What do railroads in the 1880s, telecom fiber in 2000, and AI infrastructure in 2026 have in common? In each case, the people selling the shovels got rich first—and lost the most when the music stopped.

Capital expenditure as a percentage of GDP is now at an all-time high: 12.5%. For context, the dotcom bubble peaked at 11% before it imploded. According to Bridgewater, AI financing needs in 2027 are projected to hit $612 billion—more than the entire investment-grade and high-yield bond issuance combined ($470 billion).

If you're wondering whether that's sustainable, Brian Belski, a perma-bull at BMO Capital Markets, recently said a capex recession might hit in 2027. When the optimists start hedging, you should pay attention.

Here's the setup: For the last two years, Wall Street has been telling us that the Magnificent Seven can self-finance their AI buildout without taking on debt. That narrative just died. Amazon is projecting negative free cash flow for the first time in recent history, entirely due to capex spend. The others have turned to debt markets, vendor financing (which is just circular financing with extra steps), and now they're preparing massive IPOs to squeeze out the last available liquidity.

The problem isn't that companies are spending money. The problem is what happens when the ROI doesn't show up.

Let's talk about profitability and compute demand. OpenAI missed its revenue projections. Anthropic is likely profitable soon through its enterprise model, but Anthropic alone can't justify a 12.5% of GDP capex cycle. Recent research from Oviedo et al shows that frontier-scale inference consumes 0.31 watt-hours per query—4 to 20 times below public estimates. Translation: these models are getting way more efficient, which means they need less compute.

There's also something called RouteLLM, which can cut inference costs by 85% while maintaining 95% of GPT-4's quality. How? By routing simple questions to cheaper models and only using the expensive reasoning for complex queries. If you're running a public company post-IPO and your margins are under pressure, what do you think you're going to prioritize—user experience or earnings per share?

I'll give you a hint: it's the one that keeps your stock price from tanking.

The historical parallels are uncomfortable. In 2000, a telecom supplier CEO said this about the sudden demand collapse: "Institutional investors will not put more money into companies because they have not started towards revenue, which made them stop purchasing equipment... and then things happened very fast."

That's the thing about capex cycles. They don't gradually wind down. They snap.

Here's what's different this time—and not in a good way. Fiber optic cable from 2000 is still valuable today. Dark fiber that was considered worthless back then is now being lit up for data centers. But GPUs? A GPU you buy today will be obsolete by 2030, maybe sooner. That's a depreciation schedule problem that accountants are currently solving by extending GPU lifespans from 3 years to 4-6 years purely for GAAP net income purposes. Anyone who understands accounting knows this is a timing game. The costs will reverse through deferred tax liabilities, probably in the next two years.

Then there's competition. When margins are this high, everyone wants in. ASICs are taking inference share from Nvidia. China just told Nvidia they're not welcome back—they want their own shovels. Even Anthropic, the AI darling, trains Claude on TPUs and Trainium chips, not Nvidia hardware. And Nvidia's top three customers represent 54% of their revenue. What happens when those customers decide the ROI isn't there anymore?

Now, the bull case is that this time is different because the technology actually works and has real applications. That's fair. I'm not saying AI is worthless. I'm saying the infrastructure buildout is ahead of the actual revenue generation, and at some point, the market will demand proof.

That proof will come when OpenAI and Anthropic go public. Right now, Nvidia and the memory makers are the litmus test for AI. Once the AI companies themselves are public and we can see their actual margins, cash flow, and ROI—that's when things get interesting.

An anonymous investment banker posted this thesis on Reddit, and they're planning to short the infrastructure names late in 2026 unless the timeline changes. I'm not recommending you do the same—timing these things is nearly impossible. But the question is worth asking: if the people building the picks and shovels are showing up at the IPO window asking for money, what does that tell you about how sustainable this is?

Capex cycles always end. The only question is when.