Microsoft가 Maia 300 AI 가속기를 공개했다.

Microsoft has announced the Azure Maia 300, a third-generation AI accelerator designed for complex AI reasoning and fast response times. It is the company's newest custom chip for AI inference, and it features a multi-chiplet design, 288 GB of HBM4 memory, and 18 TB/s of memory bandwidth. Microsoft is packing 96 accelerators into a liquid-cooled rack that draws 167 kW.

Microsoft introduced the chip in a post on X on October 12, 2026:

> "Introducing Maia 300, Microsoft's next-generation accelerator. Building on the magic of Maia 200, it delivers 40% more inference tokens per megawatt compared to leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200. It's another step toward rethinking innovation and unlocking new possibilities."

>

> — Microsoft (@Microsoft)

**A New Multi-Chiplet Design**

The Maia 300 changes how Microsoft builds its chips. Rather than building one large block, Microsoft is assembling the chip from connected pieces, an approach known as a multi-chiplet design.

**Splitting the Brain and the Memory**

The design separates the main computing "brain" from the memory components. The main compute die will use TSMC's new 2nm technology to save power. By splitting these parts, Microsoft can fit more computing power and speed into a single chip, and it keeps data close to the processor, so information moves faster.

**Handling AI Tasks**

To manage AI workloads, the chip compresses data into smaller sizes and uses specialized formats such as MXFP4. Because the data takes up less space, the AI can retain longer conversations and serve more users at the same time.

**Packing More Chips Together**

Microsoft is placing more chips in each server. A single rack will hold 96 Maia 300 chips, which together will draw 167 kW.

**Cooling with Liquid**

Because so many chips generate a great deal of heat, air fans alone are not enough. Microsoft will use a liquid cooling system for more than 80% of the hardware.

**Connecting Thousands of Chips**

Microsoft has built a custom network that will allow more than 12,288 chips to communicate within a single group, which the company calls a superpod.

**Maia 300 vs. Maia 200 Specifications**

| Feature | Maia 300 | Maia 200 |

|---|---|---|

| Architecture | Multi-chiplet (TSMC N2P 2nm class) | Monolithic (TSMC N3P) |

| Power (TDP) | 950 W | 750 W |

| Memory | 288 GB (8 × HBM4) | 216 GB (6 × HBM3e) |

| Memory Bandwidth | 18 TB/s | ~7 to 7.1 TB/s |

| Chips per Rack | 96 (using 167 kW) | 48 (using ~64 kW) |

| Max Scale-Up | ~12,288 chips (superpod) | ~6,000 chips |

**A Strategy for Hardware Independence**

The Maia 300 is part of Microsoft's long-term plan to reduce costs and energy use by building its own chips and data centers. By making its own hardware, Microsoft will be less dependent on outside companies such as Nvidia and AMD for AI workloads. The company noted, however, that it will continue to use traditional GPUs for training AI models.