# The AI capex boom is unlike anything America has built before and the funding structure explains why > Source: > Published: 2026-07-27 03:46:50+00:00 *The Bank for International Settlements has a blunt warning about the AI buildout: it is faster than canal mania, railways, electrification, and dotcom. The useful question isn't whether the boom is real. It is who is paying for it, and how long they can keep paying.* The AI capex boom is no longer a future risk tucked inside earnings-call language. It's here. The BIS, in its Annual Economic Report published on June 28, compared today's AI investment race with the canal mania of the 1830s, British railway mania in the 1840s, electrification in the late 1920s, and the dotcom boom of the late 1990s. Its chart starts each cycle at its pre-boom trough. AI has reached 4.5 times that level in three years. That should get your attention. Canal mania peaked at 4.1 times over five years. Railway mania hit 2.7 times over four. Electrification and dotcom both topped out near 1.9 times, and the BIS noted that those earlier cycles ended with reversals in investment and economy-wide recessions. The point is not that AI is fake. It isn't. The point is that real technologies can still attract more capital than their commercial returns justify. The spending numbers make the comparison less academic. Reuters reported on July 22 that Microsoft, Alphabet, Amazon, Meta Platforms, and Oracle are expected to spend more on capex than they generate in free cash flow by 2027, based on LSEG consensus estimates. The same Reuters analysis said current-year capex estimates for those five companies had risen from about $485 billion in January to around $730 billion in July. FactSet, in research published July 23, put aggregate FY26 hyperscaler capex above $690 billion. We're talking about a real industrial buildout. Data centers, servers, networking gear, power contracts, cooling systems, chips, land. Not pitch decks. ## The cash cushion is getting thinner The strongest argument against the bubble comparison has always been simple: the big AI spenders are not the weak telecom borrowers of 2000. Alphabet, Amazon, Meta, and Microsoft came into this cycle with cash-rich balance sheets and large operating cash flows. S&P Global Ratings said in May that the five rated hyperscalers, including Oracle, still had investment-grade profiles, though Oracle was already the outlier with weaker free cash flow and more debt on the books. That's not nothing. The railways were built on debt. The late-1990s telecom boom was built on debt. When revenue failed to cover the financing, the result was dark fiber, bankruptcies, and assets changing hands at painful prices. The early AI buildout looked different because so much of it came from earnings. But that structural advantage is now less clean. FactSet said incremental annual debt among Alphabet, Amazon, Meta, Microsoft, and Oracle rose from 9% of capex in FY24 to 32% for the twelve months to mid-2026. It also said Alphabet priced an $84.75 billion equity raise in June 2026, while Oracle planned roughly $40 billion of combined debt and equity for FY27. That is not panic financing. It is still a change you shouldn't ignore. Look at Oracle. Reuters, citing LSEG data, said Oracle's capex rose from 47% of operating cash flow in fiscal 2022 to 174% in fiscal 2026, which ended in May. Its capex reached $55.7 billion against $32 billion of operating cash flow, and free cash flow has turned negative. Oracle is not the whole AI economy, but it is the clearest example of what happens when infrastructure ambition outruns internally generated cash. Reuters also put the forward squeeze in one useful ratio. Between 2025 and 2027, operating cash flow across the five hyperscalers is expected to rise by about $340 billion, while capex rises by about $534 billion. That's $1.57 of extra investment for every new dollar of operating cash flow. The risk has shifted. It is no longer only whether companies can afford today's buildout. It is whether investors keep accepting more bonds, more equity, more leases, and more long-dated commitments before AI revenue proves large enough to carry the load. ## The winners may not be the biggest spenders History is harsh on the layer that absorbs the most capital. Railway investors did not all get rich because Britain got railways. Dotcom fiber did not create durable value for every company that laid it. The internet still won. Plenty of investors did not. That distinction matters for you if you're sizing a startup idea or a public-market position - or working out where your fundraise fits into all this. The BIS said the AI boom is being driven by expectations of large productivity gains, but it also warned that competitive pressure can push capex so high that sector-wide surplus falls or turns negative in adverse scenarios. A useful product can exist inside an ugly investment cycle. Those two facts can sit together. The cloud cycle is the cleaner counterexample. AWS, Azure, and Google Cloud spent brutally for years and came out with scale advantages that smaller rivals couldn't match. Scale matters. In AI, the same companies are trying to own the data centers, the chips or chip supply, the cloud platform, and the model services sold on top. If that vertical structure holds, they may avoid the old pattern where infrastructure owners build the road and someone else captures the tolls. Frankly, nobody knows yet. The honest bet is that hyperscalers have already made this a game for balance sheets few companies can match - compute is simply not a fight worth picking at the application layer. Use the overbuild. Prices and model quality are moving in your favour right now, and access is only getting easier - that won't last forever. The durable companies after prior booms were not always the ones that spent first. They were the ones that used the infrastructure after someone else had paid for the excess. **Also read:** [Meta removed 344,000 AI nudify ads in six months and the operators kept coming back](https://startupfortune.com/meta-removed-344000-ai-nudify-ads-in-six-months-and-the-operators-kept-coming-back/), [SK Hynix just moved $26.5 billion onto Nasdaq and then moved South Korea's currency](https://startupfortune.com/sk-hynix-just-moved-265-billion-onto-nasdaq-and-then-moved-south-koreas-currency/), [Moonshot AI releases Kimi K3 open weights making the world's largest open-weight model free to download](https://startupfortune.com/moonshot-ai-releases-kimi-k3-open-weights-making-the-worlds-largest-open-weight-model-free-to-download/)