15
0Hi~~ Good evening, everyone! It's Saturday, September 19, 2026, and you're listening to Goodnight Coffee. I'm your host, An'an. I'm so happy to be with you again! Tonight, let's talk about some fresh, hot business and tech updates—easy listening, happy chatting. Ready? Let's go～
@@@
According to 36Kr, citing AI Frontline, on September 18, Google skipped the minor version updates in the Gemini 3 series and went straight to Gemini 4. The previously planned 3.5 Pro, originally slated for a June release, was delayed; in July, Google said it had begun pretraining Gemini 4. Several developers were matched with "gemini-3.8-flash" in Arena's anonymous Battle Mode, and its capabilities far exceeded those of the publicly released 3.8 Flash on September 2. Testers say it was actually Gemini 4 Pro under testing, codenamed "Argon," though Google has not confirmed this. A leaked Arena blind-test leaderboard claims Gemini 4 Pro ranked highest in all 13 test groups, against rivals including GPT-6 Astra and others. It scored 88.7% on DeepSWE v1.1, above Astra's 86.9%; 2,064 Elo on GDPval-AA v2, versus Astra's 1,994; and 72.1% on HLE, nearly 5 percentage points ahead. Its pricing is $2.25 per million input tokens and $11.25 per million output tokens, about one-fifth of Astra's. The results table has not been officially acknowledged and differs from public leaderboard metrics.
@@@
According to Sina Tech, Anthropic is moving ahead with what could become the largest IPO in history. Although CEO Dario Amodei has called for more guardrails on AI development, the company's IPO preparations are continuing. According to four people familiar with the matter, it expects annualized revenue this year to exceed $100 billion, up from $65 billion as of July, and investors are using that to argue for a potential $2 trillion valuation. Two sources say the company could publicly file financial documents as soon as the coming weeks, with shares possibly beginning to trade as early as November; plans may still change. Investors expect it to have about 5 gigawatts of compute by the end of this year, roughly doubling by the end of next year. On Thursday, Anthropic held a customer event at its San Francisco headquarters, with about 100 venture investors attending. Co-founders Jared Kaplan and Ben Mann, Claude Code lead Boris Cherny, and researcher Andrej Karpathy presented. Company executives and bankers have also met potential IPO investors in recent weeks. The event emphasized AI's positive side, such as in biology research, and played down open-source competition. Amodei's safety warnings have raised potential liability concerns. OpenAI has pushed its IPO to 2027.
@@@
According to TechWeb, 21st Century Business Herald, citing people in the industrial chain, says Tesla's team arrived in Ningbo on September 16 and launched an Optimus humanoid robot mass-production factory audit on September 17. This round of audits will effectively lead to mass production, and Tesla has already placed relevant orders with supply chain companies. The report says Tesla plans to roll about 50,000 Optimus units off the line in 2026 and deploy them to global Gigafactories; ordinary consumers still won't have a channel to buy them. Tesla also started a round of robot factory audits in China last September, involving several traditional auto supply chain companies.
On September 18, Tesla did not respond. Tuopu Group said it had not yet received notice of the audit, but the two sides cooperate on linear and rotary actuators, dexterous-hand actuators, and other areas, and its robot component base is ready. Joyson Electronics declined to comment, citing confidentiality. Sanhua Intelligent Controls said it was not aware of the new audit round and is working with a major customer to develop robot joint actuators. The report also says Tesla has placed orders for about 5,000 Optimus units with multiple Chinese suppliers—its first thousand-unit-level mass-production order. Musk previously said Optimus production would be extremely slow in the early stage, with high-volume production beginning in 2027.
