Microsoft’s latest Maia 300 chip, unveiled this week, promises a 40% boost in token‑per‑megawatt efficiency over leading GPUs and is slated for production in 2027 with a multi‑gigawatt rollout. The company plans to make Maia the primary inference engine across its services, marking a major shift toward in‑house silicon.
At the same time, AMD rolled out the Helion blueprint, an open‑compute rack that can house up to 32,768 Zen 6 cores. The design, targeted for a November 2026 release, underscores a move to massive, CPU‑centric AI clusters that can handle agentic workloads.
Contrasting these pushes, Nvidia’s China shipments dropped 44% YoY in the first half of 2026, shrinking its market share to 34% as U.S. export restrictions fuel domestic competition. This slump highlights the growing fragmentation of the AI‑chip market and the strategic importance of regional supply chains.
Analysts note that inference will dominate roughly 60% of AI compute by 2030, driving demand for custom accelerators, high‑bandwidth memory, chiplet integration, and silicon‑photonic interconnects. The rise of silicon photonics, especially as the U.S. and China vie for optical‑transceiver leadership, adds another layer of complexity to chip‑design roadmaps.
What to watch next:
- The scale‑up of Microsoft’s Maia 300 in 2027 and its impact on GPU‑dominant data‑center economics.
- AMD’s Helion rack adoption and performance benchmarks once the November release is live.
- How Nvidia adapts to its shrinking China footprint and whether new packaging or photonic solutions can recover market share.
- Continued policy shifts around export controls and their effect on the global silicon‑photonic supply chain. [[1, 2, 3]]