ByteDance secures approximately $29.6 billion in syndicated loans to fund massive AI infrastructure expansion and compute scaling.
Evaluation results indicate that improvements in certain domains for this new model can no longer be described as incremental. On FrontierMath Tier 4, a benchmark for advanced mathematical reasoning, Astra scored 97.6%, effectively mastering the research-grade test. On ARC-AGI-3, which emphasizes learning in novel environments and abstract reasoning, its performance surged from 7.8% with GPT-5.6 Sol to 99.9%, nearing a perfect score.
On September 3, Xiaomi Group President Lu Weibing addressed questions regarding the pricing of the upcoming Xiaomi 18 Fold during a live stream. He noted that soaring memory costs have made the device notably expensive, with prices starting above 10,000 RMB—likely significantly higher than that baseline. Lu added that Apple is reportedly developing a mid-size foldable phone; consumers are encouraged to compare the two, asserting that the Xiaomi 18 Fold will undoubtedly offer exceptional value. Notably, he revealed that the CXMT LPDDR6 memory will be exclusive to the top-tier 16GB RAM + 1TB storage variant. The device will debut Xiaomi’s self-developed Xuanjie O3 AI flagship chip, marking the industry’s first AI SoC to surpass 5 million points on benchmarks. The Xuanjie O3 is also the world’s first mobile processor supporting LPDDR6 memory, delivering a bandwidth of 113.8 GB/s. This configuration pairs Xiaomi’s proprietary SoC with China’s next-generation memory. In terms of design, the Xiaomi 18 Fold features a mid-fold form factor with ultra-narrow four-sided bezels and large super-elliptical rounded corners. Both the inner and outer displays utilize a standard paper-like aspect ratio of approximately √2:1. The outer screen measures 5.38 inches, expanding to a 7.58-inch inner display when unfolded. It incorporates dual-layer UTG ultra-thin glass, striking a balance between portability and an expansive viewing experience. (KuaiTech)
On September 3, media reports indicated that ByteDance has secured a $29.6 billion syndicated loan. Initially targeting $20 billion, the financing was scaled up due to strong bank demand. The funds will be allocated for general corporate purposes. The transaction has not yet been formally signed, with banks still finalizing allocation quotas. Once completed, it will rank as Asia’s second-largest dollar-denominated syndicated loan this year, trailing only SoftBank Group’s $40 billion bridge loan in March. ByteDance is accelerating its AI investments, considering raising capital expenditures to as high as $70 billion in 2026—more than double last year’s figure—primarily to expand data centers and other AI infrastructure. The company’s previous major overseas borrowing occurred in 2024, raising approximately $10.8 billion. As of press time, ByteDance had not commented. (Yicai Global)
On the afternoon of September 3, WeChat’s Director of Public Relations posted about the “view single-deleted friends when contacts exceed 10,000” feature, noting widespread user interest. The feature was not designed to help users identify who deleted them, but rather to provide a way for those overwhelmed by friend requests to manage their connections. Addressing calls for a one-click tool to check deletions across all contacts, the post explained that this deliberate stance aims to counter “social anxiety,” preventing the paranoia of being deleted from becoming a Pandora’s box accessible to everyone at any time. “It’s not about hiding information, but preventing it from becoming a new burden. Clear out some relationships to make room for what matters. If some connections fade, let them fade. Knowing doesn’t mean you need to care.” On September 3, the hashtag #WeChatSingleDeleteNotification# trended online. Some users reported discovering a new thoughtful feature: once the friend limit is reached, the system prompts users about who deleted them, allowing for one-click removal. (Yicai Global)
