Kioxia and Sandisk lift Japanese flash spending to 2.6 times their 25-year pace

Count that 60 percent in dollars and it comes to 62, $31bn against the $50bn the joint venture has put into Japan over 25 years ^[4][2][12]; count it in yen and it comes to 56, 5 trillion against 9 trillion ^[13]. The gap is the exchange rate: the new package implies roughly 161 yen to the dollar, the quarter-century total roughly 180 ^[23]. A yen capex plan quoted in dollars has a moving part in it.

The comparison I would rather use is the run rate. Fifty billion dollars over 25 years is about $2bn a year ^[19]; $31bn over six years is about $5.2bn a year ^[20], or 2.6 times the pace the partnership has sustained across its life ^[21]. That multiple, not the press release's language about meaningful multi-year bit growth ^[11], is the actual claim being made about AI storage demand.

What has been committed is smaller than $31bn in two specific ways, neither of them hidden. Fab3 is 1.8 trillion yen, 36 percent of the package ^[25][28], and Kioxia said the construction schedule and equipment spending for the new facility will be determined by market trends ^[7]; the whole thing is contingent on government support ^[4], with roughly a third of the total, call it $10.3bn, sought from Tokyo after both chief executives met the prime minister ^[26][27]. So a third of the money is timing-optional and a third of the funding is somebody else's appropriation. The joint venture itself was extended in January 2026 to run through December 2034 ^[14], which at least gives the spend a contractual home well past Fab3's start.

Note where the money is not going. Both plants already exist ^[3], so this deepens Mie and Iwate rather than diversifying geography, and all of it sits on the long-term storage side of memory rather than DRAM ^[18], while SK Hynix committed about $38bn to Korean semiconductor capacity earlier in the same month ^[15]. That is $69bn of announced memory capacity inside four weeks ^[24], from a pair who have historically been the second-largest NAND producer behind Samsung ^[17].

Where this goes turns on who blinks first. Tokyo funds its third, agentic demand holds up the way Kioxia says it will ^[8], and by 2030 the $31bn reads as conservative. Or the subsidy negotiation drags, Fab3 slips past the fiscal year starting April 2029 ^[6], and flash stays tighter for longer than the announcement implies, which is the version that suits the two announcers best. Or demand cools around 2028, the market-trends clause gets used, and 5 trillion yen becomes three-point-something. This is probably wrong, but I would weight the third highest, on the grounds that the clause was written for a reason. What would change my mind is dull and checkable: a named line in a Japanese budget, firm tool orders at Kitakami, and a Sandisk capex disclosure consistent with half of a $5.2bn-a-year pace rather than merely with the financial guidance it says this already fits ^[9].