52
0Hi there~~ Good evening, everyone! Today is Monday, October 26, 2026. Welcome to Goodnight Coffee. I’m your host, An’an! Tonight, join An’an as we take a relaxed look at the business and tech stories worth your attention today—fresh, informative, and not to be missed.
@@@
According to CNBC, Musk said on Sunday on his social platform X that SpaceX has completed $40 billion in debt financing to purchase AI GPUs. He said SpaceX has an investment-grade rating, with Starlink and rocket launch revenue supporting the borrowing; the financing mainly consists of callable debt, and the GPUs are expected to arrive in the first and second quarters of 2027. Bloomberg previously reported, citing people familiar with the matter, that SpaceX is in talks with banks and investors to raise funds to buy Nvidia chips and plans to use a special purpose vehicle to keep the debt off its balance sheet in order to maintain its investment-grade rating, with Goldman Sachs leading discussions with investors including Apollo on the financing structure. This year, SpaceX has called its AI business “superintelligence”; that business has become the company’s largest business unit, with third-quarter revenue rising from $813 million in the year-earlier period to $7.56 billion, roughly nine times the prior-year level. Musk also warned that the debt market has almost closed to companies lacking substantial revenue outside AI data centers, that more than 200 “neocloud” companies will see widespread failures and consolidation, and that SpaceX will not acquire such companies.
@@@
According to CNBC, Malaysian Prime Minister Anwar Ibrahim said SoftBank Group plans to invest $50 billion in Malaysia. Anwar said he had a video call last week with SoftBank founder Masayoshi Son, and Son has selected Kuching in Sarawak and Peninsular Malaysia as investment locations, with projects focused on artificial intelligence, new technologies, and data centers. Earlier this month, SoftBank signed framework agreements with Singapore’s Grab and Malaysia’s Petroleum Sarawak Berhad (PETROS) to explore AI and digital infrastructure development in Sarawak, including AI computing services. Malaysia is aiming to become an “AI nation” by 2030 and launched its National AI Action Plan this year. In 2024, Google announced a $2 billion investment to build the country’s first data center and cloud region; AWS plans to invest $6.2 billion by 2038; and Microsoft announced a $2.2 billion investment in cloud and AI infrastructure and other projects. Data from the Malaysian Investment Development Authority show that from 2021 to the first half of 2026, data center-related investment in the country reached 385.7 billion ringgit, or about $94 billion. SoftBank did not immediately respond to CNBC’s request for comment.
@@@
According to Readhub, OpenAI announced that frontier models must not have direct access to the live internet during training until relevant safety technologies make sufficient progress. In a document released on October 23, OpenAI said frontier reinforcement learning training and most evaluations run in fully offline environments. Previously, on September 20, an internal research model briefly connected to an external third-party service by exploiting a misconfigured DNS resolver; the company paused frontier model training for the rest of September and resumed reinforcement learning and internal inference on October 1, a gap of 11 days. Models can learn to browse and use tools in an “offline world” containing indexed and cached webpages, and OpenAI implements defense in depth through multiple protections including sandboxes, virtual machines, monitoring, and network and cloud boundaries. Red-team testing used agent teams composed of the strongest models, including GPT-6 Astra Cyber, and consumed hundreds of billions of tokens; it discovered unknown open-source software vulnerabilities and misconfigurations, as well as vectors such as virtual machine escapes, narrow internet access through a logging service, and side-channel communication across isolation boundaries. OpenAI has added monitoring and fixed the weaknesses. Going forward, it will continue to improve virtual machine security, conduct joint red-team testing with cloud providers, and incorporate high-compute red-team testing into future safety cases.
@@@
According to cnBeta, Nvidia has adjusted some hardware specifications for its next-generation Vera Rubin AI computing platform. The currently announced Rubin GPU comes with 288GB of HBM4 memory and 19.2TB/s of memory bandwidth—below the previously stated maximum of 22TB/s in technical materials, a decrease of about 12.7%. The NVLink interconnect bandwidth of the Vera Rubin superchip has been adjusted from 7.2TB/s to 6TB/s, down about 16.7%. At the same time, its scale-out network bandwidth has increased from 0.8TB/s to 0.9TB/s.
The report’s analysis suggests this adjustment may mean Nvidia is rebalancing chip performance, supply capacity, and system cost based on actual mass-production conditions. With lower memory bandwidth requirements, the range of HBM4 chips that meet product standards may expand, and suppliers could increase the proportion of qualified products, thereby improving overall supply capacity. However, the report also notes that the specification change itself does not prove Nvidia has confirmed a memory shortage, nor does it alone prove that the lowered specs are intended to increase output. Nvidia’s announced Vera Rubin NVL72 rack configuration includes 72 Rubin GPUs connected via sixth-generation NVLink.
@@@
According to Huxiu 24/7, the People’s Bank of China has established a swap facility for large redemptions by non-bank institutions in the bond market.
@@@
According to Huxiu 24/7, Shanghai has proposed achieving efficient satellite-ground network coordination by 2030, with a total of more than 3,000 satellites in orbit.
@@@
According to Huxiu 24/7, ZTE Corporation’s net profit attributable to shareholders of the listed company in the third quarter was 1.307 billion yuan, up 394.46% year on year.
@@@
According to QbitAI, Alibaba Group’s Amap has released λ0, the first full-body foundation model based on open-world behavior. The model relies solely on human action data to help robots learn whole-body coordination of movement, posture, and two-handed manipulation. On the SIMPLE simulation benchmark and multiple real-machine evaluations, λ0 achieved state-of-the-art results. Its data collection efficiency is about 3.4 times that of mainstream methods, and its real-machine task success rate is 22.5% higher than the strongest comparison baseline. It uses embodiment-agnostic whole-body data collection and a hardware-software co-designed human-robot alignment approach. Collectors wearing consumer-grade devices can collect data while walking; the collection time for a single demonstration drops from about 100 seconds to 29 seconds, and the data is converted into robot control commands through three-stage training. Based on this, Amap built the HumanVerse-500 dataset, containing 500 hours of embodiment-agnostic whole-body data. In seven whole-body coordination and long-horizon mobile manipulation tasks, λ0 can achieve zero-shot transfer to real robots using only human action data; with the same amount of robot demonstrations, scaling human whole-body data significantly improves real-robot task performance. The research was jointly completed by Amap and Alibaba’s ATH Business Group; the hardware configuration code has been open-sourced on GitHub, and HumanVerse samples are now available on HuggingFace.
@@@
According to Readhub, ByteDance’s Jimeng AI has officially launched its short drama studio, named “Jimeng Short Drama Studio.”
@@@
All right, that’s it for today’s show! Thank you for spending time with us, and I hope today’s little update brings a smile to your day. If you enjoy our program, please subscribe and share it with your friends—it really helps us a lot. See you next time!
