down, hyperscalers whose capital expenditure as a share of operating cash flow is above 100% will need to cover the difference from their cash reserves, by cutting buybacks and dividends, or by borrowing more. This would affect stock price and perceived credit risk. You can track Capex as a share of operating cash flow in Figure 2 above.
6. Figure 4 shows what hyperscalers’ capex would look like if they paid for their current off-balance-sheet commitments evenly between now and the end of 20305. It’s a scenario, not a forecast: it counts only commitments already signed and adds no new spending. Compared with today’s operating cash flow, it shows that Amazon and Microsoft could cover these payments without borrowing. Hyperscalers only have reason to add new commitments if lab revenue keeps growing, so watch for commitments rising faster than the revenue in Figure 1.
7. Some companies may need to refinance or borrow more if datacenter projects cost more or take longer than planned, or if AI lab revenue grows more slowly than expected. The bond market is where that risk would show up first6. Figure 5 tracks each company’s GZ spread: the extra yield investors demand to hold its bonds instead of US Treasuries paying the same cash flows. The higher the spread, the more investors worry the company won’t repay. Each company is compared with the average US corporate bond. To assess a company’s risk of defaulting, check its capital expenditure in Figure 2, its off balance sheet commitments in Figure 3, and its accumulated war chest.
8. Finally, the AI buildout is now a large part of any ordinary index fund. Compute suppliers such as Nvidia, Broadcom and AMD make up 15% of the S&P 500 and 23% of the Nasdaq-100 (Figure 6, as of June 2026). The business of these compute suppliers has historically been cycle based; therefore, it’s important to keep track of any potential slow down in the number of chips sold as a signal of the current cycle’s end.
Closing thoughts
The AI buildout has been compared to the dotcom bubble of 2000 and the housing crisis of 2008. In 2000, telecom companies borrowed heavily to lay fiber for demand that took years to arrive. In 2008, risk built up off balance sheets and in credit markets.
If you’re an optimist, it’s worth taking any comparison seriously: in both cases, it took up to 7 years for safer investments in indexes to recover from the drawdown. If you had the misfortune to buy Intel or Cisco stocks at the height of the dotcom bubble, it would have taken 26 years to recover your original investment. Conversely, if you’re pessimistic, consider this: individual investments in a stock like Amazon in 2000 would still have yielded a 63x return.
What’s new today is that retail investors have access to more information than ever. It won’t make decisions for you on when to buy or sell. But it can help you stay clear of the froth while still taking part in what could become the biggest engine of growth humanity has ever built.
Footnotes
- This is likely a temporary phenomenon. Just like technology companies have provided vastly better returns than any other companies in history, AI companies will have orders of magnitude more impact over the long term, which means that holding index funds is still a good strategy. However, because of the levels of risk associated with the current buildout, there is a possibility of a significant drawdown that could make it difficult to get liquidity without losses in a less than 5 year time horizon. ↩︎
- Anthropic reports ~80% before revenue sharing schemes with partners. ↩︎
- The exact period is not known. We will have more visibility on this metric when Anthropic publishes their S-1 filing and goes public in November 2026. ↩︎
- That bet is biggest at Meta and Alphabet, whose commitments are far larger than their contracted revenue. Both companies have offerings competing directly with the AI labs, therefore plan to be their own consumer of compute. ↩︎
- Which is fairly conservative; we don’t have full visibility on when a lot of these commitments are due, but some leases have been mentioned to be up to 30 years in the hyperscalers’ SEC filings. ↩︎
- Bond investors care about one thing: getting paid back. If the borrowing company is cash-rich, short term bad news is unlikely to increase perceived risk on the bonds market. However, if the borrowing company has liquidity problems, it will be seen as at risk of defaulting, even if its future revenue generation prospects are attractive. This is important to keep in mind when analysing companies like Oracle. ↩︎