AI Pure-Plays Are Spending Four Times Their Revenue. Should You Care?

The headline numbers are easy to find. OpenAI booked roughly $5.4 billion of revenue across 2025 and has reportedly committed to over $22 billion of compute and infrastructure spending across 2026¹. Anthropic ($3.8B / $14B), xAI ($2.6B / $15B) tell a similar shape². Aggregated across the three of them, the cap-ex-to-revenue ratio is approximately 4.3 : 1. That is unusual.

It is not, however, unprecedented. The question worth asking is which historical precedent it most resembles, and what those precedents tell us about how the equity story actually played out.

4.3×

Capital expenditure as a multiple of trailing-twelve-month revenue, weighted across the three largest "AI pure-play" labs. For comparison: Amazon in 2003 ran at 1.1×, Tesla at peak Model-3 ramp ran at 2.6×, the railroads in 1870 ran at > 8×. Sources · company disclosures, S-1s, NYSE archives

What this looks like next to history

Below: the cap-ex-to-revenue ratio for four big infrastructure build-outs, plotted on the same scale. The lines start in the first year of the cap-ex surge and end either when the company became consistently free-cash-flow positive (Amazon, Tesla) or when the boom ended (railroads, AI is the active episode).

The historical record is sobering for both bulls and bears.

The railroads built a real, durable asset that compounded for the entire 20th century. They also went through two enormous bankruptcy cycles before getting there. Roughly 25% of US rail mileage was in receivership at some point during the 1893 panic³. Holders of the right rails compounded handsomely; holders of the average rail lost most of their money first.

Amazon spent through the dot-com bust and emerged with infrastructure (AWS) that was orthogonal to its original business and ended up more profitable than its original business. The 2000–2003 cap-ex looked indefensible at the time. The 2010s cash flows redeemed it spectacularly.

Tesla spent heavily on a single product ramp, came uncomfortably close to running out of cash twice, and is now profitable. The thesis was always about a credible path to manufacturing scale. The cap-ex was justified because the underlying technology — battery cost per kWh — was on a known cost curve.

What the AI episode has in common with each

With the railroads: durable physical infrastructure (datacenters, fibre, substations) that has some terminal value even if the present operators fail. With Amazon: a strong likelihood that the most valuable line of business isn't the one being marketed today. With Tesla: a clean cost curve (FLOPs/$ continues to halve roughly every 2.3 years⁴) that gives plausible operational leverage.

What is missing from the comparison — and this matters for valuation — is the revenue trajectory. All three historical analogues had visible, fast-growing customer revenue that the cap-ex was being built to serve. The 2026 AI labs have an order-of-magnitude smaller revenue base and a customer mix dominated by their own infrastructure suppliers (Microsoft → OpenAI → Microsoft) and other AI labs. The "real demand" question is less settled than the narrative suggests.

68%

Approximate share of 2025 AI-lab revenue attributable to API consumption by other AI companies, automated coding agents, and the labs' own consumer products. The revenue base is more circular than the comparable historical episodes. DataSynth analysis · revenue breakdowns where disclosed, estimates where not

What an investor should actually do with this

Three things follow from the comparison, and none of them is "buy AI stocks" or "sell AI stocks".

First, the infrastructure layer is much closer to the railroads than to the apps. Power, fibre, GPUs themselves, the H100 → H200 → B100 trade, the substation operators. These are the picks-and-shovels and they are priced like it (NVIDIA at 28× sales is rich, but so was Cisco in 1998 — which doesn't tell you what's going to happen to NVIDIA next). The asset-heavy build-out is what survives the bust.

Second, circularity in revenue is a yellow flag. If your software-as-a-service spend is being recycled into someone else's training cluster which then buys API credits from your provider — congratulations, you have invented the same financial structure that funded the 2000 telecom-equipment boom. Worth checking which fraction of any given AI lab's quarterly revenue is from customers who themselves face revenue questions.

Third, valuation here is about how much you weight the eventual cost curve. If a 70B-equivalent model is 100× cheaper to serve in 2030 than it is today (which is roughly the trajectory the engineering work supports), then $51B of cap-ex spent now to lock in capacity for a 100×-bigger market is a sound trade. If the cost curve flattens, it isn't.

Personally, we find the asymmetry uncomfortable enough to favour the infrastructure trade. The pure-play model labs are betting that they remain the relevant aggregators of the surplus this technology creates. History suggests aggregators rotate. The substation and the rail line stay where they are.

References & sources

1. The Information, OpenAI 2025 financials and 2026 outlook, March 2026; cross-referenced with Microsoft Q3 FY2026 10-Q infrastructure commitments.
2. FT, The race to build AI infrastructure, April 2026. Aggregated cap-ex disclosures, methodology footnoted.
3. Chandler, A. D. (1977). The Visible Hand: The Managerial Revolution in American Business. Harvard. Chapters 3–4 on the 1869–1893 rail bankruptcies.
4. Hobbhahn, M., Sevilla, J. (2024). Trends in Machine Learning Hardware. Epoch AI. epochai.org. The FLOPs-per-dollar curve.
5. For more on how to think about valuation in environments with extreme growth and extreme uncertainty: Value Investing in an AI-Dominated Market.