57
0Hi everyone, good evening! Today is Sunday, October 25, 2026. I’m An’an, your host at Goodnight Coffee—welcome to today’s show! Coming up, I’ll walk you through the freshest business and tech updates, packed with substance, so stick around!
@@@
According to QbitAI, Tesla delivered 486,500 vehicles in the third quarter, up 1.3% quarter over quarter but down 2.1% year over year. Revenue was essentially flat at $28.046 billion; gross profit rose 9% to $5.494 billion, and automotive gross margin recovered to 18.5%. In the same period, operating profit fell 56% to $718 million, and net income attributable to common shareholders dropped 43% to $787 million, mainly due to a 39% increase in operating expenses to $4.776 billion, including a 69% jump in R&D spending to $2.761 billion. After the earnings release, Tesla shares fell more than 4.5% at one point in after-hours trading, wiping out more than $68 billion in market value. Paid active FSD users reached 1.74 million, up 67% year over year, adding 700,000 in a year; monthly subscriptions exceeded half for the first time, and in North America more than 75% of new vehicles in the third quarter came with FSD subscriptions. Musk said the order backlog is at its highest in four years; Cybercab has already entered Robotaxi operations, and next year private vehicle owners are expected to join the driverless fleet; Optimus has begun production, and investment in the Terafab chip plant could reach tens of billions of dollars. Tesla says it is pushing for broader FSD permits in China and claims its technology lead is unmatched by other Chinese OEMs.
@@@
TechWeb reports, citing foreign media coverage based on an Andreessen Horowitz podcast, that Amazon CEO Andy Jassy said on October 23 local time that, driven by continued strong demand for AI infrastructure, AWS cloud computing capacity is basically sold out through 2027, with a large volume of long-term compute orders locked in through 2028. The podcast was hosted by Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg. Jassy said demand for compute from AI labs and enterprise customers keeps coming, and Amazon has been doing its best to keep up. Reports say AWS’s 2026 capacity is already sold out, most of its 2027 capacity has been booked, and a significant portion of long-term orders for 2028 has been locked in. Against the backdrop of rapid growth in AI model training and inference demand, cloud service providers are facing sustained infrastructure pressure; Amazon and other cloud providers need to keep expanding data centers and procuring compute resources to meet growing order demand from AI labs and enterprise customers.
@@@
According to Readhub, SpaceX is discussing a $40 billion off-balance-sheet debt financing plan, intending to buy Nvidia chips through special purpose vehicles to support its AI business expansion. The structure is designed to keep the related liabilities off SpaceX’s own balance sheet in order to maintain its investment-grade rating. Pricing negotiations have become more complicated due to recent AI market turmoil that has pushed up industry borrowing costs. Currently, banks led by Goldman Sachs are in talks with investors led by Apollo; Apollo-backed Atlas SP is proposed to manage the entity, and Pimco is also considering participating. Earlier, some subordinated debt yields were discussed in the mid-8% to 9% range; recent market volatility has prompted investors to reassess pricing, and both sides are still negotiating. In the background, xAI previously tried to raise up to $20 billion in off-balance-sheet debt through a sale-and-leaseback arrangement, but auditors deemed it a failed arrangement, and about $13 billion ultimately ended up on SpaceX’s own books; the proposed structure this time is intended to avoid a repeat.
@@@
According to Readhub, OpenAI CEO Sam Altman recently cited security again in saying the company will not open-source its most advanced large models in the short term. He said that once open-sourced, the internet security system would need to be rebuilt from the ground up, and a biosecurity protection layer would also need to be added; but he also acknowledged that open-source models of the same level from other teams are not far behind OpenAI.
The report notes that the U.S. AI large-model field initially took the open-source route, but now the major players are turning to closed source. OpenAI and Anthropic both currently hold closed-source positions and are aiming their criticism at Chinese AI companies. The report says the two companies once published articles disparaging the domestic GLM-5.3 large model, which instead gave it free publicity and confirmed that its capabilities have reached the world’s top level. The two companies’ claims of promoting “safe and responsible” AI are no longer convincing: most of the disclosed AI safety incidents have come from their products, and some even question whether they are using stunts such as AI escape and website attacks for marketing. Domestic open-source AI large models have not had such accidents, and have even stepped in to help when the AI of two overseas companies launched attacks, preventing the situation from escalating.