On September 3, netizens shared an email from “Renren Movies” welcoming back former users with the message, “We finally meet again in the App,” offering a one-month free membership. Reports confirm the app has been rebranded as “Renren Movies Shares Excellence” and is now available on iOS and Android stores, prominently displaying “officially licensed translations.” The platform continues to focus primarily on foreign film and television content. Staff members stated the project underwent extensive preparation before its official launch, partnering with Wasu Media to ensure all content is legally licensed. Similar to mainstream streaming services, the app offers tiered subscriptions: 25 RMB/month, 60 RMB/quarter, 175 RMB/year, and 888 RMB for lifetime access. Originally launched in 2006, Renren Movies gained prominence in China for its overseas subtitle resources. Under pressure from international copyright holders, it shut down download services in 2014 and pivoted to a drama community. In 2021, Shanghai police announced the dismantling of the “Renren Movies Subtitle Group” over copyright infringement involving over 16 million RMB, leading to arrests and the complete removal of its client apps. (National Business Daily, BianNiuShi)
On September 3, insiders revealed that Sun Tianxiang, head of Baidu’s Base Model Unit (BMU), internally announced the recruitment of a prominent AI researcher known by the pseudonym Wu Qin from a North American frontier lab. Wu Qin will collaborate with the team to strengthen pre-training capabilities for the Wenxin (Ernie) large language models. According to Sun, Wu Qin contributed to the development and iteration of several generations of large models over the past two years, though his real name and background remain undisclosed. Additionally, Wei Haoran, a former DeepSeek researcher who joined Baidu earlier this year, has transferred to BMU to lead multimodal algorithm development for the Wenxin models. Since joining Baidu, Wei has driven significant breakthroughs in OCR, open-sourcing the Unlimited OCR model in June and setting a new state-of-the-art record for end-to-end OCR performance. From 2024 to 2025, Wei served as a researcher at DeepSeek, contributing to the development of DeepSeek V3.2 and V4, and spearheading the open-source release of the DeepSeek-OCR series. (Shenma’s Moments)
On September 3, at the IFA 2026 Xiaomi Global Launch Event, Xiaomi officially introduced the Mijia brand to European and global audiences, announcing its large-scale expansion into Europe. IFA (International Consumer Electronics Fair) is Europe’s largest consumer electronics exhibition. Making its debut at the event, Xiaomi showcased a massive booth exceeding 3,000 square meters, representing the company’s largest-ever centralized presentation of its “human-car-home” ecosystem. During the event, Xiaomi Group Partner and President Lu Weibing reiterated the company’s ecosystem philosophy and reaffirmed that Xiaomi EVs will enter the European market next year. (ITHome)
Leiphone exclusively learned that Hu Han, head of smart cockpit development at Li Auto, recently resigned and will join an AI company. A graduate of China Agricultural University, Hu previously served as a product manager at Kaixinc.com, Baidu, and Tencent, accumulating extensive consumer-facing product experience. After joining Li Auto in 2020, he became a core architect of its cockpit intelligence initiatives. Over five years, he progressed from smart voice product manager to head of smart space products, leading the definition and iteration of the Li Auto Assistant voice interface and overall cockpit experience. Separately, Leiphone’s New Smart Driving channel reports that Xie Baojun, Vice President and head of the Intelligent Center at Chery Automobile Co., Ltd., appears to have departed, with his internal Feishu account showing as suspended. As of press time, Chery had not commented. Xie joined Chery in October 2024, initially working at subsidiary Dazhuo Intelligence on autonomous driving R&D. In May 2025, Chery consolidated Lion Technology, Dazhuo Intelligence, and central R&D operations to form the Intelligent Center, appointing Xie as its operational lead to oversee smart cockpits, advanced driver assistance systems, and digital architectures. Prior to Chery, Xie held positions at Volvo Cars and Zeekr. (Author note: The author has long tracked developments in autonomous driving and the automotive sector. Connect via WeChat: jinyuan-0428 for discussions.)
