62
0Hi there~~ Good evening, everyone! It’s Monday, October 12, 2026, and I’m An’an, your host at Goodnight Coffee. A new week has begun! Tonight, let’s talk about the freshest business and tech developments and see what new trends and opportunities are emerging. Relax your ears, get your mood ready—An’an is here to spend a fun and informative stretch of time with you!
@@@
According to The Wall Street Journal, Amazon is in talks to acquire San Francisco AI startup Decart, in a deal that could value it at about $7 billion. People familiar with the matter say the deal could be reached soon if negotiations do not fall apart and no rival bidder emerges. Decart’s software makes it easier for developers to switch between different AI chips. The company was founded in 2023 by three Israeli engineers—the Leitersdorf brothers and Moshe Shalev. According to PitchBook, Decart has raised more than $400 million in venture capital from investors including Sequoia Capital, Benchmark, and Radical Ventures, and was valued at nearly $4 billion in May. The company is also developing world models called Lucy and Oasis, which can generate and simulate real-world 3D environments. Amazon has developed a custom AI chip, Trainium3, to compete with Nvidia. The report says AWS was impressed by the founders after meeting Decart in 2025 and gave it early access to Trainium3 chips. Another report says Anthropic had previously discussed acquiring Decart for about $6 billion and conducted due diligence, before abandoning the effort in September. Amazon’s deal with Decart is still in negotiations.
@@@
According to VentureBeat, Google announced Wednesday that it is launching two versions of Gemini 3.8 Flash: the standard version for agentic tasks, software development, and multistep reasoning, and Flash Cyber, which focuses on vulnerability detection and remediation. Google CEO Sundar Pichai said 3.8 Flash is a significant leap over 3.7 Flash, beating many large frontier models on the DeepSWE coding benchmark at far lower cost. Flash Cyber is Google’s strongest cybersecurity model, scoring 86.2% on CyberGym and 47.2% on CWE-Bench. Internal benchmarks show it successfully finds vulnerabilities in more than 70% of cases across 20 programming languages. This is Google’s third Flash model release in six weeks. 3.8 Flash is already available in Gemini Enterprise, and developers can try it through Google AI Studio, Android Studio, and other platforms. Pricing is $0.75 per million input tokens and $3.75 per million output tokens, the same as 3.7 Flash, with adjustable model effort. The model supports 1 million tokens of input and 64,000 tokens of output, can process text, images, audio, video, and PDFs, and scores 54.9% on HLE-Verified. Google also says it has used Flash Cyber to protect its own code, producing 2.6 times as many correct patches for Chrome vulnerabilities as a larger commercial model.
@@@
According to CNBC, Intel CEO Lip-Bu Tan and Broadcom CEO Hock Tan warned that access to capital could become the biggest constraint on the AI boom. Speaking at DBS Bank’s Asia Future Outlook conference in Singapore, Hock Tan said financing could hinder AI expansion more than shortages of semiconductors, memory chips, or electricity. Lip-Bu Tan is also worried about financing, noting that the cost of capital is rising and that over the long term people want to see returns on investment.
The report says the AI industry has invested enormous sums in computing demand, building data centers, advanced chips, and power infrastructure, but rising borrowing costs could push up financing costs and put pressure on projects that take years to pay off. Hock Tan said that over the past two years the industry has repeatedly hit bottlenecks but overcome them: the memory shortage may ease in a year, the wafer shortage may ease in 2028 to 2029, and power supply will improve as new infrastructure comes online.
Despite financing concerns, both remain bullish on near-term demand. Hock Tan said demand for computing infrastructure exceeds supply, and the imbalance could last until 2029 or even 2030. Lip-Bu Tan said the industry is unlikely to slow significantly over the next two to three years, but longer-term growth depends on how quickly enterprises adopt AI and whether the technology can deliver economic returns. Hock Tan also warned that if adoption of advanced AI models disappoints or capability improvements stall, demand for computing infrastructure could slow.
@@@
cnBeta news: Mobility Global data shows that in the first half of 2026, global sales of pure gasoline new cars were about 20.25 million units, down 10% year-on-year, accounting for 49% of new car sales. That was below 52% in the same period last year and 73% in 2021, falling below half for the first time. Traditional hybrids sold about 7.27 million units, up 10% year-on-year, accounting for 18%; pure EVs sold about 6.87 million units, up 12%, accounting for 17%; diesel, plug-in hybrids, and a few other powertrains accounted for about 16%. Electrek estimates that about 83% of new cars still have an internal combustion engine. By region, gasoline car sales fell 26% in China and 13% in Europe. Europe’s pure EV sales rose 32% to 1.81 million units, Southeast Asia rose 81% to 350,000 units, and Oceania more than doubled to 110,000 units. China’s pure EV sales fell 3% to 3.44 million units but still accounted for about half of global pure EV sales, while North America’s pure EV sales fell 15%. Nikkei said Middle East conflicts drove up oil prices and consumers focused on car ownership costs, accelerating the shift to electrification; the global auto market shrank about 5% overall in the first half. Electrek noted that this data covers only six months and falls during a period of oil price volatility, so whether the full-year share can fall below 50% remains to be seen. U.S. Cox Automotive second-quarter data shows pure EV share fell from 7.4% to 5.8%, plug-in hybrid share fell from 2% to 1.1%, gasoline and diesel vehicles combined slipped slightly from 77.6% to 76.8%, and traditional hybrids rose from 13% to a record 16.3%. After subsidies expired, some consumers shifted to hybrids rather than simply returning to pure gasoline cars. Overall, multiple powertrain routes are squeezing gasoline-only vehicles together, rather than pure EVs having fully replaced fuel-burning cars.
