节目简介
来源：小宇宙
Hi everyone~~ Good evening! Today is Sunday, November 1, 2026, and welcome to Goodnight Coffee. I’m your host, An’an. Tonight, let’s talk about new developments in business and technology—relaxed, fun, and full of substance—to keep you company for a pleasant evening. Come along with me!
@@@
According to Readhub, Apple has reclaimed the title of the world’s most valuable company, overtaking Nvidia. Amid concerns over AI spending, investors turned to Apple for relative safety. Apple shares fell 1.2% that day, giving it a market value of $5.03 trillion, while Nvidia fell 4.2% to a market value of $4.93 trillion. This is the first time Apple has overtaken Nvidia again since briefly surpassing it in late July.
@@@
According to Readhub, OpenAI has lowered its latest fundraising target from an original $30 billion by $10 billion to about $20 billion, with the corresponding valuation adjusted from $1.4 trillion to $1.25 trillion, and has asked investors to complete their funding commitments by Friday. Affected by the selloff in AI stocks in public markets and the impending listing of top rival Anthropic, the fundraising process is under pressure. Some investors have reduced their subscriptions, and the first closing is expected to bring in less than $20 billion. A UAE-related investment consortium will be the largest investor in this round, while Founders Fund, Coatue, Sequoia Capital, Andreessen Horowitz and SoftBank also plan to participate. A retail allocation for wealthy individual investors will follow, and total fundraising is expected to reach around $20 billion.
@@@
According to 36Kr, Meta plans to raise nearly $100 billion, and Oracle has put in the “final offer.”
@@@
According to 36Kr, Google engineers recently sent a prompt to a prototype satellite in orbit via a ground station: “Write a haiku about being in space.” The satellite’s TPU started up and ran the pre-deployed Gemma 4 to complete inference. As the satellite passed near Svalbard, Norway, the ground station received the returned haiku. Google CEO Pichai called it Project Suncatcher’s “Hello World” moment, and the first time a data-center-class TPU had generated and sent back a complete AI response in orbit. The satellite was developed by Google in partnership with Planet and launched into orbit aboard SpaceX Transporter-18 on October 1. The team completed debugging and started the TPU within weeks, and only this week carried out the actual generative AI task. The satellite carries four Google TPUs and can draw up to about 1 kilowatt of power, but only one TPU can be started at a time, for about 20 minutes. Google originally planned to launch two prototype satellites in early 2027 to test
in-orbit TPUs and inter-satellite optical communications, but adjusted the schedule and sent the TPU into orbit ahead of time in October. Next, it will move to more practical system tests to observe how the TPU performs under orbital thermal stress and radiation.
@@@
According to cnBeta, the FBI has confirmed that a North Korean hacker group carried out the Binance theft, involving $6.3 billion.
@@@
According to Readhub, the official DeepSeek Harness team published an article introducing the design approach of its Super-agent architecture for engineering practice. The article notes that traditional single-session agents have limitations such as handling multiple tasks, and summarizes the new autonomous agent paradigm with four characteristics: continuous, efficient, proactive and persistent. The architecture consists of five parts, including a dual-layer architecture and event-driven design, with multiple optimization mechanisms. Experiments show that the architecture outperforms traditional single-session solutions in task completion efficiency and routing throughput. At present, an internal instance based on this approach is already in operation. The team has also released a prototype implementation for technical validation, while noting that the architecture still has several unresolved issues.
@@@
According to Sina Tech, ChangXin Memory Technologies said at its Q3 results briefing that its sixth-generation process technology platform based on the 4F² vertical-channel DRAM architecture has completed engineering sample validation. This is the first time ChangXin has announced that its sixth-generation platform adopts the 4F² architecture. Completion of engineering sample validation means the platform’s core performance indicators have met design expectations on real wafers, and after client validation it can move into mass production. ChangXin previously said in a public speech that it expects to launch a new-generation DDR5 RDIMM product incorporating 4F² technology for the server sector by the end of the year. According to publicly disclosed progress, Samsung and SK hynix have listed 4F² as a key direction below the 10nm class, and ChangXin is the first to complete engineering sample validation. If subsequent progress goes smoothly, ChangXin is expected to become the world’s
first company to implement 4F². A 4F² cell area is about two-thirds that of a mainstream 6F² cell, and is seen as a key path to breaking the DRAM scaling bottleneck and moving toward 3D DRAM. The core of ChangXin’s approach is a self-developed VSG vertical surrounding-gate transistor, which can alleviate short-channel effects and leakage. The architecture decouples DRAM and CMOS manufacturing processes and uses wafer-level hybrid bonding for high-speed interconnection, allowing logic performance, memory density and reliability to be optimized separately.