On the evening of September 3, Enflame Technology announced its lottery winning numbers for its initial public offering. The issue price is set at 142.18 RMB per share, with 43.035173 million shares issued, projecting total proceeds of 6.119 billion RMB. Investors participating in the online subscription whose ticket serial numbers match the published winning digits are considered successful applicants. There are 20,657 winning numbers in total, each entitled to purchase 500 A-shares of Enflame Technology. Founded in March 2018, Enflame has spent nearly eight years independently developing and iterating through four architectural generations comprising five cloud AI chips. The company has established a comprehensive product portfolio spanning AI chips, acceleration cards and modules, intelligent computing systems and clusters, and AI computing and programming software platforms, solidifying its position as a leading domestic provider of cloud AI chips. (ITHome)
On the evening of September 3, the IPO review status for Yangtze Memory Technologies Co., Ltd. (YMTC) shifted to “Inquiry Issued” as it seeks a listing on the STAR Market. The prospectus indicates that following a return to profitability in 2024, earnings accelerated rapidly, reporting 47.042 billion RMB in revenue and 33.379 billion RMB in net profit attributable to shareholders for Q1 2026. Based on TrendForce calculations, YMTC ranked third globally and first in China among NAND Flash manufacturers by both sales volume and shipment volume during the same period. (National Business Daily)
On September 3, Anker unveiled a suite of new audio products at IFA 2026 and announced its strategic entry into the hearing aid market. The company’s inaugural hearing aid will feature its proprietary Thus AI chip, released earlier this year and currently integrated into select Soundcore earphones for features like intelligent noise cancellation. The new devices will emphasize extended battery life and rapid charging, with further details to follow. Among the newly launched audio lineup, the standout product is the Soundcore Sleep Earbuds 4, priced at €249.99 (approximately 1,949 RMB), scheduled for release next month. (ITHome)
On September 3, Tmall officially launched the AI Space Station (Token Recharge Center), enabling users to directly purchase subscription plans from leading domestic large language model providers. The hub aggregates services from Alibaba Cloud, Zhipu AI, Kimi, and MiniMax, offering token plans, coding plans, and pay-as-you-go recharge options delivered via physical/digital keys or direct account credit. The day prior, Zhipu AI opened its official flagship store on Tmall, launching coding plan subscriptions. Search traffic surged 40 times week-over-week on opening day, signaling the integration of LLM subscriptions into mainstream e-commerce. Beyond virtual AI credits, Tmall is rapidly expanding into AI hardware. The platform now hosts computers pre-installed with AI agents, learning tablets powered by the Qwen model, smart speakers, and AI glasses. This move extends Tmall’s AI commerce from smart devices to online model services, providing consumers with a one-stop purchasing experience. (Sina Tech)
On September 3, media reports highlighted that during its H1 2026 earnings briefing, Moore Threads demonstrated a fully integrated pipeline for embodied AI, spanning synthetic data generation, large model training, simulation validation, and edge hardware deployment. The company introduced MT Lambda, a full-stack embodied AI simulation platform, and open-sourced MuJoCo Warp MUSA, a GPU-accelerated physics simulation backend built on its MUSA architecture. Successful Sim2Real deployments were verified on both quadruped robot dogs and bipedal humanoid robots. Moore Threads has established an Industrial Embodied AI Innovation Center in Wuxi in collaboration with ecosystem partners and co-founded a Joint Laboratory for Embodied AI Computing and Simulation with the National Embodied AI Application Pilot Base. Embodied AI requires synergistic AI computation, physics simulation, graphics rendering, and edge inference—a workflow that underscores the distinct advantages of full-function GPUs. (Sina Tech, ITHome)
On September 3, Horizon Robotics announced that cumulative production of its Journey® smart driving chips has surpassed 15 million units, leading the domestic market in both volume and market share. To date, Journey® chips have empowered 8 million vehicle owners, with partnerships spanning over 40 brands, including all of China’s top ten automakers. The company has secured design wins for 500 mass-produced vehicle models, nearly 130 of which target mid-to-high-level autonomous driving, bringing total order volumes to over 20 million sets. (Sina Tech)
