SK hynix has started shipping HBM4E samples as Samsung and the AI chip crowd turn memory into a knife fight.
The AI boom has pushed DRAM makers to accelerate their roadmaps because nobody wants to be the outfit left selling yesterday’s memory. SK hynix is racing Samsung to get HBM4E into the hands of next-generation data centre customers.
HBM4E will be a key part of future AI platforms, including Nvidia Rubin Ultra and AMD Instinct MI500. Those chips are expected to bring in serious revenue, so the memory suppliers are clustering around them like gulls near chips.
SK hynix started sampling the parts only weeks after Computex 2026. At the show, it previewed HBM4E with up to 48 GB capacities and speeds of up to 16 Gbps.
Samsung was strutting around the same show with its own HBM4E and HBM5 technology. Its pitch included HPB, or Heat Path Block, because even acronyms need cooling now.
SK hynix says the 12-layer HBM4E improves performance and efficiency. The product offers a maximum data processing speed of 16Gbps per pin and power efficiency up more than 20 per cent from previous models.
In a statement, the company said it had delivered samples of the 12-stack HBM4E on schedule thanks to its advanced HBM development and production expertise for HBM.
“We will work closely with partners for mass production in a timely manner,” it said.
The faster memory is aimed at AI training and inference, where bandwidth bottlenecks can make expensive accelerators look rather less clever. HBM has become the part nobody can get enough of, which is handy for memory makers after years of cyclical misery.
The new HBM4E reduces transfer latency through a newer interface and design optimisation. SK hynix says it remains stable in high-bandwidth environments, which is useful when data centres are being asked to act like industrial furnaces with invoices.
SK hynix is using Advanced MR-MUF technology to hit 48 GB capacity in a 12-layer stack while keeping the package structurally stable. It says heat resistance is 17 per cent better than HBM4, which should help in high-performance computing boxes that already run hotter than a CFO near a capex spreadsheet.
The company has built its pitch around its track record in HBM3, HBM3E and HBM4. That matters because AI chip suppliers do not just want flashy specs, they want memory they can qualify without losing a year and several board meetings.
Samsung, is hardly sitting on its hands. It has been pushing its own 12-layer HBM4E samples, turning the next AI memory round into another Korean semiconductor punch-up.
Whoever wins HBM4E design slots gets a fat slice of the next AI server buildout, while everyone else gets to explain why “strategic positioning” did not pay the bills.
