Will The AI Boom End Like The Railroad Crash Of 1873? — ft. Liaquat Ahamed

NVIDIA

Across the shows: 36 speakers in other episodes, 29 bullish and 9 bearish calls, past 30 days

• Liaquat AhamedBearish

  Nvidia is effectively lending to its customers, making data center financing hard to gauge.

  At the centre of the data center financing chain is Nvidia, which Ahamed says is essentially lending to its customers even though it does not call it lending. Because this vendor financing is not visible, it is hard to know how much of the construction is debt-financed. He treats this opacity as a warning sign, so his stance is cautious, though he states no explicit price or fundamental view on the stock.

  “At the center is a company like NVIDIA, which is essentially lending to its customers. It doesn't call it lending, but it is a form of lending.”at 21:59

AI infrastructure & data centers

• Liaquat Ahamed

  Data center buildout is increasingly debt-financed and opaque, with no clear sense of total debt.

  Until recently hyperscalers funded data center construction out of earnings. The scale of the build-out has now outrun their resources, so they have turned to debt. Much of this debt is opaque, and nobody knows how much there is. Nvidia is effectively lending to its customers, though it does not call it lending, so the share of construction financed by debt is unknown. Ahamed sees this as the key risk, drawing a parallel to 1873, when about three quarters of railroad construction was financed with borrowed money.

  “the key risk in the data center boom is overbuilding and overbuilding finance by debt. ... We don't actually know how much debt there is.”at 21:07

• Liaquat Ahamed

  Data center overbuilding is structural: each firm's caution still adds up to excess capacity.

  Satya Nadella said Microsoft would prevent overbuilding by being careful, citing the railroad book. Ahamed argues Nadella got the wrong end of the stick. Each company acts carefully on its own, but the collective outcome of their decisions causes overbuilding. This is a prisoner's dilemma: no firm wants to give up market share by slowing construction, so each keeps building. Preventing it would require collective action, which he doubts is possible, so overbuilding is inherent in the competition to be one of the two or three winners.

  “I think he got the wrong end of the stick. Because each individual railroad company, in the same way that each individual tech company, is trying to be careful. It's the collectivity of their decisions that leads to overbuilding.”at 24:57

• Liaquat Ahamed

  Data center investment is about 3-4% of GDP, below prior 5% boom peaks, so the boom may be mid-way.

  Past private investment booms ran at 2-4% of GDP and peaked at about 5% of GDP, and each lasted two to four years. The current data center boom has run about two years, and next year's spending is projected at around $1.5 trillion, roughly 3-4% of GDP. Ahamed concludes that the boom is not yet at historical extremes and could be halfway through, though he says history gives no precise timing. He also notes that AI now dominates the stock market and crowds out other investment.

  “We've had booms go on for four years. So we could be halfway through the process. ... This current boom is not at 5% of GDP... That's probably three to 4%. We're not at the sort of extremes that we've experienced in the past.”at 29:10

• Liaquat Ahamed

  Local community protests against data center construction are a modern random risk that could derail financing.

  In 1873, Jay Cook's failure to finance the Northern Pacific was partly caused by running into wars with Native American tribes in the northern territories, which made the railroad's goal impossible. Ahamed sees the modern equivalent as political protests against data center construction by local communities. He calls this one of the random effects that could shake the boom, and says it is one of the things he would worry about.

  “I suppose the modern day equivalent is the political protest against the construction of data centers.”at 32:39

• Liaquat Ahamed

  A major AI buildout player may fail to raise capital, triggering a giant market shock.

  Drawing the railroad parallel, he expects one of the major players in the AI infrastructure build-out to hit financing problems and suddenly be unable to raise capital. He names Oracle and OpenAI as possibilities without committing to either. He expects this to hit the market as a sudden shock, similar to the way the railroad financing failures of 1873 triggered the panic.

  “One of the major players will face financing problems. ... Will suddenly, you know, will suddenly not be able to raise the capital. ... so I think that will come as a giant shock across the market.”at 52:20

AI pricing & unit economics

• Liaquat Ahamed

  Falling token prices and eroding profitability are the warning signs of AI data center overcapacity.

  In 1873, railroad overbuilding showed up as excess capacity and eroding profitability: only a quarter of US railroads could pay dividends by the end of 1872. The analogy for AI is that data center construction is pushing down the price of its output, which Ahamed takes to be tokens, and token prices are falling. He says profitability erosion is the other big factor and is where he would look for trouble, but admits he does not know how to measure it or whether demand for tokens is inexhaustible.