@@@
According to cnBeta, citing Kuai Technology, the iPhone 18 Pro series officially went on sale on September 18, starting at 9,999 yuan. When media visited an Apple retail store in Wuhan, a staff member said the black version of the iPhone 18 Pro Max sold out first and was the most sought-after color; restock timing is still unclear, with an expected arrival in November. Apple's website shows that orders placed now for the Pro Max 256GB require a 15-business-day wait, while the Pro requires 11 business days; JD.com's Apple self-operated store expects the Pro to be in stock after September 25. In terms of specs, the series debuts Apple's first 2nm phone chip, the A20 Pro. It has a 6-core CPU with 2 performance cores and 4 efficiency cores, a 7-core GPU with about 40% better graphics performance, 50% higher memory bandwidth, dual 16-core neural engines, and direct contact with a VC vapor chamber inside the chassis for more stable sustained performance. The Pro Max uses a 6.9-inch Super Retina XDR display, while the Pro uses a 6.3-inch display, and the Dynamic Island is further reduced. Apple Intelligence is still unavailable on the mainland China version and will need a later rollout; mainland China users cannot yet experience the full AI features. The Pro Max offers up to 43 hours of video playback, setting a new iPhone battery-life record.
@@@
According to The Verge, in The New York Times lawsuit against OpenAI and Microsoft, recently unsealed court documents show that both companies internally warned their AI content strategy was triggering a "doom loop" that would harm both model performance and the web as a whole, and was pushing the web toward a "Google Zero" situation with zero traffic. A 92-page document called scraping data to train models "the largest labor theft in human history" and said it "completely mocks the concept of fair use," adding that it could pose an "existential threat" to publishers such as The New York Times. Many of these quotes come from Brent Hecht, Microsoft's director of applied sciences. A Microsoft spokesperson responded that the comments reflect only one employee's personal views, are not legal analysis, and do not represent the company's position. Jordan Usdan, Microsoft's general manager of AI data strategy and operations, said in another court filing that Hecht holds dissenting, academic, and forward-looking views and does not specifically represent Microsoft. The documents also show that Microsoft CEO Satya Nadella acknowledged chatbots have largely replaced search, so users no longer need to visit the original sources directly. One internal document called it "highly unusual" for an end product to threaten the economic foundation of its key suppliers, but said the large language model business and content supply chain had already created that situation. OpenAI co-founder Greg Brockman was said to be more focused on the astronomical revenue commercial AI could generate. The documents also quote Sam Altman and other OpenAI employees.
@@@
According to Readhub, ChangXin Memory Technologies (CXMT) is planning to enter the flash memory market. It intends to build a NAND flash R&D production line and establish related R&D operations at a new plant in Beijing, and will later compete head-on with Samsung, SK Hynix, and Micron. CXMT has already approached customers about its NAND business plan; one startup plans to purchase its NAND chips for the R&D and deployment of storage products for AI systems and supercomputers. Meanwhile, CXMT's performance in the DRAM market is also growing. In the second quarter of 2026, its DRAM revenue reached $14.624 billion, up 99.3% quarter over quarter, with market share rising from 7.6% in the first quarter to 9.5%.
@@@
According to Sina Tech, on the afternoon of September 18, during Huawei Connect 2026, Zhou Yuefeng, a Huawei board member and CEO of Huawei Cloud, said in a speech that in the agentic era, the core of infrastructure is no longer just providing compute. Instead, it must make every token more efficient, enable general computing and intelligent computing to work in concert, allow models to keep learning, and let agents run safely and reliably in real production environments. Huawei Cloud defines a new Agentic Infra paradigm with "efficient tokens, enhanced memory, integrated general-intelligent computing scheduling, and secure autonomy," and has already served more than 3,500 customers. The Lingqu Ascend 950 intelligent computing cluster service ensures stable cloud training for 40 days and fault recovery within 10 minutes; token throughput is 20% higher than the previous generation of compute, with domestic commercial availability on September 30 and overseas commercial availability on November 30. CMS memory storage builds petabyte-level memory space, with storage capacity twice that of comparable industry products and 50% higher terabyte-level read performance. Agentic MaaS supports deployment-free, one-click access to SOTA models from leading vendors, and MiniMax demonstrated multimodal large-model capabilities integrated with Huawei Cloud. Huawei Cloud uses Zhiguo AgentArts and openJiuwen to build an enterprise-grade agent platform, opening more than 5,000 general assets and more than 1,000 industry assets, and has served over 100 enterprises. Zhiguo AgentArts will be commercially available overseas on December 30; the openJiuwen community has more than 50,000 stars and over 3.29 million downloads. The Industry AI Dream Factory covers five major industry zones, with more than 1,000 implemented projects; the government intelligence and AI hardware zones are newly launched.
