In the same week, SpaceX's cost of insuring its debt hit a record and Amazon sought to move $8bn of Nvidia chips off its books. Both are funding the same bet on AI, but they carry the risk in opposite ways, and credit markets are starting to price the difference.

Two deals, one bet

SpaceX is seeking $40bn to buy Nvidia chips: $10bn in bank loans and $30bn in investment-grade debt. After the FT reported the plan, the five-year credit default swap (CDS) on its debt rose to 194 basis points, a record. That means $19,400 a year to insure $1mn of its debt against default, up from about $11,000 when the contract began trading in June. Its 2056 bond now pays 2.36 percentage points over US Treasuries, up from 1.75 in June (FT, citing MarketAxess data).

A few days earlier, the FT reported that Amazon was looking to offload $8bn of Nvidia chips to investors.

They are not isolated. Oracle's CDS is at 244bp, Nvidia's has climbed from 45bp to about 81bp since early July, and the contracts for Alphabet, Meta and Microsoft are close to record highs (FT). Amazon, Alphabet, Microsoft and Meta are expected to spend more than $700bn on AI this year (Reuters), and AI-linked debt sales are running at roughly $335bn globally, more than double 2025 (Bloomberg).

Amazon's way: move the chips off the balance sheet

As I read the FT's report, it works through a special purpose vehicle (SPV): a separate company set up for one job. The SPV borrows from lenders, buys the chips and leases them to Amazon. Lenders are repaid from the lease payments and from the chips themselves. On Amazon's own balance sheet, the chips show up as a lease rather than as debt.

To me, this is financial engineering to keep the balance sheet light. The hidden risk is the backstop. If Amazon guarantees the value of the chips, a fall in chip prices comes back to Amazon as a liability, and it only becomes visible when it is called. The risk has not gone away. It has changed form.

The interest rate backdrop probably plays a part. With the Federal Reserve not signalling cuts, plain corporate bonds are expensive, and a structure that gives lenders a claim on specific assets can lower the yield they ask for.

SpaceX's way: carry the risk, with losses

SpaceX does the opposite. The debt and the chips sit on its own balance sheet. At 30 June 2026, weeks after a record $75bn IPO that valued it at about $1.77 trillion (Reuters), it reported roughly $193bn of assets against about $40bn of debt (balance sheet data). It had also issued $25bn of investment-grade debt in June (FT). On paper, another $40bn is manageable.

That was my first reaction too: $40bn against a market value of about $2.2 trillion (StockAnalysis, 9 October close) looks small. But lenders do not lend against a share price. They are repaid from cash flow, and SpaceX is losing money: over the last twelve months it earned about $23bn of revenue and lost about $8.9bn (StockAnalysis). Its AI business reported $3.2bn of revenue in 2025, while Starlink's connectivity business made up 61% of the total (Satellite Today, from the S-1 filing). The chips are a bet on AI revenue that has not arrived yet.

So SpaceX has two problems that Amazon does not. The chips lose value quickly, and there is a repayment question on top. Morgan Stanley's Adam Jonas said some investors doubt SpaceX will succeed as a major provider of frontier AI models (FT). The CDS market is saying the same thing in numbers.

Two notes on scale. 194bp is still modest, below Oracle's 244bp: the market is saying "riskier", not "danger". And these figures come from public filings as summarised by data providers, so treat them as approximate.

What my MBA taught me to ask

During my Executive MBA at POLIMI Graduate School of Management, the lesson that stayed with me is to ask what the money is repaid from, and where the risk really sits. Five questions follow from that.

  • Debt capacity, not valuation. A market value in the trillions is what investors will pay for the shares. Bondholders care about cash flow to cover interest.
  • Asset life versus debt life. The bonds run for decades. The chips become obsolete in a few years. When the asset dies before the debt, the mismatch lands on someone.
  • Substance over form. An SPV with a value backstop may move the debt off the balance sheet without moving the risk. It becomes a contingent liability.
  • Herding. Each company spends because the others do, and nobody wants to be the one that fell behind. That is how capital races build up.
  • What CDS really says. A rising spread shows demand for protection. It is a price for risk, not a forecast of default.

Is it a bubble?

I do not know, and I would treat anyone who is certain with caution. Here is how the evidence looks to me.

The case for worry. CDS spreads are rising across the companies borrowing most for AI, and several are near record highs (FT). I see echoes of the period before 2008: complex structures built to spread risk, high credit ratings on the senior pieces because a strong sponsor stands behind them, and insurers and pension funds among the buyers. It is not the same and not on that scale, but the ingredients feel familiar.

The case against. Amazon earns real money: $27.5bn of operating income in a single quarter (Fi-desk). Its $25bn bond sale in July drew $62bn of orders, even if that was about half the demand of its March deal (Bloomberg). Spreads at 194bp or 244bp are elevated, not distressed.

Why it reaches Australians. About 55 per cent of Australia's overseas share holdings are in the US, down from a 2024 peak of 57.7 per cent, and super funds drove much of the recent net purchases (ABS). My concern is that the US market is heavily weighted to a handful of AI-exposed technology names. A correction would probably not be a calamity, but it would show up in super returns.

Every chip needs a building

This is where the story comes home for me. Behind every chip is a data centre: land, power, steel, concrete, cooling and commissioning. The $700bn of AI spending is a construction programme before it is anything else, and the debt is repaid only if those projects are delivered on time and on budget.

A data centre that opens late is a pile of expensive chips losing value in a warehouse.

I spent a decade on metro and rail projects, where lenders have long used project companies and ring-fenced structures. They learned the discipline the hard way: independent technical review, schedule risk analysis, and money released against milestones. The lesson was that financing risk and delivery risk are two sides of the same problem. A structure is only as good as the schedule and cost certainty behind it.

There is also a contrast for founders. The hyperscalers and SpaceX need tens of billions to build the physical layer. A company working at the application layer can create value from AI without carrying that balance sheet, and that discipline is worth remembering when capital is cheap.

Takeaways

  • Same bet, different risk. Amazon moves chip risk toward lenders and investors through a structure. SpaceX holds it directly, while losing money.
  • Ask what repays the debt. Valuation tells you what the shares are worth. Cash flow and asset life tell you whether lenders get paid.
  • Watch the delivery chain. The AI trade is financed in capital markets and built on construction sites. Delay and cost overrun are credit risks.