@@@
According to Huxiu 24 Hours, a source says Tencent plans to issue about $5 billion in bonds, including $1.5 billion to $2 billion in dollar-denominated bonds. The proceeds will fund AI spending, most of which will go toward chip purchases.
@@@
According to Huxiu 24 Hours, CICC says that on October 19, Dongxing Securities and Cinda Securities will be delisted.
@@@
According to 36Kr, citing Zhidongxi, AI-native infrastructure company Infinigence AI recently completed a 2.3 billion yuan Series C round, bringing its total financing to 4.3 billion yuan. Both the single round and cumulative total set new highs for China’s AI Infra industry. The financing has closed, and the company is preparing for an IPO. Spun out of Tsinghua University’s Department of Electronic Engineering, in 2024 it completed the world’s first “4+2” heterogeneous mixed training of a 70-billion-parameter large model, using four domestic and two imported chips to reach 97.6% compute utilization. New chip adaptation was shortened to two working days, domestic adaptation costs dropped by more than 70%, trillion-parameter MoE training throughput rose 41% to 75% over mainstream open-source frameworks, and inference costs fell 10-fold in a year. At the 2026 World AI Conference, the company proposed a “front store, back factory, one center” strategy. Its cross-domain training system, together with China Telecom, completed China’s first mixed training of a large model over more than 4,000 kilometers, improving two-location cluster performance by 165%. It has reached more than 37,000P of computing power and covers 16 types of chips. As of July, Token Factory’s daily call volume had grown 40-fold from December last year. The company operates asset-lightly and maintains an independent third-party positioning.
@@@
According to Zhidongxi in 36Kr, Zhipu recently started testing its high-speed flagship model GLM-5.3 HighSpeed in the AI coding tool ZCode. AI coding blogger Leo said he had been given gray access to the model, and its output speed exceeds 490 tokens/s. AI Ping data shows DeepSeek V4.1 Flash’s official API runs at 121.34 tokens/s, while Zhipu’s official GLM-5.3 API runs at 70.84 tokens/s. GLM-5.3 HighSpeed is several times faster than mainstream domestic models. The model is currently available only randomly to ZCode subscribers, with no clear beta application entry point. The beta speed is for reference only; ultimately, what matters is the speed when it is scaled into service after launch. A month ago, it appeared on the official website of the open-source coding agent Pi harness, and pricing has not yet been disclosed. Its context window is 1 million tokens, and it can output up to 131,072 tokens at once, consistent with GLM-5.3. In May this year, Zhipu launched a high-speed API, GLM-5.1-highspeed, for some enterprise users, with output speeds reaching 400 tokens/s. It was jointly built by the Zhipu GLM team and the TileRT team, with system-level optimizations across the inference engine, scheduling system, and underlying infrastructure.
@@@
36Kr news: Swancor Advanced Materials’ subsidiary Qiyuan Robotics began selling two consumer-grade humanoid robots, the Q1 and T1, on September 20, and will ship them in order of orders starting October 1. The Q1 is priced at 19,999 yuan, with an Explorer Edition at 26,999 yuan; the T1 is priced at 19,999 yuan, and the T1 Pro at 29,999 yuan. Swancor Advanced Materials originally made chemical new materials. In 2025, AgiBot Innovation gained control through a negotiated transfer and partial tender offer. AgiBot serves industrial and other B2B markets, while Qiyuan targets consumers. Its 2026 half-year report showed that all four core new materials subsidiaries were profitable. The Q1 focuses on small-size whole-body force control. It is about 88 cm tall and weighs about 15 kg, is foldable, and uses egg-sized miniaturized QDD quasi-direct-drive joints. The Explorer Edition opens up the SDK and hardware interfaces. The T1 focuses on transformability. It is about 100 cm tall and weighs 16 kg, and can switch between a wheeled-legged humanoid and a quadruped. The T1 Pro is equipped with Nvidia’s Orin Nano. Qiyuan has also launched a skills store and developer community, with more than 1,000 developers and more than a hundred new skills added each month. Unitree’s R1 starts at 29,900 yuan, while Booster Robotics’ BoosterK1 is priced at $5,999 overseas.
@@@
Okay, that’s all for today’s show. Thank you for joining us. May you fall asleep in a good mood and have sweet dreams too~ If you enjoy our program, you’re also very welcome to subscribe and share it with your friends—it helps us a lot. See you next time!
