{
  "id": 14038064,
  "title": "Microsoft debuts multichiplet Maia 300 inference accelerator",
  "url": "https://urgent.news/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-10-12T20:11:12.000Z",
  "source": {
    "name": "SiliconANGLE",
    "slug": "siliconangle",
    "url": "https://siliconangle.com/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator/"
  },
  "original_language": "en",
  "account": "Microsoft showcased its latest artificial intelligence accelerator, Maia 300, at an event today. The chip is specially designed for inference tasks, with Microsoft estimating it can handle 40% more tokens per megawatt than leading GPU-based systems, likely referring to Nvidia products. The Maia series will eventually become Microsoft's main source of inference capacity. The new chip utilizes four computing chiplets made with Taiwan Semiconductor Manufacturing Co.'s most recent two-nanometer node. These chiplets are surrounded by eight HBM4 modules, each containing 288 gigabytes of memory. Though Maia 300 has 20% less memory bandwidth than Nvidia's flagship Rubin accelerator, it compensates through a dataflow architecture that reduces data movement. This architecture allows computing chiplets to retain information in an internal SRAM cache when possible, or use a nearby memory stack when HBM transfer is necessary. Additionally, Maia 300 compresses data into lightweight formats to enhance speed. The chip is backed by a device control chiplet that manages power, collects data, and handles other tasks. Two auxiliary chiplets manage data movement in and out of the accelerator. Microsoft intends to install Maia 300 in trays, each containing four accelerators, which will be placed in liquid-cooled racks with a total of 96 accelerators. Microsoft anticipates the racks will offer about seven times the token throughput of its previous Maia 200 chip. The racks will be grouped into superpods, each containing 12,288 accelerators, connected via a custom Ethernet-based network stack. This network can adjust bandwidth allocation for in-rack and inter-rack communications, improving overall efficiency. Microsoft plans to roll out Maia 300 in production next year, aiming to provide multiple gigabytes worth of AI computing capacity.",
  "summary": "Microsoft Corp. today pulled back the curtains on Maia 300, its newest internally developed artificial intelligence accelerator. The chip is optimized for inference. Microsoft estimates that it can process 40% more tokens per megawatt than “leading GPU-based systems,” which is likely a reference to Nvidia Corp. products. The company stated the Maia chip series will […] The post Microsoft debuts…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}