[비즈한국] Microsoft (MS) has released its latest proprietary artificial intelligence (AI) accelerator, optimized for inference. As hyperscalers (large-scale data center operators) like Google and Amazon Web Services (AWS) move to reduce their dependence on NVIDIA, Microsoft has joined the ranks of companies developing custom AI chips in-house. This follows two years after the debut of its first AI chip, 'Maia 100,' in November 2023.
On the 26th (local time), Microsoft announced the launch of 'Maia 200,' a next-generation accelerator designed to boost the efficiency of generative AI inference. Built on TSMC's 3nm process, Maia 200 is characterized by improved cost-effectiveness in generating 'tokens,' the fundamental units of AI inference.
Microsoft emphasized that Maia 200 is "the highest-performing semiconductor ever created by a hyperscaler." The company explained that Maia 200's lightweight computing (FP4) performance for generating AI responses is three times higher than that of Amazon's third-generation proprietary chip 'Trainium,' and its computational efficiency exceeds that of Google's seventh-generation Tensor Processing Unit (TPU) 'Ironwood.'
CEO Satya Nadella stated, "Designed for industry-leading inference efficiency, this product delivers 30% higher performance per dollar compared to existing systems."
Upon its release, Maia 200 is expected to serve as the backbone supporting Microsoft’s meticulously built vast AI ecosystem. It will be dedicated to the inference infrastructure for high-spec models such as OpenAI’s next-generation model GPT-5.2 and Microsoft’s own AI assistant 'Copilot,' with a strategy to maximize computational efficiency across the entire Azure cloud infrastructure.
The integration speed into actual hardware is also rapid. Microsoft has already deployed and activated Maia 200 at its data center in Iowa, U.S., and plans to sequentially expand its application to major western hubs, including Phoenix, Arizona. Through this, the company aims to ensure that Azure customers worldwide can stably utilize high-performance services based on proprietary AI accelerators.
Innovations at the system level are also notable. Maia 200 introduces a proprietary two-tier scale-up network design based on standard Ethernet. It provides 2.8TB/s of bidirectional bandwidth per chip, enabling efficient large-scale inference tasks through a massive cluster connecting up to 6,144 accelerators.
Microsoft previously unveiled the Maia 100 in November 2023. However, the Maia 100 was solely for internal use within the Azure cloud and was not offered to external customers, limiting its market impact. Consequently, Microsoft was evaluated as trailing behind competitors Amazon and Google in the proprietary chip market.
With the release of Maia 200 as effectively its first commercial chip, Microsoft is expected to strengthen its leadership in the AI infrastructure market. Based on this roadmap for next-generation accelerator adoption, Microsoft plans to increase power efficiency and reduce Total Cost of Ownership (TCO).
On this day, following domestic media reports that SK Hynix000660 would be the sole supplier of HBM (High Bandwidth Memory) for the Maia 200, SK Hynix’s stock price turned to an upward trend despite remarks by U.S. President Donald Trump regarding the "restoration of reciprocal tariffs." This is interpreted as a result of the large order windfall and earnings expectations, as it was confirmed that the company secured a supply advantage for the HBM3E volume going into the Maia 200. As of 11:30 a.m. on the 27th, SK Hynix stock is trading at 775,000 won, up 5.29% (39,000 won) from the previous trading day.