No Statistics on My Railway

No Statistics on My Railway

Mara: York, 1842. George Hudson has just joined the board of a railway company. His first piece of business, as the finance blog A Wealth of Common Sense tells it, is to change how the company keeps its accounts. And he puts it plainly: "I will have no statistics on my railway!"

Eli: That's where a listener's question starts: is the AI boom a bubble like the railroads and dot-coms, or is this time actually different? Today's angle is the trail of capital expenditure across technological buildouts.

Mara: Hudson started with a real need. York wanted better links to the industrial towns around it. He backed railway schemes, learned the business travelling with surveyors, and had a talent for joining short local lines into something more useful.

Eli: So a real builder, not just a salesman.

Mara: A very big one. By 1844 he oversaw more than a thousand miles of track. That was a third of all the railway running in Britain.

Eli: The figures he did offer were promises. On some shares, Hudson promised fifty percent dividend yields, which pushed prices up and encouraged insider trading by him and his directors. And getting in was cheap, because railway shares could be bought partly paid. Due by law, if the company called for it. And the money flooded in. By 1850, about £250 million had gone into British railways. In its day, that was almost half of everything Britain produced in a year.

Mara: Half a nation's yearly output in one kind of asset, with its biggest operator refusing to count. The companies building AI point to one big difference, Eli: they pay for it out of profits.

The Fortress Balance Sheet

Eli: The nineties already tested whether a buildout this big can avoid a crash when spending runs far ahead of what customers pay.

Mara: And one investor saw it coming. In 1999 and 2000, a veteran investor and his partner at their investment firm decided to get out of everything they could. He later recalled that they knew markets were overheated, the dot-com boom had gotten ridiculous, and the S&P 500's valuation was nonsense. So they got out.

Eli: Selling while everyone else was still buying.

Mara: The telecom carriers were doing the opposite, laying fiber on borrowed money and on vendor financing, which means an equipment maker lending customers the cash to buy its own gear. At its peak, the equipment maker Lucent had lent out more than fifty times the cash its own operations brought in. One industry analysis calls the telecom bubble "heavily debt-financed and vulnerable once capital markets closed." When credit tightened, it all unravelled.

Eli: To its own customers. So a chunk of its sales ran on its own credit.

Mara: Which is the weakness today's defenders say this boom doesn't have.

Eli: Their case starts with who's paying. Opus Interactive says that many dot-com companies had little revenue and no clear path to cash. This build is led by the hyperscalers, the four biggest spenders on AI data centers: Microsoft, Alphabet, Amazon, Meta. They came in with some of the market's highest free cash flow, the cash left after running and investing in the business, and they were paying mostly from that, with little net new debt.

Mara: So no Lucent-style borrowing. The money is their own.

Eli: And by the old yardstick, the bulls say, the share prices look tame. Price-to-earnings measures what investors pay for each dollar of profit. Those four traded at about 26 times their expected earnings two years out. At the dot-com peak, Cisco traded at roughly 201 times.

Mara: Twenty-six against two hundred. That gap anchors the bulls' case, the one Joseph Briggs at Goldman Sachs makes: rich companies spending their own money, rationally.

Eli: And it holds only as long as the profits keep paying the bills.

The Trillion-Dollar Hurdle

Mara: Hudson didn't want the statistics kept. On AI, David Cahn has spent three years keeping them.

Eli: Cahn is a partner at the venture firm Sequoia, and his tool is the revenue hurdle: how much customers have to pay each year to justify what's being built. In September 2023, he put the gap between that hurdle and real AI revenue at 125 billion dollars. By June 2024, he wrote that the hole was "now going to become a $500B hole."

Mara: Four times bigger in under a year.

Eli: And spending kept running ahead of revenue. The hyperscalers were expected to spend roughly 650 to 700 billion dollars in 2026, nearly double the year before. Every dollar of that raises the bar customers have to clear. By July 2026, Cahn's hurdle had reached one and a half trillion dollars.

Mara: Against that, even optimistic counts put yearly AI revenue at a sixth of it, at most.

Eli: That gap is the reading Cahn shares with skeptics who saw the dot-com bust up close: spending so far past returns that a sharp correction has to come.

Mara: Joseph Briggs reads the same climb another way, Eli. He puts AI capital spending at nearly two percent of US GDP in 2026. And he writes that these levels are "consistent with the 2%-5% of GDP peak investment impulses observed in prior general-purpose technology buildouts." A surge this size, in his reading, is what a big new technology looks like.

Eli: Railways were that kind of surge. Whether the people funding them got their money back was a separate question.

Mara: What nobody knows yet is whether businesses outside software coding will pay for AI at anything like that scale. And a hurdle that high stays harmless only while the builders can cover it from their own cash.

Cash Runs in Circles

Eli: For years they could. The cloud and software businesses threw off cash, and the cash built the data centers. In 2023, building took about 40 percent of what those businesses brought in. PIMCO expected that share to reach 94 percent in both 2026 and 2027. Almost nothing left over. By the third quarter of 2026, spending had passed the cash coming in. Man Group found a growing share of it was being paid for with borrowed money.

Mara: So the profits that were the whole case for this boom stopped covering the bill.

