HBM Remains a Hyperscaler-only Memory Tied Tightly to AI Accelerators

HBM remains a hyperscaler-only memory tied tightly to AI accelerators

SPHBM4 uses serialization and organic substrates to lower HBM packaging costs while keeping bandwidth memory exclusive to hyperscale AI systems.

New SPHBM4 Tech Slashes HBM Costs 30%, But AI Boom Fuels RAM Shortages

In the relentless pursuit of artificial intelligence dominance, a quiet revolution is unfolding in the memory technology sector, where high-bandwidth memory (HBM) has long reigned as the elite fuel for AI hyperscalers. But a new innovation promises to slash costs dramatically, potentially reshaping how these tech giants build their data centers—while casting a shadow over everyday consumers scrambling for affordable RAM. At the heart of this shi…

SCIENCE & TECH: JEDEC’s SPHBM4 plan shifts HBM economics toward cheaper substrates without changing who actually gets to use it today

SPHBM4 cuts pin counts dramatically while preserving hyperscale-class bandwidth performance Organic substrates reduce packaging costs and relax routing constraints in HBM designs Serialization shifts complexity into signaling and base logic silicon layers High bandwidth memory has evolved around extremely wide parallel interfaces, and that design choice has defined both performance and cost constraints. HBM3 uses 1024 pins, a figure that already…

New 'serial' tech will significantly reduce the cost of memory — HBM memory, that is, the sort of RAM only AI hyperscalers can use, but hey, at least they won't go after consumer RAM, or would they?

SPHBM4 uses serialization and organic substrates to lower HBM packaging costs while keeping bandwidth memory exclusive to hyperscale AI systems.

Samsung and SK are planning to begin mass manufacturing of new AI memory chip HBM4 chips in February

Samsung and SK Hynix plan to mass-produce sixth-generation high-bandwidth memory (HBM4) chips designed for artificial intelligence (AI) uses. Samsung is anticipated to start manufacturing in February 2026, followed by SK Hynix in September 2026. NVIDIA’s Vera Rubin AI accelerator system will primarily utilise these processors, providing significant performance and customisation improvements compared to previous versions. According to a report, S…