@@@
According to Huxiu 24 Hours, the People’s Bank of China will hold a symposium on relending for stock buybacks and increased holdings.
@@@
According to Readhub, Xiaomi Auto delivered more than 50,000 vehicles in October.
@@@
According to QuantumBit, XHEART made its first public appearance and launched the X7 Ultra chip, saying this is the first time a world model has guided chip design. The X7 Ultra delivers 700 TOPS of computing power, with NPU utilization of 98.7%; the output of quantized models has a similarity of over 99.99% to the FP32 baseline. Compared with industry flagships, it reduces memory bandwidth consumption by 80% and doubles energy efficiency. Its “large-model compute” evaluation directly uses production intelligent-driving models such as the Momenta R7 world model for testing, rather than relying only on laboratory benchmarks. The chip adopts a super-core NPU architecture and provides dedicated engines for high-frequency operators such as LayerNorm and Softmax. Its AI toolchain reduces data movement through computational graph fusion and parallel execution; the lidar accelerator processes 20 million points per second, and camera pipeline stability is improved by 100%. The CPU adopts the
Arm v9 automotive architecture and is implemented in China for the first time; the SoC reaches ASIL-B, and the integrated MCU reaches ASIL-D. XHEART has no direct equity affiliation with Momenta; Momenta founding team members Cao Xudong and Xia Yan hold indirect stakes. The X7 has already entered mass production in new SAIC Volkswagen vehicles. Cao Xudong said that “large models define chips” is Momenta’s third important judgment.
@@@
Alright, that’s all for today’s show! Thank you all for staying with us, and may good moods always be with you. If you enjoy our program, you’re very welcome to subscribe and share it with your friends—that helps us a lot. See you in the next episode!
@@@
According to Readhub, Apple has reclaimed the title of the world’s most valuable company, overtaking Nvidia. Amid concerns over AI spending, investors turned to Apple for relative safety. Apple shares fell 1.2% that day, giving it a market value of $5.03 trillion, while Nvidia fell 4.2% to a market value of $4.93 trillion. This is the first time Apple has overtaken Nvidia again since briefly surpassing it in late July.
@@@
According to Readhub, OpenAI has lowered its latest fundraising target from an original $30 billion by $10 billion to about $20 billion, with the corresponding valuation adjusted from $1.4 trillion to $1.25 trillion, and has asked investors to complete their funding commitments by Friday. Affected by the selloff in AI stocks in public markets and the impending listing of top rival Anthropic, the fundraising process is under pressure. Some investors have reduced their subscriptions, and the first closing is expected to bring in less than $20 billion. A UAE-related investment consortium will be the largest investor in this round, while Founders Fund, Coatue, Sequoia Capital, Andreessen Horowitz and SoftBank also plan to participate. A retail allocation for wealthy individual investors will follow, and total fundraising is expected to reach around $20 billion.
@@@
According to 36Kr, Meta plans to raise nearly $100 billion, and Oracle has put in the “final offer.”
@@@
According to 36Kr, Google engineers recently sent a prompt to a prototype satellite in orbit via a ground station: “Write a haiku about being in space.” The satellite’s TPU started up and ran the pre-deployed Gemma 4 to complete inference. As the satellite passed near Svalbard, Norway, the ground station received the returned haiku. Google CEO Pichai called it Project Suncatcher’s “Hello World” moment, and the first time a data-center-class TPU had generated and sent back a complete AI response in orbit. The satellite was developed by Google in partnership with Planet and launched into orbit aboard SpaceX Transporter-18 on October 1. The team completed debugging and started the TPU within weeks, and only this week carried out the actual generative AI task. The satellite carries four Google TPUs and can draw up to about 1 kilowatt of power, but only one TPU can be started at a time, for about 20 minutes. Google originally planned to launch two prototype satellites in early 2027 to test
in-orbit TPUs and inter-satellite optical communications, but adjusted the schedule and sent the TPU into orbit ahead of time in October. Next, it will move to more practical system tests to observe how the TPU performs under orbital thermal stress and radiation.
