@Elliptica@poa.st @merchantHelios@poa.st I don't follow the memory industry closely (maybe I should change that

) so I just discovered Samsung is claimed to have gotten its HBM act together (South Korean news sources can be ... bad).
It had previously failed Nvidia's tests multiple times for HMB3E, which is projected to be majority of 2026's HBM production, is for example in Nvidia's Blackwell datacenter GPUs, and the GB300 third generation uses 1.5x the HBM of the GB200. Those two run so hot they dissipate up to 135 kW per rack using liquid cooling all the way to inside the GPU assembly

And now it's claimed Samsung has passed Nvidia's tests for HBM4, and will start production in February. That said, it'll be many months before any of these show up in customer's hands, these are for example going to be in the next Nvidia Rubin generation, scheduled for 26Q3 but Nvidia is aggressive in setting timetables.
Then again, this is both very hard and complicated. Due to limits in TSMC's simulation/verification process for its masks, with these dies being a big as possible (max first EUV generation), Blackwell volume production was delayed because yields were lower than they should have been. Requiring investigation, design and manufacture of some new masks, and then the months long process starts to make the new ... not really a stepping maybe??
Nvidia's design transistors, wires etc. didn't need modification which is what we usually refer to as a stepping, just TSMC's implementation of it....