Generating 1 minute of audio in 1 second with 1 GPU: Microsoft's internal AI announcement ─ Will the "heart" of Copilot be self-developed?

The "Coexistence Competition" Pioneered by In-House Models

Microsoft has decided to fully deploy its in-house AI. The announcement was made on August 28, 2025 (local time). The Verge described it as "a new twist in the complex partnership with OpenAI," positioning it as a "competing model" alongside GPT-5 and DeepSeek. In other words, the company has moved to a stage of "coexistence competition," collaborating with OpenAI while also standing at the forefront itself.The Verge

MAI-Voice-1: 1 Minute of Audio in Under 1 Second with 1 GPU

One of the highlights is the voice generation model "MAI-Voice-1." According to the official announcement, it achieves efficiency by synthesizing one minute of audio in less than a second using a single GPU. It is already integrated into "Copilot Daily," which reads news aloud, and a feature that explains topics in a "podcast style." In Copilot Labs' demo, users can experiment with changing the speaker's voice quality and narration style.Microsoft AI

MAI-1-preview: MoE LLM Trained with About 15,000 H100s

Another newcomer, "MAI-1-preview," is a Mixture-of-Experts type large language model adept at following instructions. Approximately 15,000 NVIDIA H100s were used for pre-training and post-training. It has begun public testing on the community evaluation platform "LMArena" and is gathering feedback through limited API access. A phased rollout for specific text uses in Copilot is also announced to occur within a few weeks.Microsoft AIPYMNTS.com

The Design Philosophy of "Consumer Optimization"

Mustafa Suleyman from Microsoft AI has consistently stated that the priority for in-house models is "consumer experience." The idea is to optimize "AI companions" by leveraging the company's data assets, such as advertising and consumer telemetry. Considering the company's shift towards a product-driven structure, this in-house development can be understood as a move to refine Copilot's "everyday pathways."The Verge

Copilot Moves Towards "Orchestration of the Best Models"

The company plans to continue using the "best models" from OpenAI and open source, while starting to use MAI-1-preview for some Copilot functions. The key is the concept of "orchestrating the optimal model for each use case." At this stage, it's not a complete replacement, but the precision of differentiation will determine success.Microsoft AI

Infrastructure Ambition: Operation of the GB200 Cluster

The announcement also mentions that the next-generation NVIDIA GB200 cluster is operational. This means the company is steadily preparing to "continuously and extensively" improve its in-house models. The ability to frequently update while keeping learning and inference costs low will likely differentiate it from competitors.Microsoft AI

Try It Out: Copilot Labs and LMArena

The experience pathways are already open. The voice generation MAI-Voice-1 can be tried from Copilot Labs. On the LLM side, users can participate in the evaluation of MAI-1-preview on LMArena, and the company is also recruiting testers through limited APIs. Before implementation into products, the process involves identifying "quirks" and "strengths" together with the community.Microsoft AI

Social Media Reactions: A "Duet" of Expectations and Caution

On X (formerly Twitter), Suleyman himself announced the "first in-house model." Technical accounts quickly spread the news, with many posts positively viewing it as "a strategic step towards independence from OpenAI."X (formerly Twitter)

On the other hand, PhoneArena and others emphasized concerns about the era where "any voice can be generated convincingly." There are many discussions demanding measures for the spread of voice deepfakes and the establishment of verification methods. On Reddit, debates are ongoing about "how to ensure safety" and redefining the relationship with OpenAI.PhoneArenaReddit

Strategic Impact: A "New Equilibrium" with OpenAI

The in-house model does not immediately "dissolve" the relationship between OpenAI and Microsoft. However, as The Verge points out, it visualizes a new phase where the company is not only a "supplier of top-tier models" but also a "competitor." If they can internalize the core of Copilot, they can accelerate differentiation at their own pace.The Verge

Future Evaluation Points

1. Performance: How well it ranks in practical measurements on platforms like LMArena.

2. Experience: Whether the phased introduction to Copilot improves answer quality, response speed, and audio naturalness as perceived by users.

3. Safety: Detection and labeling of voice forgery, rate control, and traceability in case of misuse.

4. Economy: How the advantage of fast generation with a single GPU impacts operations.

5. Governance: Whether the design of using OpenAI and other models together maintains transparency and responsibility boundaries.Microsoft AI

Conclusion: In-House Development is a "Means," Not an "End"

The adoption of in-house models is not an "end" to compete on the same stage as OpenAI. It is a "means" to refine the Copilot experience that integrates into users' daily lives at the company's own pace. The release of MAI-Voice-1 and MAI-1-preview is merely the prologue of this story. The key to the next chapter lies in how well they can execute on performance data from the field, safe operations, and the "orchestration of the best models."Microsoft AI

Powered by Froala Editor