Microsoft meets SK hynix on AI chips

AI chips are the story here, but the real choke point is memory. Microsoft can design more of its own silicon, but those chips still need huge amounts of high-bandwidth memory to do useful work. That is why SK hynix CEO Kwak Noh-jung showing up at Microsoft’s CEO Summit in Redmond matters. He is expected to meet Bill Gates and Satya Nadella this week, and the point is pretty clear — deepen a relationship that already sits inside Microsoft’s AI infrastructure. ### Why is SK hynix in this story? SK hynix is not just another chip supplier. It is the company with the strongest grip on HBM, the stacked memory used beside top-end AI accelerators. SK hynix itself says it held 62% of HBM shipments as of Q2 2025 and 57% of revenue as of Q3, which is why every large cloud company has to care about its roadmap. (en.yna.co.kr) ### What does Microsoft already buy from it? The existing link is Maia 200, Microsoft’s in-house AI accelerator for inference. Microsoft says Maia 200 is already deployed in its US Central datacenter region near Des Moines, with more regions to follow. Industry reporting says SK hynix is the sole HBM3E supplier for that chip family, which makes this less like a first date and more like a supply-chain check-in with expansion on the table. (news.skhynix.com) ### Why does memory matter so much? Because AI accelerators are basically starved without it. The processor can be brilliant, but if data cannot move in and out fast enough, performance falls apart. HBM is the expensive, hard-to-scale layer that keeps giant models fed. There are only a few companies that can make it at volume, and ramping capacity takes time, so memory becomes a strategic bottleneck — not a commodity add-on. (blogs.microsoft.com) ### Is Microsoft trying to replace Nvidia? Not exactly. The better way to read this is optionality. Microsoft’s own Maia 200 materials pitch the chip as a way to improve inference economics inside Azure, not as a declaration that Nvidia is out. The company is building a mixed fleet — some Nvidia, some custom silicon — because AI demand is too big and too expensive to leave on one architecture. That makes deeper SK hynix ties a hedge as much as a partnership. (siliconanalysts.com) ### Why now? Because custom AI chips have moved from lab project to production infrastructure. Microsoft introduced Maia 200 in January 2026 and has already put it into live Azure deployment. Once that happens, supply questions get real fast — volume, yield, packaging, future generations, and who gets priority when HBM is tight. A summit meeting at this level usually means the conversation is no longer just technical. (blogs.microsoft.com) It is commercial and strategic too. ### What does SK hynix want from this? Longer visibility. If Microsoft is serious about scaling Maia and future chips, SK hynix wants to lock in demand across multiple product cycles. That matters because HBM capacity is precious, and suppliers would rather allocate it to customers with durable roadmaps than to one-off buyers chasing the latest hype. This is also a way for SK hynix to stay embedded as hyperscalers diversify beyond standard GPUs. (blogs.microsoft.com) ### So what is the bottom line? This meeting is really about control. Microsoft wants more control over AI costs, supply, and performance. SK hynix wants to be the memory partner inside that shift. If the tie gets deeper, it strengthens a broader industry move away from buying one company’s complete AI stack and toward mixing custom chips, outside foundries, and scarce premium memory into something the cloud giants control themselves. (news.skhynix.com) (koreaherald.com)