What Is Maia 300? Microsoft's Inference Chip

Inside Maia 300: Microsoft's 288GB HBM4 Inference Chip

Maia 300 is Microsoft's third-generation Azure AI accelerator, a custom inference chip announced on October 12, 2026 that the company says delivers 40% more tokens per megawatt than leading GPU-based systems. It pairs a TSMC N2P (2nm-class) compute die with eight HBM4 stacks for 288GB of memory at 18TB/s, inside a 950W power envelope. Production deployments begin in 2027, when Microsoft plans to make the Maia family the primary inference system for its own products and services.

Key Takeaways

• Maia 300 is Microsoft's third in-house AI accelerator, following Maia 100 (2023) and Maia 200 (January 2026), unveiled at the OCP Global Summit 2026 in San Jose.
• Microsoft claims 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale token throughput of Maia 200.
• Each accelerator carries 288GB of HBM4 at 18TB/s of bandwidth on a 950W TDP, built as a multi-chiplet part on TSMC's N2P process.
• A full rack holds 96 accelerators at 167kW with 27.6TB of HBM, and the scale-up fabric reaches a superpod of 12,288 accelerators over Ethernet.
• Maia is inference-first and stays entirely in-house; third parties cannot rent it, unlike Google TPUs or AWS Trainium.
• Production starts in 2027; no independent benchmarks, price, or hands-on tests exist yet as of October 2026.

What Is Maia 300?

Maia 300 is a custom silicon accelerator Microsoft designed to run AI inference in its own data centers. Inference is the process of generating responses from a trained model, as opposed to training the model in the first place. Microsoft positions Maia 300 as the third generation of a multigenerational program: Maia 100 was a 2023 pilot part, Maia 200 shipped in January 2026 and now serves production workloads, and Maia 300 extends that line with a far larger memory system.

The chip was announced, not yet shipped. In the official Microsoft blog post, Scott Guthrie wrote that Microsoft will "scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services." CEO Satya Nadella called it "our biggest bet yet on our own silicon."

The strategic goal is independence from merchant GPUs. The Verge and CNBC both framed the launch as Microsoft reducing reliance on Nvidia and AMD. Evercore ISI analysts estimate Maia will represent only about 10% of Microsoft's AI-chip spending by Q1 2027, so the shift starts small.

Maia 300 at a Glance

Attribute	Detail
Developer	Microsoft (compute die fabbed by TSMC)
Type	Custom AI inference accelerator (ASIC)
Generation	Third (Maia 100 → 200 → 300)
Process	TSMC N2P (2nm-class), multi-chiplet
Memory	288GB HBM4 across 8 stacks, 18TB/s
Peak compute	13.8 POPS FP4, 6.9 POPS FP8
Power (TDP)	950W per accelerator
Rack config	96 accelerators, 167kW, 27.6TB HBM
Scale-up domain	Superpod of 12,288 accelerators over Ethernet
Host interface	PCIe Gen6 x16; 32 ports of 400Gb/s Ethernet
Availability	Production deployments begin 2027
Pricing	Not published
License	Proprietary; Microsoft-internal only, not rentable

How Maia 300 Works

Maia 300 extends an architecture Microsoft introduced on Maia 200 called the Software Defined Locally Accessed Dataflow Architecture (SDLA). The idea is that moving data, not raw compute, is the main cost in modern inference, so software explicitly coordinates where data lives, when it moves, and how it is converted across local memories, DMA engines, compute engines, and the network.

Three changes define the new generation. First, memory moves from a uniform layout to a locality-aware NUMA organization across chiplets, so software can place data near the compute that uses it. Second, narrow-precision formats such as enhanced MXFP4 with hardware data-type conversion cut the bytes moved per token. Third, the rack doubles to 96 accelerators and the scale-up domain grows to a 12,288-accelerator superpod.

The physical package is a departure from Maia 200's monolithic die. According to a third-party specification writeup, Maia 300 splits into a main compute die on N2P, two I/O dies, and four memory dies, with eight HBM4 stacks raising capacity from 216GB to 288GB and bandwidth from 7.1TB/s to 18TB/s.

Source: @mark_k

A device control chiplet manages power delivery and telemetry, while two auxiliary chiplets handle data movement in and out of the accelerator. The scale-up fabric runs over a custom Ethernet-based transport with two-tier switching, and it uses the same communication semantics whether a job spans 16 accelerators or four racks, which reduces stranded capacity.

What You Can Do With Maia 300

Maia 300 is built for a specific class of work: inference, not training. Microsoft describes it as "a complete intelligent system designed for the new types of AI workloads emerging today: long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks."

Source: @testingcatalog

In practice, that means serving the models behind Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Dragon Copilot. The predecessor Maia 200 already runs OpenAI models, Microsoft's own MAI models, and leading open-source families in production. Maia 300 is designed for the decode-heavy, memory-intensive serving patterns that large mixture-of-experts models generate.

