The spread is the trade. A book that borrows in yen and holds duration abroad is now looking at a domestic alternative at 3% [1] against a US ten-year that peaked overnight at 4.76% [9], which leaves 176 basis points of gross carry [1] to pay for hedging, borrowing spread, and whatever leverage justifies the desk space. Against Korea, where the ten-year added six hundredths of a point to clear 4.37% [11], the gap is 137 [4]; against Australia's five-week high above 5.1% [12] it is wider still, and neither of those is denominated in the currency whose short rate is about to move.
Multiply Totan's 93% [2] by FedWatch's better-than-60% [10] and, treating them as independent, you get roughly 56% [2]: better than even odds that both central banks tighten in the same month, which is the branch where the differential survives and only the cost of the short leg rises. That is the benign case, and it is the modal one on these numbers. The unpleasant branch is the Bank of Japan moving in September [3] into a market already absorbing 3% while the Fed does not follow, because then the gap compresses from the near side and the funded position gets smaller by arithmetic rather than by conviction.
What a 3% domestic yield changes is what Japanese capital is not obliged to do. It is not obliged to reach abroad, not obliged to wear currency risk, and not obliged to lend the funding currency to anyone else's levered equity exposure, which is the quiet channel through which a bond auction touches a data-center trade.
The counter-thesis is stated cleanly by Nomura's Wataru Akiyama, who reads the semiconductor selloff as "little more than a pretext or trigger for selling" after a significant build-up of open positions in margin trading, and who says the earnings-growth outlook for AI and semis has not changed [17]. He may be right, and there is an entirely separate cost channel that has nothing to do with yields: Nvidia has told its largest customers to expect AI servers to cost more than 15% more on rising memory prices [16]. If margins compress because DRAM got dear, no amount of cheap yen would have saved the multiple.
This is probably wrong, but the honest read is that funding cost is the binding constraint here and positioning is the transmission mechanism, not the cause. Three ways it breaks: Akiyama is right and the margin book clears in a fortnight with the ten-year still at 3%, which would show up as AI names rebounding while yields sit still; the Fed hikes alongside the Bank of Japan and the 176 keeps its width, in which case the squeeze is a rent increase rather than a liquidation; or the Bank of Japan holds in September and 93% turns out to have been a price rather than a forecast. Note that the tape is not yet cooperating with any single story, and the source itself attributes the pressure on AI shares to rising financing costs [18]. If the yen leg gets dearer while the gap it funds stays 176 wide, the positions survive at lower size; if the gap narrows, they do not survive at this size at all.