On the evening of September 3 (Beijing Time), major U.S.-based AI platforms including ChatGPT, Claude, Gemini, and Grok experienced widespread service disruptions, leaving millions of global users facing HTTP 403 blocks and infinite CAPTCHA loops. According to Downdetector, over 12,000 incident reports were filed for OpenAI, approximately 1,200 for Claude, and around 1,000 for Grok. Technical diagnostics confirmed that this “domino-style outage” was not caused by physical hardware failure. Instead, the root cause traced to Cloudflare, the unified frontend gateway for much of the AI industry, which triggered an edge risk-control circuit breaker. Minor localized service fluctuations inadvertently maxed out security thresholds, triggering cascading network-wide risk controls and false positives. Official explanations are pending. All affected services have since resumed normal operations. (Quick Command, Cailian Press)
On the evening of September 3, U.S. Secretary of Commerce Howard Lutnick stated that the administration is evaluating a new round of import tariffs on semiconductors, potentially raising duties on foreign chips to incentivize manufacturing reshoring. In media interviews, Lutnick suggested the government may introduce tariff exemptions for companies building facilities in the U.S., mirroring current policies applied to the pharmaceutical sector. The President previously imposed a 25% tariff on select advanced semiconductor products. Following a January trade investigation by the Department of Commerce, import duties were enacted. The proposed measures may extend beyond chips alone, likely encompassing imported consumer electronics and data center servers. Critics warn that further tariff hikes could inflate data center construction costs. Lutnick countered that tariff relief for domestic investors could mitigate this impact. He noted that companies like TSMC and Micron are actively constructing U.S. facilities, with approximately $1.2 trillion (roughly 8.08 trillion RMB) already committed to American semiconductor infrastructure. (ITHome)
On September 3, foreign media reported that Commerce Secretary Howard Lutnick stated Anthropic has restored a positive relationship with the Trump administration. Lutnick told reporters the company has repaired ties with the White House. This follows a late August 27 federal court ruling supporting Anthropic’s position regarding the application of AI in combat during disputes with the U.S. military. The conflict originated when Defense Secretary Pete Hegseth terminated contracts with certain defense firms linked to Anthropic after the company refused to permit its Claude model for domestic surveillance or autonomous weapons systems, prompting Anthropic to file a lawsuit in California. Neither Anthropic nor the Department of Commerce commented. (BianNiuShi)
On September 3, foreign media quoted OpenAI CEO Sam Altman stating, “We will definitely develop humanoid robots, as well as other forms of humanoid robots.” However, Altman emphasized that not all machines require a human shape. When anthropomorphic designs offer no practical advantage, OpenAI will engineer non-anthropomorphic robots tailored to specific use cases. Explaining his stance, he noted, “There will certainly be data center robots in various forms. But the robot’s shape is not the most critical factor; what truly matters is the robot’s brain.” (STAR Market Daily)
On September 3, media citing confidential documents submitted to the supervisory board reported that Volkswagen will phase out the SEAT brand no later than the end of 2029. “The SEAT brand will be gradually withdrawn from the market in an orderly and low-cost manner by the end of 2029, while ensuring continued support services for existing customers and fulfilling all existing obligations,” the document states. A VW spokesperson declined to comment on the specifics. Reports indicate that SEAT’s sister brand, Cupra, will continue operations. (Sina Tech)
On September 3, Meta executives Maher Saba and Santosh Janardhan circulated an internal memo updating expectations for engineers, explicitly stating that “we will not use AI adoption dashboards or token usage to evaluate employee impact.” While noting that 93% of code changes are currently assisted by AI agents, the memo clarified that adopting AI “is not the end goal in itself.” (Cailian Press)
On September 3, Porsche initiated a restructuring of its sales and marketing divisions, overhauling its international sales network. The number of regional sales territories will be reduced from five to four. The consolidation also involves dissolving headquarters functions for the European region and the Overseas & Emerging Markets division, decentralizing responsibilities directly to individual markets. The new organizational structure is slated to take effect on October 1, 2026. (Cailian Press)