SoftBank’s young memory venture has found a Japanese packaging partner for its attempt to break into a market ruled by three Korean and American giants.
Saimemory, the memory subsidiary SoftBank set up in December 2024, signed a memorandum of understanding with Shinko Electric Industries on September 28 to jointly develop and commercialize Z-Angle Memory, a next-generation memory technology the two companies call ZAM. The design aims for high bandwidth and low power consumption, and Saimemory is positioning it squarely at the AI server market.
The target is the same one that has made high-bandwidth memory the hottest corner of the semiconductor industry. AI accelerators are starved for the memory that feeds them, and the companies that supply it, Samsung, SK Hynix and Micron, have turned that shortage into enormous profits. Those three control virtually the entire market for the memory used in AI servers, and much of their output is already spoken for by the chipmakers that build the accelerators. Saimemory wants to cut into that share with a different approach.
ZAM is built on stacked DRAM, the same foundation as high-bandwidth memory, but Saimemory argues its architecture can deliver the bandwidth AI servers need while drawing less power, a combination that matters when a single AI data center can consume the electricity of a small city. The technology was first made public in February, when Saimemory signed a collaboration agreement with Intel, and the two companies have talked about a goal of reaching commercialization around 2030.
Intel’s role in the project says something about the stakes. Intel was once a memory maker itself and still retains the design expertise, even after selling most of its memory business. For Saimemory, the Intel partnership brings credibility and a path into the data-center customers Intel already serves. For Intel, it is a way to stay relevant to a technology that could eventually compete with the high-bandwidth memory its rivals supply.
Shinko brings a different missing piece. The company, one of Japan’s largest makers of semiconductor packages and substrates, specializes in the physical work of connecting a chip to the outside world. It traces its roots back to 1946 and grew up inside the Fujitsu group, and its packaging is used in everything from processors to communications equipment. Memory designed for AI servers has to be stacked and packaged with extreme precision, and that is exactly the craft Shinko has practiced for decades. The deal gives Saimemory a route from design to a product that can actually be assembled and shipped.
The memorandum is a framework, not a contract. Neither company disclosed a production timeline or how much money it intends to spend, and memorandums of this kind can fall apart before any chip is built. What it signals is that Saimemory is moving from research into the long, expensive phase of turning a technology into manufacturing.
The odds are long. Samsung, SK Hynix and Micron have spent billions and years building the factories and supply chains that dominate memory today, and they are not standing still. SK Hynix said in the same week that its next-generation memory had passed joint validation with TSMC, a sign of how deep the incumbents’ ties to the AI ecosystem already run. A newcomer must match the technology and also persuade chipmakers and server builders to switch suppliers, a decision no one makes lightly for a component as critical as memory.
SoftBank is betting that the AI boom is big enough to reward a challenger. The company has tied its own AI ambitions to the deal, and a memory product designed for AI servers would give it a claim on a market that has largely been controlled by suppliers it does not own. Whether the incumbents can be dislodged is another matter.
The market Saimemory is aiming at is large, and lopsided. High-bandwidth memory is the single most constrained part of the AI supply chain, and the companies that make it have been sold out for quarters ahead. Nvidia’s accelerators each carry stacks of the memory, and the chipmaker’s own growth has been limited at times by how much its memory suppliers can ship. A third credible source of that memory, even one that arrives years from now, would redraw the market. Saimemory’s bet is that ZAM can be that source, arriving just as demand for AI memory is expected to keep climbing.
The partnership with Shinko is a step toward an answer, but only a step. A framework agreement with no numbers and no date is a long way from a factory, and the companies still have to show that ZAM can be produced at a cost and a scale that matters. The open question is whether a new architecture and a trusted packager are enough to crack a market where three companies have spent a decade digging in.
