24-hour AI news: OpenAI and AMD sign 6GW computing power contract, Purple East Taichu 4.0 released

In the past 24 hours, overseas, OpenAI and AMD reached a multi-year chip and computing power cooperation, California signed the AI Security and Transparency Act, and medical AI startup Heidi completed US$65 million in financing; domestically, the Institute of Automation of the Chinese Academy of Sciences and the Wuhan Institute of Artificial Intelligence released the multimodal reasoning model Zidong Taichu 4.0, the 2025 World Intelligent Connected Vehicle Conference was officially announced to be held in Beijing this month, and enterprise-level AI manufacturers released a number of new products.

1. OpenAI and AMD reach a long-term partnership for 6GW of computing power, with equity options up to approximately 10%

1. On October 6, the two parties signed a multi-year AI chip supply agreement, planning to deploy a total of 6GW of computing power in the next few years, with the first batch of approximately 1GW expected to be deployed in 2026 (US time).
2. The agreement includes supporting equity arrangements: OpenAI obtains warrants linked to the progress of chip deployment, with a total amount corresponding to options for up to approximately 10% of the equity (based on the current outstanding shares).
3. This cooperation has been interpreted by many media as a milestone event in the AI infrastructure landscape, which is expected to bring "billions of dollars" in new revenue to AMD and significantly diversify OpenAI's upstream supply risks.

2. Zidong Taichu 4.0 Release: Towards the Deep Reasoning Stage of "Fine-Grained Multimodal Semantic Thinking"

1. On October 6, Zidong Taichu 4.0 was jointly released by the Institute of Automation of the Chinese Academy of Sciences and the Wuhan Institute of Artificial Intelligence, emphasizing explainable multimodal reasoning of "seeing, recognizing, and thinking at the same time".
2. Example capabilities include: segment-level positioning and summarization of 180-minute videos, automated complex operations in real-world apps such as medical registration, and hands-on execution in embodied scenarios.
3. The implementation directions cover embodied intelligence, low-altitude economy, smart medical care, etc., marking the advancement of domestic multimodal reasoning towards real tasks and operability.

3. California AI Safety and Transparency Act (SB 53) has a clear path to signing and taking effect

1. On October 5 (local time), the Governor of California signed SB 53, requiring cutting-edge model developers to disclose and follow a safety and security assessment process, which will be overseen by the State Office of Emergency Management.
2. The core of the bill focuses on the protection of "high-risk capabilities" (such as cyber attacks on critical infrastructure and biosecurity risk control) and is called "the first state-level framework of its kind."
3. The industry believes that the bill will improve bottom-line security without inhibiting innovation, and will work with federal and EU regulations to shape compliance boundaries.

4. The 2025 World Intelligent Connected Vehicle Conference will be held in Beijing from October 16 to 18

1. On October 6, authoritative media confirmed that the theme of the conference is "Gathering Wisdom and Energy for Infinite Connections", focusing on autonomous driving, AI empowerment, vehicle-road-cloud integration, chips and communications, etc.
2. A number of industry reports will be released at that time, including the construction achievements of "artificial intelligence + automobile" and "vehicle-road-cloud integration".
3. The forum structure includes policy governance, technological innovation, security protection, data application, etc., to deepen industrial collaboration and international exchanges.

5. Heidi Health completes $65 million Series B financing, with a valuation of $465 million

1. On October 6, Australian medical AI company Heidi announced that it had received US$65 million in financing led by Point72 and others, bringing the total financing to nearly US$100 million and the latest valuation to approximately US$465 million.
2. Its "AI Nursing Partner/Medical Document" product covers clinical records, evidence-based retrieval, follow-up communication and other aspects, and is said to have returned tens of millions of hours of work time to frontline doctors.
3. The funds will be used for global expansion and product iteration, focusing on standardized implementation in multi-language and multi-specialty scenarios.

6. The EU promotes a new draft strategy for "Apply AI": reducing dependence, strengthening application, and achieving independent control.

1. Between October 5 and 6, multiple media outlets revealed that the European Union plans to release a draft "Apply AI" strategy, with the goal of reducing dependence on US and Chinese technology and promoting the application of domestic AI in the public sector and key industries.
2. It plans to mobilize approximately 1 billion euros of existing funds, giving priority to deployment in defense C2, medical care, manufacturing, etc., emphasizing open source ecology and accessibility for small and medium-sized enterprises.
3. This move and the regulatory system of the AI Act form a "regulation + application" dual wheel to enhance industry competitiveness and digital sovereignty.

7. Minglu Technology launched DeepMiner, an enterprise-level large-scale model product line, and Mano, a GUI intelligent agent, ranked on both SOTA lists.

1. On October 6, Mininglamp Technology announced the "DeepMiner" product line for commercial data analysis, focusing on "trusted intelligent agents" and verifiable processes.
2. On the same day, it was disclosed that the self-developed GUI large model Mano achieved SOTA results in the Mind2Web and OSWorld benchmarks, emphasizing the paradigm of online reinforcement learning and automatic data collection.
3. Positioning multi-agent collaboration in ToB production scenarios, focusing on enterprise implementation standards of traceability, auditability and human-machine collaboration.

8. Alibaba Cloud Tongyida Model and Agent Toolchain Update (According to Media Reports)

1. On October 6, multiple reports summarized the "seven consecutive updates" of Tongyi Qianwen during the Yunqi Conference, saying that Qwen3-Max has a parameter scale of over one trillion and 36T tokens of pre-training data, covering all sizes and all modalities.
2. Reports indicate that the open source ecosystem has accumulated more than 300 models, with over 600 million downloads and over 170,000 derivative models; the average daily call volume of the Bailian platform has increased by about 15 times a year.
3. Official and third-party evaluations of its ranking as "top three globally" still need to be verified, and the industry is paying attention to its replicability and cost-effectiveness in agent and enterprise-level tool chains.

Frequently Asked Questions (Q&A)

Q: How big is the 6GW computing power cooperation between OpenAI and AMD?

A: Planning in GW units is extremely rare, which means the construction of large-scale GPU clusters in batches over several years; the agreement also gives OpenAI the option of up to approximately 10% equity, which helps to bind long-term supply and reduce the risk of a single supplier.

Q: What are the key changes in Zidong Taichu 4.0 compared to the previous generation?

A: From "being able to see and answer" to "being able to see, think and do", emphasizing fine-grained multimodal semantic reasoning and explainability. Typical scenarios include long video retrieval and summary, real App automation operation, and embodied intelligent execution.

Q: How will California SB 53 affect AI companies?

A: Auditable security processes and information disclosure requirements are proposed for the development of "high-risk capability" models, and are supervised and implemented by state-level agencies; the impact on companies that already have mature security processes is mainly in compliance alignment and documentation.

Q: What is the relationship between the EU’s “Apply AI” and the AI Act?

A: The AI Act provides a "risk classification + compliance" framework. "Apply AI" is more like a supporting deployment of "application and autonomous capability building", emphasizing funding orientation and implementation in key areas.

Q: What should companies pay attention to when implementing Agent?

A: There are two main lines: one is trustworthiness and traceability (process and results are verifiable), and the other is the governance and cost control of human-machine collaboration; evaluating "SOTA on the list" does not mean that it can be used in production, and a comprehensive evaluation is required based on latency, stability, and TCO.