|
|
The US AI lead over China is measured in single digits. The costs to defend it are not. |
|
One Number
The capability gap between Anthropic’s leading model and China’s DeepSeek. Stanford report, cited by Fortune/ September 2026 That is the edge achieved against a country locked out of the Nvidia chips and now defended with more than $3.1 trillion in commitments. |
|
|
One Argument When looking at the 2.7% gap, and the cost of protecting it, the AI race stops looking like a story about whose model is better.China has no avenue open to buy its way to parity. The US export controls block it from the Nvidia chips that train and run the West’s most advanced models. Instead Chinese labs have found a cheaper route to a similar result. This technique focuses on training a new, cheaper model by learning from the outputs of an expensive one already built rather than assembling the raw computing power to train from scratch. This method is called distillation. Two of the resulting models, GLM 5.2 and Kimi, now handle roughly 75% of the same engineering tasks as their US equivalents at one-fifth of the cost, according to Futurum Group analysis, cited by Fortune. A country without the chips getting within 2.7% of one with them is not a small gap. It should worry you. Even holding that 2.7% edge is not cheap, and increasingly not visible either. Meta, Broadcom and Nvidia have each adopted a financing tool called a residual value guarantee. It is a promise from a strong-credit company to make a lender whole if a data centre or a batch of chips is worth less than expected when eventually sold. The guarantee lets a special purpose vehicle carry the actual debt. It is a separate entity set up just to own the infrastructure, so that the tech company does not have to take the debt on its own balance sheet. Meta's $28 billion guarantee behind its Hyperion data centre in Louisiana helped raise $27 billion of debt priced within 150 basis points (1.5 percentage points) of Meta's own borrowing costs. Morgan Stanley's Richard Myers, who designed the structure, says it does this while barely touching Meta's own balance sheet. Morgan Stanley’s own analysis separately now counts more than $3.1 trillion of these kind of commitments across just seven hyperscalers and chipmakers, according to the Financial Times report a few days ago. A thin gap today does not mean a thin gap tomorrow. Western labs are betting that raw compute compounds in a way distillation cannot indefinitely match. If frontier compute is what unlocks the next jump in capability, paying up for it now is simply the price of staying in the race. However, S&P’s Pierre Georges says the safety margin behind these guarantees is now "closer to zero most of the time," and KBRA's Doug Colandrea points to a "very high degree of complexity" built into Big Tech's credit risk. A shrinking margin doesn't prove the bet is wrong — it proves the market isn't sure it's right.
|
|
One Position The US AI lead is smaller than it looks. And the money defending it has very little room to be wrong. If you hear the AI buildout described as unstoppable, the more useful question is what happens if it merely disappoints. Right now, few of the people financing it seem prepared for that. I will change my mind if banks started demanding bigger guarantees before lending into new AI data centres. That would mean they see real safety margin in the assets, not just confidence that demand keeps going. Where in your own company’s AI spending would you want to ask “what happens if this doesn’t pay off” — and has anyone actually asked that yet? |
|
If this was forwarded to you and you'd like to receive Corporate Financier's Notes every Thursday, you can subscribe here. Reading on your phone? Tap the three dots or overflow menu in the top corner of your email app and select View in browser for the best experience.
|