24-hour AI news: litigation and financing go hand in hand, Jetson Thor launches new sales and AI DingTalk

24-hour AI news at a glance: domestic enterprise-level and industrial-grade applications are new, and overseas computing power and regulatory wind direction have jumped at the same frequency

In the past 24 hours, Huawei on the domestic side has announced the release time of AI SSDs, DingTalk released AI 1.0 and strengthened office scenarios, and Wuhan's "AI+" industry has achieved results. Overseas xAI filed a lawsuit against Apple and OpenAI, AirTrunk completed a $100 billion refinancing, NVIDIA Jetson Thor went on full sale for physical AI, and in terms of methodological research, Apple proposed the RLCF checklist-style alignment paradigm to achieve balanced coverage of China and foreign countries as a whole.

1. xAI sues Apple and OpenAI, focusing on system-level AI and application distribution competition

1. On August 25 (local time), a case was filed in the U.S. Federal Court in Texas, seeking "billions of dollars" in damages and behavioral relief.
2. The complaint alleges the impact of system-level integration, default portals, and App Store ranking mechanisms on competition.
3. Short-term or trigger regulatory review of AI access rules in the mobile ecosystem, which will affect the neutrality boundary of system-level AI in the long term.

2. AirTrunk received $10.4 billion in refinancing, expanding its Asia-Pacific ultra-large-scale AI data center

1. On August 26, it announced the completion of an A$16 billion (approximately US$10.4 billion) sustainable-linked refinancing, covering business expansion in Asia-Pacific and Japan.
2. Supporting Singapore's S$22.5 billion green loan framework to strengthen the supply capacity of regional low-carbon data centers.
3. It is expected to alleviate the bottleneck of AI training/inference computing power and electromechanical support, and drive the demand for power, liquid cooling, network and upstream equipment.

3. NVIDIA Jetson Thor is officially on sale for "physical AI" and robots

1. The Jetson AGX Thor development kit and mass-produced Jetson T5000 module will be released and sold on August 25.
2. The official nominal FP4 computing power is about 2070 TFLOPS, the development kit starts at $3499, and the real-time fusion capabilities of multi-model generative inference and multi-sensor on the single machine side are enhanced.
3. The early ecology covers humanoids and industrial robots, medical equipment and automation scenarios, and the proportion of end-side inference increases.

4. Apple proposes an RLCF checklist-based feedback alignment paradigm to improve complex instruction execution

1. The study proposes to replace a single ruler reward with "task list scoring", combined with AI referee and programmatic verification to generate reinforcement learning signals.
2. In the dimensions of complex instruction compliance and constraint compliance, it is emphasized that the method is not replaced by "safety alignment".
3. Provide new tools for enterprises to align models within an interpretable and auditable framework, and reduce the dependence on manual preference labeling.

5. Huawei officially announced the release of AI SSDs on August 27, facing the bottleneck of HBM

capacity 1. Positioning "high capacity + high performance" AI-specific SSDs, facing the shortcomings of the storage layer for large model training and inference.

1. It will collaborate with the whole machine/server ecosystem to serve data-intensive and inference acceleration scenarios.
2. If mass production is implemented, or improve the domestic AI storage chain to improve cost performance and availability.

6. DingTalk released 8.0 (AI DingTalk 1.0), a comprehensive upgrade of enterprise-level AI workflows

1. Released in Hangzhou on August 25, 10 new AI products such as enterprise AI search, AI forms, listening and application creation were added.
2. The official caliber positions the "AI native" collaborative platform to strengthen information retrieval, knowledge management and automated processes.
3. According to media reports, DingTalk has served more than 26 million enterprises and more than 190,000 paying enterprises, and the scale of ecological partners and AI applications has expanded simultaneously.

7. Progress of Wuhan's "AI+" action: the efficiency of industrial welding recognition has increased by about 50% 1

. The weld recognition and tracking system based on the industrial model has been implemented, and the robot welding recognition efficiency has been improved by about 50% compared with the traditional solution.

1. Promote industry-university-research collaboration in the direction of display panel design simulation and yield optimization.
2. The "computing power× application" in the region is parallel, and the digital transformation of the manufacturing industry is accelerating.

8. The release of domestic Wanka cluster large model inference technology to tackle domestic chip adaptation

1. On August 25, the "domestic Wanka cluster large model inference technology" was disclosed, facing the adaptation problem between domestic chips and mainstream frameworks.
2. Systematically tackle key problems around the technical stack such as parallel inference, memory management, and compilation optimization.
3. Provide an integrated "end-edge-cloud" path for multi-scenario deployment to reduce the latency and cost of large-scale inference.

Frequently Asked Questions (Q&A)

Q: What are the direct impact points of xAI's prosecution incident?

A: The focus is on the competitive impact of system-level AI default entry and distribution rankings, as well as the transparency of potential training data channels and platform rules. In the short term, it may prompt the refinement of platform disclosure and access mechanisms, and in the long term, it may trigger antitrust and neutrality discussions in the mobile ecosystem.

Q: What are the core parameters and application threshold of Jetson Thor?

A: The official nominal computing power is about 2070 TFLOPS (FP4), and the development kit starts at $3499; It supports multi-model real-time inference and multi-sensor fusion, making it suitable for scenarios with high latency requirements, such as humanoid/industrial robots, automation, and medical equipment.

Q: What does AirTrunk refinancing mean for the AI industry chain?

A: Sustainable refinancing of $100 billion will accelerate the launch of data centers in the Asia-Pacific region, benefiting supporting links such as power supply, liquid cooling, networking, and cabinets, while reducing the unit cost and latency of training/inference through scale effects.

Q: What is the difference between AI DingTalk 1.0 and traditional collaborative office?

A: From "application-driven" to "agent-driven", the core is AI search, AI documents/forms, automatic listening and process orchestration, and enterprise knowledge and workflows can be uniformly retrieved and called by AI, significantly reducing the cost of cross-application switching.