  “I suppose the analogy today is, you know, is all of this data center construction leading to a fall in the prices of their output, which I suppose is tokens, and token prices are going down.”at 23:00

AI bubble & tech valuations

• Liaquat Ahamed

  Past private investment booms all ended in excess capacity, rate spikes and crashes; AI looks similar.

  Ahamed lists four or five US private investment booms: the 1870s railroad booms, the 1920s, the dot-com bubble and the mid-2000s real estate bubble. Each lasted two to four years, each left excess capacity, each saw interest rates spike as everyone competed for capital in the late stages, and each ended in a crash. He leaves the question of whether this time will be different as a rhetorical one, implying the AI boom follows the same pattern.

  “All of them lasted two to four years. All of them ended up with excess capacity. And all of them, interestingly, ended up with a spike in interest rates as everyone was competing for capital in the final stages of the boom, and all of them ended up with a crash.”at 26:49

US Treasuries & yields

• Liaquat Ahamed

  AI capex absorbing world savings is pushing long-end US yields up about 150 basis points.

  Private investment booms absorb such a large share of the world's savings that they push interest rates up. Long-end US rates have already risen about 150 basis points, and by more in other countries. Ahamed says this is typical late in these booms, when everyone competes for capital. The rise in rates then causes knock-on effects, including defaults on other debt.

  “Interest rates have risen, whatever, a 150 basis points in the long end in The US, even more in other countries. ... These booms absorb such a large percentage of the world's savings that it causes interest rates to spike up.”at 31:08

US stock market outlook

• Liaquat Ahamed

  US equities at 250% of GDP make consumption vulnerable to a stock market decline via wealth effects.

  The equity market is now about 250% of GDP, versus roughly half that in 2000. The economic boom has become self-sustaining over the last two years partly because of wealth effects from rising equity prices, a factor that did not drive the economy in the past. If equities fall, the consumption hit from those wealth effects is hard to predict, and it is unclear how the Federal Reserve would manage it.

  “One is the size of the equity market relative to GDP. I mean, the equity market is now 250% of GDP. ... In 2000, the equity market was half the size relative to GDP.”at 41:25

US deficit & government debt

• Liaquat Ahamed

  A 6% of GDP budget deficit means the US has probably run out of fiscal room for the next downturn.

  With the US budget deficit at 6% of GDP, he argues authorities have probably used up their fiscal options. Combined with the Fed's constraints, this leaves policymakers with fewer tools than in the last crises, so the next downturn would be harder to manage.

  “We've got a budget deficit that's at 6% of GDP. ... We've probably run out of bullets in terms of fiscal policy.”at 42:29

US dollar

• Liaquat Ahamed

  Reversal of foreign inflows into US equities could trigger a dollar selloff that complicates Fed easing.

  Roughly $1 trillion moved from foreigners into US equities over the last twelve months. If that money reverses and flows out of the dollar, the currency falls. That would make it hard for the Fed to ease, since easing could worsen the dollar's decline and threaten its inflation goals.

  “A trillion dollars has gone from foreigners into The US equity market in the last twelve months. If that turns tail and there's a flood of money out of the US dollar, what does the Federal Reserve do?”at 42:46

US rates & the Fed

• Liaquat Ahamed

  In the next downturn the Fed faces a dilemma: easing risks inflation, unlike 1929's gold-standard constraint.

  In 1929 central banks were tied to the gold standard and felt unable to ease, which made the crash worse. Today no one is bound to gold, but the Fed is bound by its inflation mandate. Easing a dollar and inflation-constrained economy could worsen knock-on effects, so the Fed's room to act is narrower than in recent crises.

  “This time around, no one's tied to the gold standard, but are we tied to keeping inflation under control? Does that make the knock on effects much worse?”at 46:18

Open-source AI models

• Liaquat Ahamed

  Competition from Chinese open models could suddenly cause a mass erosion in AI industry profitability.

  He sees competitive pressure from open models produced in China as a second possible source of crash in the AI boom. If cheaper open models arrive, they could sharply cut the profitability of the companies that have invested heavily in AI infrastructure. He presents this as one of several scenarios and does not forecast it firmly.

  “Will there be competitive pressures from, I don't know, open mod... Models produced by the Chinese that suddenly causes a mass erosion in profitability.”at 52:53