Index 2025 spending at 100 and the forecast becomes legible: 98% growth puts 2026 at 198, another 50% puts 2027 at 297 [1][1], and the published shares put memory at 93 of that 198 and then 202 of that 297 [2][2]. Read it from the increment rather than the level, or rather from the more useful version of the same sum, and the total rises 99 index points while the memory line rises 109, so memory takes about 110% of the new money and then starts on the money that was already there [4].
Then the bits, which is where the compute actually lives. TrendForce has combined server DRAM and HBM bit supply up 27% in 2027 on process migrations and second-half fab ramps [8]; set 2.17 times the memory dollars [2] against 1.27 times the bits and the buildout is paying roughly 71% more per bit than it did the year before [5]. That division is crude, since the bit figure is industry supply while the dollar figure is buyer capex and includes NAND, but the sign is not in doubt, and it is the whole reason a 50% capex print stops working as a proxy for accelerators installed.
Worth noting what the note does not settle: it assigns memory a share of capital spending without saying whether a cloud provider writes that check to a memory maker or pays it inside the price of a server or an accelerator [14], and that distinction decides who has the leverage. TrendForce's own first implication points to the answer it fears, namely that elevated memory costs give server and AI chip suppliers, NVIDIA among them, justification to raise prices, after which buyers must lift capex again just to hold their target chip volumes [10].
There are three ways the 68% does not hold. Price ceilings written into long-term agreements from the second quarter of 2026 onward could bite harder than assumed, though HBM contracts are still modelled up 70 to 140% in 2027 [5][6]. Buyers could take the second route TrendForce sketches and cut memory capacity per system, re-specifying RDIMM configurations, trimming HBM per chip, or pushing work onto ASICs with the model architecture hardwired into the silicon [11][12]. Or boards simply approve more, in which case the share holds and the non-memory pool grows again in absolute terms.
This is probably wrong, but I would take the de-spec branch. Procurement organisations facing contract moves of the size logged in the second half of 2025 [3][4] change the bill of materials well before they change the budget, which would put 2027 memory share under 68% and surface the cost as less capacity per accelerator rather than as a bigger capital plan. What kills that thesis: 2027 landing at or above 68% with server DRAM and HBM bits up the forecast 27% and no reported change in per-system configuration, which would mean the buyers paid the price and kept the spec.