Eli: Alphabet shows it most starkly. In the second quarter of 2026, its free cash flow was negative 5.9 billion dollars. Its first negative reading since it went public in 2004.

Mara: Its whole life as a public company, never below zero until then.

Eli: And it wasn't alone. Over twelve months, Amazon's free cash flow came out negative too. Meta's barely cleared zero. Microsoft was the outlier, still well in the black.

Mara: Three of the four giants at or near zero, after spending.

Eli: So the money had to come from elsewhere. In a single year, Alphabet's debt grew about sixfold, to roughly 100 billion dollars.

Mara: And it sold shares on top of that, announcing an 80-billion-dollar equity raise in June 2026. That included a ten billion dollar private placement with Berkshire Hathaway.

Eli: The shift from paying for data centers out of operating cash to heavy debt and off-balance-sheet leasing points to a capital structure that's changing underneath these companies.

Mara: And some of the money runs in a loop. Circular financing is a supplier funding its customers, who then spend that money buying from the supplier. Microsoft pays OpenAI. OpenAI pays Nvidia. Nvidia invests back into AI startups and cloud providers, and it has disclosed about 110 billion dollars in investments and financing commitments.

Eli: That's Lucent's nineties playbook, at a far larger scale. The buildout that was supposed to pay for itself now runs partly on borrowed money and partly on money going round in circles.

Tracks Across the Ruins

Mara: Hudson's reckoning started in 1847. Poor harvests and expensive food imports drained Britain's gold reserves. The Bank of England tightened credit, and money got hard to find. Now those small deposits had a sting. Shareholders still faced capital calls, the company's demands to pay in the rest of what they'd pledged, and investors short of cash had to sell.

Eli: By early 1850, railway shares were down 85 percent on average from 1845.

Mara: Then investigators looked at the numbers Hudson hadn't wanted kept. Some of his companies had paid dividends out of money raised from shareholders, and presented it as profit. One of them, the Eastern Counties Railway, paid about £200,000 that way. Evidence that fraud drove the boom-year dividend rises is limited, and it wasn't everywhere. But in 1849, Hudson was forced out of his railway chairmanships.

Eli: Shareholders' own money, handed back to them and called a return.

Mara: And his own fall kept going. A few years after rising to the top, Hudson was fleeing creditors. He was eventually arrested for unpaid debts, and he died broke.

Eli: The track stayed, though. Between 1844 and 1850, even as shares fell from their 1845 peak, Britain's railway network roughly tripled. Two economists at BCG put the general point: bubbles can leave transformative infrastructure behind even when investors lose money. Here, lines from the boom years came to make up about 90 percent of the British system.

Mara: Shareholders wiped out, rails still carrying the country, Eli. Whether today's buildout leaves anything that lasts like that iron is still the open part.

The Silicon Depreciation Trap

Eli: Iron lasts. Chips don't, and chips are where most of the money goes. GPUs and CPUs, the processors doing the AI work, take roughly two-thirds of data-center spending, by one estimate as much as 80 percent. Builders say they'll depreciate them over five years, spreading that cost on their books across five years. One skeptic points to Nvidia's own claim that each new generation makes the old ones obsolete within two. That's economic obsolescence: the chip still runs, but newer, cheaper ones make it not worth running.

Mara: That's David Cahn's reading, and other skeptics share it: chips that age out leave stranded costs, not rail beds.

Eli: The other reading has a powerful backer in the Fed chair, Kevin Warsh: this buildout leaves foundations that last decades, the way rails and fiber did.

Mara: And Warsh spoke up for it in a rough week. On Friday, October 16th, 2026, chip stocks sold off for a ninth straight session as the main indexes slumped. That same day, Warsh backed the boom, calling AI "probably the most consequential change to the U.S. and the global economy in my lifetime." The markets that day looked more worried about a bubble.

Eli: And the worry is how far a fall would travel. A US credit-rating agency warned that a sharp correction in AI stocks could push the country into recession. The IMF sketched a downside scenario, not a forecast, where global GDP ends 1.1 percent lower by 2028.

Mara: The IMF also named an AI-driven correction as a financial-stability risk, with many channels between the AI industry and traditional banks. That's where Cahn's trillion-and-a-half hurdle stops being a spreadsheet gap: if borrowed money fills it, the risk has a road into the banks.

Eli: So the answer for the listener: this isn't a simple repeat of the railroads or the dot-coms, and it isn't immune either. It's anchored by profitable giants, unlike the dot-com startups. But it's picked up the old habits: spending past its cash, new debt, money moving in circles. And unlike iron rails, half or more of what it buys ages out within a few years.

Mara: Hudson wanted no statistics. This time the statistics are what to watch. The coding question from earlier, first. Then how deep the write-downs go on aging GPU clusters, and what those clusters sell for secondhand.

Eli: Default rates in the private credit funds behind off-balance-sheet data centers. And how much chipmakers disclose of revenue that rests on financing they provided themselves.

Mara: Hudson's rails outlived his fortune. Whether these chips outlive the loans taken out to buy them is what those numbers will show.

Eli: Dig into vendor financing.

Mara: And follow economy with us.

Eli: That's it for today. I'm Eli.

Mara: And I'm Mara. Thanks for listening, and until next time on Angle.