Microsoft Azure

Full briefing

Microsoft Corporation (NASDAQ: MSFT) operates Azure, the world's second-largest cloud platform by revenue, trailing only AWS. Its January 2023 $10 billion investment in OpenAI (extended from the original 2019 partnership) locked in Azure as the exclusive provider of compute infrastructure for ChatGPT, GPT-4, and subsequent OpenAI models — instantly making Azure the most GPU-hungry cloud on earth. In 2023–2024, Microsoft deployed an estimated 150,000–300,000 NVIDIA H100 GPUs across its Azure AI infrastructure, in clusters designed for large-scale transformer model training. These are assembled into complete server nodes by ODMs including Supermicro, Wiwynn, and Quanta Computer, cooled by Vertiv liquid cooling systems, and connected via Broadcom or NVIDIA high-speed networking. Recognizing the strategic and cost risk of GPU dependency, Microsoft developed two custom silicon chips unveiled at Microsoft Ignite 2023: (1) Maia 100 — a custom AI training and inference ASIC designed in collaboration with Broadcom and manufactured by TSMC on N5 (5nm). Maia 100 is optimized for transformer model training and is deployed in dedicated Azure Maia 100 instances; and (2) Cobalt 100 — an Arm Neoverse V2-based CPU for Azure general compute workloads, replacing Intel and AMD server CPUs in some Azure instance types. Microsoft's AI infrastructure strategy embodies a trend across all major hyperscalers: using NVIDIA GPUs for frontier model training while progressively shifting inference and commodity compute to lower-cost custom silicon. This puts Microsoft in direct supply chain relationships with TSMC (Maia fabrication), Broadcom (ASIC design collaboration), and the Japanese ODM/materials ecosystem (server assembly and components).