Nvidia halves the memory planned behind each Vera CPU

Put last year's Vera figure next to this year's and the cut divides too cleanly to be a design refinement. The planned 1.5TB per CPU is 1536GB in binary terms, and the new number, 768GB, is exactly half of it ^[1][15]. The Rubin Ultra change has the same shape: going from 16-high HBM4E stacks to 12-high or 8-high removes a quarter or a half of the dies in a stack ^[2][16]. JPMorgan's phrase for this is content optimization as a coping mechanism, and the bank is explicit that the systems did not suddenly need less memory ^[3].

The same allocation runs down the price list. Omdia's Runar Bjorhovde puts memory components at four or five times their cost of about a year ago ^[4], and says memory can now be more than half the bill of materials in some cheaper devices, against 10% to 15% before the spike ^[5]. That is roughly a threefold to fivefold rise in the share of a device that one component eats ^[17]. A budget handset has nowhere to put that, so the answer is a 4G radio, which CCS Insight's Ben Hatton says gets away with less memory than an equivalent 5G design ^[7], alongside reduced storage and older Qualcomm or MediaTek parts ^[8]. Hatton's read on the sub-$500 tier is that buyers will get a less powerful product ^[9]. New capacity does not arrive inside the planning window either: plants take two to three years to build, and are increasingly specified for higher-return GPUs rather than basic components ^[6].

Here is what users actually do while that plays out. Samsung's biggest seller by volume in Europe is the Galaxy A16 4G, a cheap model already around 18 months old, according to Bjorhovde ^[10]. Xiaomi shipped about 25% fewer units year over year in the first half of 2026 while revenue fell only about 5% ^[11], which works out to roughly 27% more revenue per unit sold ^[18].

Here is what hardware plans tell themselves: that a part a generation out is a superset of the one you can order today, so capacity per socket is a floor you build on. What the Vera and Rubin Ultra revisions show is capacity per socket being set by what the supplier can obtain ^[3]. The forcing function for anyone with a 2027 refresh in a spreadsheet is two questions per line item. First, is the capacity number in a configuration you can order and take delivery of, or in a slide. Second, if the shipped part lands at half the planned memory, does the workload fail or does it just run slower. A slide number attached to a workload that fails is the cell to re-plan this quarter. An orderable number attached to a workload that fails means buying earlier than you wanted. Everything that merely runs slower can wait and absorb the cut.

Neither analyst gives a date when a spec sheet becomes a reliable input again, and they are not measuring the same clock: Bjorhovde doubts meaningful price declines within 18 months, while Hatton expects the squeeze to last until at least mid-2028 ^[12]. Bjorhovde's own framing is that at this scale he has not seen anything like it before ^[14]. For the person defending a fleet or cluster budget on Friday, the usable version is to write the out-year plan with less capability per dollar and book any restored spec as upside.