NuAxiSemiconductor trade intelligence

[News] AI Chip Memory Strategies Diverge: Microsoft’s Maia 300 Fuels HBM4 Demand as Intel Bets on LPDDR5X

What happened

Reported October 13, 2026: Microsoft and Intel on Monday unveiled their latest in-house AI accelerators, stepping up their challenge to NVIDIA. According to Reuters, Microsoft's Maia 300 features a multi-chiplet architecture using TSMC's 2nm (N2P) process for core computing chiplets, with memory controllers and I/O functions on 3nm dies; per Seoul Economic Daily, Maia 300 packs eight HBM4 stacks, up from six HBM3E stacks in Maia 200, raising memory capacity from 216GB to 288GB. Microsoft said the architecture delivers 36% more compute FLOPs, 2.5 times the HBM bandwidth and 14% higher scale-up bandwidth. According to Reuters, Microsoft plans to roll out potentially more than one million Maia 300 accelerators next year, with each rack housing 96 chips and consuming around 167kW, and claims roughly 40% more computing output per megawatt than leading GPU-based systems; per Seoul Economic Daily, total HBM stacks per rack jump 2.7-fold from 288 to 768, potentially opening a major new growth opportunity for memory suppliers. SK hynix is seen as a leading contender (sole HBM3E supplier for Maia 200), and Samsung could also secure orders with its 36GB 12-high HBM4 now in mass production. On Intel's side, per Tom's Hardware, at OCP 2026 Intel officially branded its Xe3P-architecture Crescent Island AI inference accelerator as the Intel GPU D-Series, unveiling performance targets of 1.3 PFLOPS in FP8 compute and 1.5TB/s of memory bandwidth; per Wccftech, it opts for LPDDR5X over the increasingly costly and supply-constrained HBM, supporting up to 480GB of LPDDR5X (reference PCIe card: 160GB), positioned as a cost-efficient option for AI inference; the 350W accelerator's manufacturing process remains undisclosed. These remain unveilings and plans; deployment and orders are not yet confirmed.