You cannot build on Maia 300 directly. It never leaves Microsoft's fleet. Developers who want to serve comparable frontier models today reach them through vendor APIs instead; a hosted chat endpoint such as Claude Opus 5.5 runs the same inference workload Maia is tuned for, without any hardware decision on your side.

How Maia 300 Compares

The clearest comparison is against Microsoft's own previous generation. The jump is weighted toward memory, not peak math.

Spec	Maia 300	Maia 200
Process	Multi-chiplet, N2P compute die	Monolithic N3P
FP4 peak	13.8 POPS	10.1 POPS
FP8 peak	6.9 POPS	5.1 POPS
TDP	950W	750W
HBM	288GB HBM4, 18TB/s	216GB HBM3E, 7.1TB/s
Rack power	167kW	64kW
Accelerators / rack	96	48

FP4 compute rises 37% while memory bandwidth jumps 2.5x. Microsoft summarizes the generation as "2.5x the bandwidth, 33% more capacity, and about 2x the HBM bandwidth per watt" of Maia 200. Against Nvidia's flagship Rubin accelerator, one third-party analysis noted Maia 300 matches its 288GB capacity but carries roughly 20% less memory bandwidth, and leaves out the CUDA software ecosystem entirely.

Availability: How to Access Maia 300

Maia 300 production deployments begin in 2027. As of October 2026 no units serve live workloads, and no third party has tested the chip. Microsoft describes the rollout as "multi-gigawatt," but has not published a region list, unit volume, or a precise date beyond 2027.

Access is the sharpest limitation. Unlike Google, which sells TPUs to external customers, and Amazon, whose Trainium chips are rentable and attract outside AI labs, Microsoft keeps Maia completely in-house. Azure customers cannot select a Maia 300 instance. The only way to benefit is to use a Microsoft product whose inference happens to run on it. For official specifications, Microsoft maintains a Maia 300 architecture deep-dive on the Azure blog.

What We Don't Know Yet

Several gaps remain open as of October 2026:

• Benchmark methodology. The 40%-tokens-per-megawatt and 7x-throughput claims come from Microsoft alone. No post or document specifies the models, batch sizes, latency targets, or GPU baselines used. Independent numbers do not exist yet.
• Deployment scale. One account cited a Reuters-sourced claim of more than one million chips during 2027, but that figure is uncorroborated by other reports reviewed.
• Supplier details. The HBM4 supplier was unannounced at launch, according to one post.
• Real displacement. How much Nvidia and AMD capacity Maia 300 actually replaces, and in which services first, is unstated. Microsoft confirms it will keep buying third-party chips for training.
• Price. No per-chip cost, system cost, or internal cost basis has been disclosed.

Frequently Asked Questions

What is Maia 300?

Maia 300 is Microsoft's third-generation Azure AI accelerator, a custom inference chip announced on October 12, 2026. Microsoft says it delivers 40% more tokens per megawatt than leading GPU-based systems and will enter production deployments in 2027.

Who makes Maia 300?

Microsoft designs Maia 300 in-house, and the compute die is fabricated by TSMC on its N2P (2nm-class) process. Microsoft unveiled the chip at the OCP Global Summit 2026 in San Jose.

When is Maia 300 available?

Maia 300 begins production deployments in 2027. Microsoft announced it on October 12, 2026, but no units are running production workloads yet; the earlier Maia 200 already serves Microsoft Copilot.

How much does Maia 300 cost?

Maia 300 has no published price. Microsoft keeps Maia chips entirely in-house and does not rent them to third parties, so there is no public per-chip or per-hour rate, unlike Google TPUs or AWS Trainium.

Is Maia 300 better than Nvidia GPUs?

Microsoft claims Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems, but no independent benchmark exists yet. Microsoft did not name Nvidia or disclose the models, batch sizes, or baselines behind the figure.

Can I use Maia 300 on Azure?

No. Maia 300 is reserved for Microsoft's own products and services, such as Copilot, GitHub Copilot, and Microsoft Foundry. Azure customers cannot select Maia 300 directly, unlike rentable custom chips from Google and Amazon.

What is Maia 300 built for?

Maia 300 is built for AI inference, specifically long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic tasks. Its architecture is optimized for decode-heavy, memory-intensive workloads rather than training.

What to Watch Next

Three signals will tell you whether the 2027 rollout matches the announcement. First, watch for independent benchmarks once early silicon reaches testers, since every performance claim so far is Microsoft's own. Second, track which services migrate first and in which regions, which will reveal real deployment scale rather than the vague "multi-gigawatt" framing. Third, follow the Evercore 10%-of-spend estimate against actual capital reports, because that ratio is the honest measure of how much Nvidia and AMD dependence actually falls.

Try related models on Kie.ai: Claude Opus 5.5, GPT-6 Sol and Luna and Claude Haiku 5.5.