Artificial Intelligence – Debt Debate pt. III

Excessive leverage usually ends in tears 

• From a purely financial perspective, there is a strong argument to use debt to finance the AI roll-out, but if something goes wrong, the problems will be far greater than they would otherwise have been if equity or cash flow had been used.
• The USA AI capex bill for 2026 is now sitting at around $750bn with more to come in 2027, which is so large that it is having a material effect on the USA’s economic statistics, such as GDP.
• The vast majority of this is being spent by the combination of Google, Microsoft, Amazon and Meta Platforms, all of whom have very strong cash flow and are mostly in a position to make the investments without raising debt.
• However, most are choosing not to finance AI capex with cash flow but instead are choosing to raise debt.
• The reasons for this are rational and make sense, but in this climate, it raises the risk of a correction greatly and also ensures that any correction will be deeper and more painful than it otherwise would have been.
• Every company is financed by a combination of debt and equity, and in most cases, it makes sense to have a balance between the two.
• There are two reasons for this:
  • First, cost efficiency: where, up to a point, adding debt to the balance sheet reduces the cost of financing the company.
  • This is because interest payments can be subtracted from profit before the calculation of corporation tax, whereas dividend payments or share buybacks cannot.
  • Hence, the impact of lowering the tax owed increases the profitability of the company by more than the interest paid.
  • Second, equity dilution: where issuing more shares reduces the stake that existing shareholders have in the company unless they put up more money.
  • Most shareholders do not appreciate being diluted, and so many companies will look to raise debt when they need money rather than conduct an equity offering.
• The net result is that raising debt is a perfectly normal practice of financing the growth of a company, but it is easy to take it too far.
• This is precisely what happened in 1999 and 2000 when the telecom equipment companies raised a lot of debt and lent it to their customers to buy their products.
• When the data traffic didn’t show up, sales fell, and there was not enough money left over to pay down the debt.
• This resulted in great financial distress, a lot of painful restructuring and a declining stock market that lasted for a few years.
• If there had been no debt, there would have been some painful write-downs and substantial valuation corrections, but there would have been no painful and drawn-out restructuring.
• This is why, when I look at Google raising $3.6bn in the Japanese market on top of the $55bn it has already raised in USD and CHF, I get nervous.
• Morgan Stanley estimates that $450bn will be raised in 2026 to pay for AI investments, and at some point, all of this will have to be paid back plus interest.
• The key difference between AI and the Internet Bubble is that demand for AI compute is extremely strong, and one reason for this is that the vast majority of those who use it don’t pay for it.
• This has to change and change soon to avoid a financial correction.  
• This is why it is increasingly urgent that the economics of the data centre are rapidly improved because, without this improvement, there will be no cash to meet the debt repayments.
• Given this problem and the inevitability that supply will exceed demand at some point, the last thing a prudent AI data centre investor should be doing is financing it with debt.
• The economics of Nvidia’s latest Vera Rubin system will be crucial, as Nvidia is promising a 5x increase in revenues per GW of AI capacity under certain circumstances.
• We have heard these sorts of things before, but they have failed to deliver on better economics, as revenue per GW has been stuck at $10bn for the last 3 years.
• I think that we are still quite far from a serious crisis in AI financing, but these things always occur very slowly and then very suddenly and all at once.
• I am steering clear of this situation and instead, I remain happy to stick with memory via Samsung, inference at the edge via Qualcomm and nuclear power.