Microsoft Courts SK Hynix as its Maia 200 AI Chip Raises Memory Stakes

TL;DR

Summit Talks: SK Hynix chief executive Kwak Noh-Jung is expected to meet Bill Gates and Satya Nadella at Microsoft’s May 12-14 CEO Summit in Redmond. Chip Constraint: Microsoft says Maia 200 uses 216GB of HBM3e and more than 10 petaFLOPS at FP4, making memory supply a practical scaling limit. Supplier Stakes: SK Hynix is reportedly tied to Maia 200 supply, so any deeper partnership could help Microsoft expand inference capacity beyond NVIDIA-heavy deployments.

SK hynix chief executive Kwak Noh-Jung is expected to meet Bill Gates and Satya Nadella during Microsoft’s CEO Summit this week. 

With the summit underway, SK hynix appears to be aiming at a sole-supplier role for Maia 200, Microsoft’s in-house AI inference chip. No Maia 200 supply agreement is public. 

Kwak’s Redmond trip also fits a longer timeline. SK hynix was already at Microsoft’s CEO Summit in 2024, and Microsoft’s June 2025 Braga setback already showed how difficult it is to turn an internal accelerator program into dependable datacenter capacity.

How Maia 200 turns the meeting into a hardware story

Microsoft already places Maia 200 in its heterogeneous AI infrastructure for Foundry and Microsoft 365 Copilot workloads. More HBM supply would let Microsoft widen that hardware lane across more inference demand instead of leaving Maia 200 as a narrower in-house option.

As Maia 200 moves from launch messaging into deployment, Iowa rollout details and a reported 30% better performance per dollar gain already pushed the chip beyond a launch concept earlier in 2026. Lower cost per inference only matters at scale, though, if Microsoft can secure enough memory, packaging capacity, cooling, and board-level supply to build out more complete systems.

January’s Maia 200 product post gave Microsoft one direct line on what the accelerator is supposed to do inside that broader AI stack.

“[Maia is] a breakthrough inference accelerator engineered to dramatically improve the economics of AI token generation.”

Scott Guthrie, Microsoft executive (via The Official Microsoft Blog)

Maia 200 uses 216GB HBM3e at 7 TB/s. Stacked high-bandwidth memory keeps large AI models fed with data fast enough to avoid turning expensive compute into idle waiting time. That hardware profile helps explain why a memory supplier can matter almost as much as the chip design itself.

Microsoft’s earlier Maia generation used far less memory, so the 2026 design marks a meaningful jump in how much data each accelerator can keep close to the compute cores. Bigger memory pools reduce the need to shuttle model state across slower links, which is especially valuable when Microsoft wants to run heavier inference workloads efficiently.

Each Maia 200 chip is specified at more than 10 petaFLOPS at FP4 within a 750W envelope. Thermal design, packaging yield, memory availability, and rack-level integration decide how much of that per-chip output becomes usable fleet capacity. Supplier depth affects cost and also shapes how quickly Microsoft can translate a paper advantage into real inference throughput for Copilot and Foundry customers.

Where Microsoft sits in the AI chip field

As Microsoft moves deeper into custom silicon, Google’s TPUs and AWS Trainium already serve training and inference workloads inside rival cloud stacks, while NVIDIA still sets the broader GPU benchmark that hyperscalers are trying to offset rather than fully replace.

Because provider demand is shifting, 44.6% growth in custom AI ASIC shipments in 2026, versus 16.1% growth for GPU shipments points to why cloud providers keep investing in internal chip programs. Faster ASIC growth raises the value of reliable memory supply because more providers are trying to secure specialized components for their own hardware roadmaps at the same time.

Against that backdrop, supplier access can become a strategic issue before a cloud company fully replaces outside GPUs. Microsoft does not need Maia 200 to displace NVIDIA everywhere at once for SK Hynix to matter. It only needs enough memory and packaging support to widen the set of Copilot, Foundry, and other inference workloads that can move onto its own silicon.

That supplier picture sharpens the pressure because SK Hynix is not negotiating with Microsoft in isolation. The company has been linked to HBM for NVIDIA GPUs and to Google and AWS partnerships, leaving Microsoft to court a supplier that already sits inside several competing ecosystems.

Redmond’s meetings matter for that reason even without a signed deal. If Microsoft eventually secures more HBM3e supply around Maia 200, it could expand inference capacity on hardware it owns more directly. If supply stays tight, the summit may simply become another step in Microsoft’s slower effort to reduce how much it depends on NVIDIA-heavy deployments.