The AI build-out has become a return-on-capital bet

The American AI build-out has stopped being a capability race and turned into a return-on-capital question, and Jefferies' latest warning names the math. In the week ended July 19, leading Chinese models processed 36.39 trillion tokens on OpenRouter against 7.39 trillion for the top US models, a nearly five-fold gap that did not exist in late April, when Chinese models were running at 4.37 trillion.

That ratio is what makes the capex scale matter. The four largest US hyperscalers are projected to spend roughly US$695 billion in 2026 and US$870 billion in 2027, with Alphabet alone lifting its 2026 guidance by another US$15 billion to a US$195 to US$205 billion range. The thesis Jefferies is publishing is that open-source Chinese systems, anchored by Moonshot AI's Kimi K3 release, are now compressing the price floor of inference compute, the exact cost curve that super-cycle is being sized against. Hyperscalers are simultaneously the largest issuer of US investment-grade debt, which is how the bill gets paid if the revenue ramp slips.

Both readings are real: the US still leads frontier capability, and the build-out may still earn its keep. The frame that survives the evidence is simpler. When the marginal cost of running a competitor's model falls toward zero, the return on a US$870 billion build stops being a technology bet. It becomes a debt bet.