@@@
According to cnBeta, the FBI has confirmed that a North Korean hacker group carried out the Binance theft, involving $6.3 billion.
@@@
According to Readhub, the official DeepSeek Harness team published an article introducing the design approach of its Super-agent architecture for engineering practice. The article notes that traditional single-session agents have limitations such as handling multiple tasks, and summarizes the new autonomous agent paradigm with four characteristics: continuous, efficient, proactive and persistent. The architecture consists of five parts, including a dual-layer architecture and event-driven design, with multiple optimization mechanisms. Experiments show that the architecture outperforms traditional single-session solutions in task completion efficiency and routing throughput. At present, an internal instance based on this approach is already in operation. The team has also released a prototype implementation for technical validation, while noting that the architecture still has several unresolved issues.
@@@
According to Sina Tech, ChangXin Memory Technologies said at its Q3 results briefing that its sixth-generation process technology platform based on the 4F² vertical-channel DRAM architecture has completed engineering sample validation. This is the first time ChangXin has announced that its sixth-generation platform adopts the 4F² architecture. Completion of engineering sample validation means the platform’s core performance indicators have met design expectations on real wafers, and after client validation it can move into mass production. ChangXin previously said in a public speech that it expects to launch a new-generation DDR5 RDIMM product incorporating 4F² technology for the server sector by the end of the year. According to publicly disclosed progress, Samsung and SK hynix have listed 4F² as a key direction below the 10nm class, and ChangXin is the first to complete engineering sample validation. If subsequent progress goes smoothly, ChangXin is expected to become the world’s
first company to implement 4F². A 4F² cell area is about two-thirds that of a mainstream 6F² cell, and is seen as a key path to breaking the DRAM scaling bottleneck and moving toward 3D DRAM. The core of ChangXin’s approach is a self-developed VSG vertical surrounding-gate transistor, which can alleviate short-channel effects and leakage. The architecture decouples DRAM and CMOS manufacturing processes and uses wafer-level hybrid bonding for high-speed interconnection, allowing logic performance, memory density and reliability to be optimized separately.
@@@
According to Huxiu 24 Hours, the People’s Bank of China will hold a symposium on relending for stock buybacks and increased holdings.
@@@
According to Readhub, Xiaomi Auto delivered more than 50,000 vehicles in October.
@@@
According to QuantumBit, XHEART made its first public appearance and launched the X7 Ultra chip, saying this is the first time a world model has guided chip design. The X7 Ultra delivers 700 TOPS of computing power, with NPU utilization of 98.7%; the output of quantized models has a similarity of over 99.99% to the FP32 baseline. Compared with industry flagships, it reduces memory bandwidth consumption by 80% and doubles energy efficiency. Its “large-model compute” evaluation directly uses production intelligent-driving models such as the Momenta R7 world model for testing, rather than relying only on laboratory benchmarks. The chip adopts a super-core NPU architecture and provides dedicated engines for high-frequency operators such as LayerNorm and Softmax. Its AI toolchain reduces data movement through computational graph fusion and parallel execution; the lidar accelerator processes 20 million points per second, and camera pipeline stability is improved by 100%. The CPU adopts the
Arm v9 automotive architecture and is implemented in China for the first time; the SoC reaches ASIL-B, and the integrated MCU reaches ASIL-D. XHEART has no direct equity affiliation with Momenta; Momenta founding team members Cao Xudong and Xia Yan hold indirect stakes. The X7 has already entered mass production in new SAIC Volkswagen vehicles. Cao Xudong said that “large models define chips” is Momenta’s third important judgment.
@@@
Alright, that’s all for today’s show! Thank you all for staying with us, and may good moods always be with you. If you enjoy our program, you’re very welcome to subscribe and share it with your friends—that helps us a lot. See you in the next episode!