@@@
According to Readhub, XPeng Group is partnering with Volcano Engine to bring agents into frontline carmaking. XPeng is building an "AI-native organization." In its first AI open competition, more than 21,000 participants took part, with support from Volcano Engine. XPeng is encouraging frontline employees to become agent creators and has built an accelerator from demo to production. Several agent applications have already been deployed: a supply chain risk management agent covers 800 suppliers, greatly improving risk screening and reporting efficiency; the Vision-X creation platform uses Volcano Engine models to visualize concepts and produce content efficiently; and an onboarding assistant agent automates the onboarding process end to end. In addition, Volcano Engine and XPeng are deeply collaborating in multiple areas. The AI middle platform already has 2,000 active users and 5,000 agents, and 50 digital employee scenarios are being rolled out. The two sides have also jointly built digital employees such as AI outbound calling, making AI a productivity tool available to all employees.
@@@
According to Sina Tech, on September 19, the Beijing Municipal Market Supervision Administration, in accordance with the work deployment of the State Administration for Market Regulation and preliminary verification, opened investigations under the Anti-Unfair Competition Law, the E-Commerce Law, and other laws and regulations into four online hotel and travel booking platforms suspected of relevant illegal acts: Hangzhou Taomei Aviation Services Co., Ltd. (Fliggy), Tongcheng Network Technology Co., Ltd., Tujia Online Information Technology (Tianjin) Co., Ltd., and Beijing Sankuai Information Technology Co., Ltd. (Meituan). All four platforms responded the same day that they would cooperate with the investigations and implement regulatory requirements. Tujia said it would improve platform governance based on the business characteristics of the homestay industry and work with hosts and merchants to build a healthy and sustainable homestay industry environment; its business is currently operating normally. Meituan said its Meituan Hotel & Travel and other online hotel booking platforms received notice and would cooperate with the investigation, working with all industry parties to build a sustainable market environment. Fliggy said it would safeguard a fair, open, and orderly market environment and promote high-quality development of the industry. Tongcheng said it would protect the legitimate rights and interests of consumers and partners; its business is running normally, and it will continue to improve compliance and jointly maintain a high-quality, sustainable market environment.
@@@
According to Readhub, Moxin Technology, founded by Chen Tianrun, a post-2000 PhD student at Zhejiang University, is about to complete a new funding round of as much as 1 billion yuan, with its valuation potentially jumping to nearly 10 billion yuan and making it a unicorn. In the past eight months, the company raised nearly 1 billion yuan cumulatively at a valuation of nearly 4 billion yuan. The report says the acceleration was driven by the growing popularity of the world-model track and Chen Tianrun's connection as a fellow disciple of DeepSeek's founder. Moxin Technology started in 3D printing and at the end of 2024 pivoted to tackling 4D world models. It has launched related products and applied them in multiple fields, with investors including state-owned capital and industrial capital. At the same time, a Gen Z founder army has emerged in AI: Hong Letong's Axiom and Lingxin Qiaoshou, in which Jia Xiaoyou is involved, have both become unicorns valued at over 10 billion yuan. The report notes that the average age of AI unicorn founders has dropped sharply, due to new tracks having no historical baggage, research results being easier to commercialize, and primary-market capital flowing toward new directions.
@@@
Alright, that's all for today's show! Thank you for being with us, and sweet dreams tonight～ If you enjoyed our program, we'd really appreciate it if you subscribe and share it with your friends—it helps us a lot. See you next episode!
