On October 12, Microsoft officially unveiled its third-generation in-house AI chip Maia 300and explicitly stated that the Maia series will become the “primary AI inference system” for Microsoft’s products and services. This means one of the world’s largest NVIDIA GPU buyers is accelerating its move away from dependence on NVIDIA and AMD, bringing AI inference capability fully under its own control.
Where Maia 300 excels: simultaneous leaps in energy efficiency and throughput
Maia 300 is the successor to Maia 200 (released just this January), with its core value proposition distilled into two words—EfficiencyMicrosoft Cloud Executive Vice President Scott Guthrie cited two key metrics:
- 40% improvement in energy efficiencyCompared to mainstream GPU systems, Maia 300 is projected to generate 40% more tokens per megawatt.
- 7x increase in throughputIn a full rack configuration with 96 accelerators, Maia 300 delivers seven times the rack-level token throughput of Maia 200.
These efficiency gains stem from an entirely new chiplet architectureGuthrie explicitly confirmed that Microsoft will begin scaling Azure’s Maia family to “multi-gigawatt” deployments starting in 2027, establishing Maia as the primary inference foundation for Microsoft’s own products.
Who is Maia 300 for: inference models and agents
Maia 300’s design targets three emerging AI workloads:
- Long-context inferenceIncreasingly large context windows place higher demands on inference hardware.
- Low-latency conversational experiencesReal-time voice and conversational AI require extremely fast response times.
- Continuous multi-step agent tasksLong-running workloads of the Agent class are at the core of the Microsoft Copilot ecosystem.
Microsoft has already proven this path viable internally: Maia 200 is already running in production. Microsoft Copilot,GitHub Copilot,Microsoft Foundry and Microsoft’s own AI models. Guthrie even called Maia 200 “the best-performing infrastructure under real-world Microsoft AI service metrics.” Maia 300 builds directly on this foundation, pushing capabilities to the next level.
The in-house chip race intensifies: Microsoft is not alone in this fight.
With the launch of Maia 300, the “cloud provider–designed AI chips” track has become fully clarified. Currently, all three major cloud computing giants have bet on in-house AI chips:
- Microsoft Maia seriesAlready evolved to its third generation, targeting “primary AI inference systems.”
- Google TPUAmazon AWS Trainium / Inferentia Gemini Already powering training and inference for models such as
- Amazon Trainium / InferentiaAlso reducing reliance on Nvidia.
The underlying logic is simple: inference accounts for the largest share of AI costs—controlling inference chips means controlling pricing power and profit margins. For enterprises running models, this implies sustained downward pressure on future inference costs—DeepSeek V4 Pro,GLM 5.3 These high-inference-capacity models are direct beneficiaries of this inference hardware advantage.
Chronologically, Microsoft has moved remarkably fast: Maia 200 launched only in January this year, and Maia 300 arrived just months later—with an explicit target of “multi-exawatt-scale deployment.” This pace reflects Microsoft’s acute urgency around slashing inference costs—after all, the Copilot ecosystem’s user scale is massive, and every cent of inference cost gets magnified exponentially.
What this means for the industry and everyday users
The signal from Maia 300 is clear:Cloud providers’ in-house chips are shifting from “supplementary” to “primary.”For ordinary users, the direct impact of this will manifest later—only after Microsoft reduces the inference costs of products like CopilotGitHub Copilot Long-term adjustments to subscription pricing and free quotas may then become possible. For developers, however, it may soon no longer matter whether NVIDIA GPUs underpin inference models running on Azure.
Notably, beyond chips, Microsoft is also building decision models (see Microsoft-Decision-1 in-depth review), pursuing a dual-track strategy spanning hardware and software to control both ends of AI—the “compute + models.”
Frequently Asked Questions (FAQ)
What is Microsoft Maia 300?
Maia 300 is Microsoft’s third-generation in-house AI chip, positioned as the “primary AI inference system” for Microsoft’s products and services. Mass deployment is expected to begin in 2027, replacing portions of NVIDIA and AMD GPUs.
Where does Maia 300 outperform NVIDIA GPUs?
Microsoft claims Maia 300 delivers 40% more tokens per megawatt and achieves seven times the rack-level throughput of Maia 200 in a 96-accelerator configuration. Its core advantages lie in energy efficiency and cost per unit—not raw absolute compute power.
When will Maia 300 be available?
Maia 300 is expected to enter production AI workloads in 2027; ordinary developers cannot yet use it directly.
Will in-house chips make AI cheaper?
Very likely. The primary goal of in-house chips is to reduce inference costs, and those savings are expected to flow through to model APIs and subscription pricing over time—this is why we consistently track model cost-effectiveness. Follow Tool Comparison Engine for updates on pricing changes across vendors.
Want to stay updated on the latest developments in the AI industry? Check our AI Model Library and benchmark listor continue reading:Microsoft-Decision-1 in-depth review · Gemini 4 Argon released · Mistral Large 4 Release Analysis.