@@@
According to The Verge, Google’s AI search feature is disrupting the online publishing industry. In the past, publishers invested time and money in optimizing web pages to rank high in Google Search and earn clicks, while Google returned exposure and traffic. In 2024, Google launched AI Overviews, showing AI-generated summaries at the top of search results that synthesize pages from across the web, with links to the sources the system considers most useful placed below the summary; users may not click through to the sources cited in the summary. The News/Media Alliance warned that the feature would have a “catastrophic” impact on web traffic. A Press Gazette survey found that in May 2024, nearly a quarter of news-related searches showed AI Overviews; by August, the feature was already affecting publisher visibility by replacing some queries that previously went to leading media websites. In May 2025, Google launched AI Mode with a chatbot-style interface, offering detailed summaries from web results. A June 2025 Wall Street Journal report showed that Business Insider, The Washington Post, HuffPost and the Wall Street Journal itself had seen traffic declines since 2022; Similarweb data showed that The New York Times’ search traffic fell from 44% to 35.5% over three years.
@@@
According to 36Kr, Today AI completed an initial funding round of more than $200 million on October 21, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, Transsion Holdings, Tongge Venture Capital, and Shanghai Jiao Tong University’s Hanyuan Assets. Today AI was founded by Qi Junyuan. Since September 9, it has launched international and domestic versions; it entered the U.S. market first in September and launched in China two weeks later, supporting platforms such as Feishu and DingTalk. The product is positioned as a “personal assistant.” With user authorization of information and app data, it accumulates memory through continuous conversation and organizes interactions around a “person” rather than tasks; users need to name the assistant at registration. The team says it will train an interaction model to lower the unit cost of intelligence and make tone, understanding and execution more closely match users. It noted that Muse surpassed 5 million downloads in 22 days, while it took ChatGPT 56 days to reach that figure, but downloads do not equal long-term value; retention, proactivity, privacy authorization and platform access remain common challenges. Manus recently launched “Cue” and completed more than $500 million in financing.
@@@
TechWeb reports that Moonshot AI released Kimi K3.1 Pro and the high-speed K3.1 Pro Highspeed in the early hours of today. K3.1 Pro has 3 trillion-level parameters, supports 1 million tokens, and is positioned as its strongest model to date; the high-speed version has comparable intelligence and is 2 to 3 times faster than the standard version. The former is a post-training upgrade of K3, with a big increase in compute and large-scale reinforcement learning to improve capabilities. Both are now available through Kimi products and the API, with full weights planned for release by November 4. The company says its code, agent and knowledge benchmarks reach or approach those of the strongest closed-source models: Terminal-Bench 4.0 scored 65.7%, above GPT-6 Astra’s 58.2% and slightly below Claude Opus 5.5’s 66.4%; the Code Bench long-horizon test scored 61.2%, above K3’s 55.95%. It scored 92.7% on the 3D modeling and game development test BenchCAD; OfficeQA Pro v2 scored 64.4, which the company calls the current best. Hallu Bench Hard’s factual statement error rate dropped from 23.4% to 8.3%, and the proportion of answers containing errors was about 42.4% of the previous generation; safety test pass@3 was 81.0% for the released version and 90.5% for the unfiltered version.
@@@
According to Sina Tech, XHEART recently announced the completion of a strategic funding round, bringing cumulative financing to nearly RMB 5 billion. Investors include automakers and industrial strategic investors such as SAIC, Chery, ChangXin Technology, CICC Renault and Joyson; investment institutions including CDH Investments, IDG Capital, GGV Capital, NIO Capital, Shunwei Capital, Blue Lake Capital, Cathay Capital, ZhenFund, Joy Capital, Gaocheng Capital, Chentao Capital, Dingpei Group, Zhongwei China, Shuimu Alumni Fund, Zhongqing Capital and Runzhang Venture Capital; and state-owned investors including the National Artificial Intelligence Industry Investment Fund, China Development Bank Financial, Suzhou State-owned Investment Group, Suzhou Venture Capital, Shenzhen Capital Group, Suzhou High-speed Rail Xincheng Capital and Shanghai Semiconductor Industry Investment. Proceeds from this round will be used for mass production delivery and commercialization of existing chip products, as well as R&D of next-generation chips. Through dedicated chips designed for large-model requirements, the company provides hardware support for physical AI applications such as autonomous driving and robotics, and works with partners including Momenta and QNX to advance coordination among chips, algorithms and basic software, helping automakers accelerate the deployment of intelligent driving.
@@@
According to LatePost, Yinwang and CATL’s first battery-swapping models are tentatively set at five, with plans to go on sale in 2027. Currently, the five models are a tentative number, and the 2027 launch is a planned arrangement.
@@@
According to Huxiu 24 Hours, the 5,000th personal robot from Qiyuan Robotics, a subsidiary of Swancor Advanced Materials, has rolled off the production line.
@@@
Alright, that’s all for today’s show! Thank you for spending time with us. I hope today’s content brought you a little relaxation and a good mood. If you enjoy our show, you’re very welcome to subscribe and share it with your friends—it really helps us a lot. See you next time!
