24-Hour AI News: Nvidia Bet on OpenAI, Alibaba Open Sources Qwen3-Omni, Synchronous Upgrades Across China and Abroad

In the past 24 hours (September 22nd to September 23rd), overseas, Nvidia plans to invest up to US$100 billion in OpenAI and jointly build a 10GW AI data center, Google introduced Gemini into Google TV, and DeepMind updated its cutting-edge security framework; domestically, Alibaba Tongyi released and open-sourced the full-modal large model Qwen3-Omni, and DeepSeek upgraded V3.1 to Terminus version. Industries and scenarios in multiple places are being implemented simultaneously to achieve basically balanced coverage at home and abroad.

1. Nvidia plans to invest up to $100 billion in OpenAI and jointly build a 10GW AI data center

1. The upper limit of investment scale is US$100 billion, mainly used to purchase NVIDIA systems and deploy at least 10GW computing power clusters.
2. The transaction design includes non-voting equity and large-scale chip procurement. The first batch of US$1 billion to US$10 billion will be launched with the first 1GW of construction, and the first batch of equipment is targeted for delivery in the second half of 2026.
3. This move is seen as a milestone in strengthening the "computing power-model-application" closed loop of the AI industry, and may trigger antitrust scrutiny and ecological reconstruction.

2. Ali Tongyi releases and open-sources Qwen3-Omni: native end-to-end omnimodal, integrating voice, video, image, and text

1. Qwen3-Omni supports text, image, audio, and video input, and can output text and natural speech in real-time streaming, emphasizing the "native omnimodal" path.
2. The same batch of updates includes Qwen3-TTS and Qwen-Image-Edit-2509. According to reports, some versions adopt enterprise-friendly open source licenses (such as Apache 2.0) to facilitate commercial implementation.
3. Alibaba said it has refreshed the open source SOTA in multiple audio and audio-visual benchmark tests, expanding the options for domestic large models in the fields of generative AI and intelligent agents.

3. Google Gemini is available on Google TV and is first available on the TCL QM9K series

1. Starting from September 23, Gemini will enable conversational interaction on Google TV, which can be used for content search, plot review, learning guidance and recipe suggestions.
2. The first batch of models is TCL QM9K, and it will be expanded to Google TV Streamer, Walmart Onn 4K Pro, Hisense and more TCL models within the year.
3. This move is interpreted as an AI node in the living room scene, and smart TVs have evolved from "voice commands" to "conversational intelligent entities."

4. DeepMind Updates Frontier Security Framework (FSF): Adds "Shutdown Resistance" and "Harmful Persuasion" Risk Control Dimensions

1. This is the third iteration of FSF, emphasizing the monitoring and mitigation of potential evasion of human intervention (shutdown/modification) and high persuasive capabilities.
2. The framework introduces a stricter critical capability level (CCL) assessment and deployment mitigation process, with annual review and updates.
3. While competing and coexisting with mainstream large models (ChatGPT, Claude, and GPT-5 series), cutting-edge security governance is being strengthened to respond to AI ethics and governance concerns.

5. DeepSeek Releases V3.1-Terminus Upgrade: Significant Improvements in Language Consistency and Agent Stability

1. Officials claim that the problem of mixing Chinese and English and abnormal characters has been greatly alleviated, and Code Agent and Search Agent have been optimized to make output more stable.
2. According to media reports, some evaluation items have improved by more than 30% compared with the previous version; the App, web page and API have been upgraded simultaneously.
3. The iteration pace on the end-user and developer sides is accelerating, and the domestic reasoning/intelligent agent ecosystem continues to expand.

6. Meta's large-scale model Llama has been granted access to the US government procurement system and has been expanded to allied institutions in multiple countries.

1. The U.S. General Services Administration has included Llama in its available list to meet federal security and compliance requirements and adapt to multi-department business scenarios.
2. Meta also announced the expansion of availability to several allied organizations in Europe and Asia Pacific, emphasizing its restricted environment deployment capabilities with partners such as AWS and Snowflake.
3. The integration of policy and compliance paths is expected to drive the application of open source/open big models in the public sector.

7. Anhui released the first batch of 18 "artificial intelligence + manufacturing" scenario projects, with a total amount exceeding 370 million yuan

1. The project list covers scientific and technological innovation, industrial upgrading and public services, with the single project amount ranging from RMB 3 million to RMB 50 million.
2. Adopt the "take the lead" approach and provide support of up to 20% of the investment, with a maximum limit of RMB 10 million, to the unit that successfully answers the call.
3. Scenario-driven and fiscal support at the local level will promote the deep integration of the AI industry and the manufacturing industry.

8. Beijing established the "Artificial Intelligence Exhibition Ecological Alliance" to collaborate with the government, industry, academia, research and application in the industrial chain

1. The alliance was launched in the second phase of the National Convention Center. Its members include more than 50 research institutions and leading enterprises, focusing on technological innovation and scenario implementation.
2. Aims to build an AI ecological cluster of data elements and the exhibition industry, and promote the improvement of standards, applications and service systems.
3. Using major exhibitions as a starting point, create a demonstration window for industrial cooperation and application in the AI industry.

9. Shenzhen proposes quantitative goals to build an "AI pioneer city"

1. By 2026, the number of AI companies will exceed 3,000 and the number of unicorns will exceed 10.
2. The AI terminal industry is expected to reach a scale of 800 billion to 1 trillion yuan (RMB), with an average annual growth rate of more than 20%.
3. Driven by both talent and industry, the regional AI industry agglomeration effect is enhanced.

10. LSEG and Databricks reach data-AI collaboration for financial AI applications and agents

1. London Stock Exchange Group will natively introduce financial data into the Databricks platform through Delta Sharing.
2. The first phase covers Lipper funds and cross-asset historical analysis, and will subsequently expand pricing, reference data and models.
3. Provide a data foundation for financial institutions to build well-governed AI applications and intelligent entities, and enhance AI engineering and compliance capabilities.

Frequently Asked Questions (Q&A)

Q: What is the “biggest news” in the AI industry in the past 24 hours?

A: Nvidia plans to invest up to $100 billion in OpenAI and jointly build a 10GW AI data center with it. The first batch of supplies is targeted for the second half of 2026. It is seen as a structural turning point in the game between computing power and models.

Q: What is the most noteworthy development in the domestic model sector?

A: Alibaba Tongyi released and open-sourced a combination of Qwen3-Omni (omnimodal) and Qwen3-TTS. DeepSeek upgraded V3.1 to Terminus. The former complements the domestic "native omnimodal" and the latter improves the stability of the intelligent body and language consistency.

Q: What are the new trends in AI governance?

A: DeepMind's cutting-edge security framework has added key risk control points such as "shutdown resistance" and "harmful persuasion", emphasizing critical capability thresholds and deployment mitigation mechanisms, providing more detailed tools for large-scale model security governance.

Q: What are some representative cases in terms of application implementation?

A: Google brought Gemini to TV (first released on TCL QM9K), upgrading the "home big screen" to a conversational portal; on the financial side, LSEG cooperated with Databricks to provide a compliant data base for AI applications and agents.