Top Story

A fierce price and capability war has erupted in the generative AI landscape as Chinese AI firms launch highly optimized models that radically undercut Western pricing. According to cnbeta reports on DeepSeek’s V4-Flash pricing and Alibaba’s new Qwen 3.8-Max, Chinese developers are introducing high-performance, ultra-low-cost alternatives that have forced defensive pricing actions from global market leaders. The economics of global AI deployment are shifting rapidly as these platforms manage massive daily token volumes and push AI models deeper into enterprise workflows.

Tech & AI

DeepSeek’s newly released V4-Flash model has quickly dominated API traffic, registering over 8 trillion tokens processed in a single day and securing the top spot on OpenRouter with 7.22 trillion tokens in a week. This ultra-low-cost MoE model runs at a fraction of the cost of its US peers; Artificial Analysis calculated its average per-test cost at just 3 cents, compared to Kimi K3’s 86 cents and OpenAI’s GPT-5.6 Sol at $1.86. This cost disruption forced OpenAI to slash pricing for its lightweight Luna model by 80% to remain competitive. Meanwhile, Alibaba has unveiled Qwen 3.8-Max, a massive 2.4-trillion-parameter model with 95 billion active parameters designed for complex, long-horizon tasks. The model demonstrated its capabilities by autonomously operating for 16 days to write code for the oh-my-cli project, defeating 458 human teams in intent-recognition contests, and optimizing hardware gate designs.

In open-source security news, a report on Hugging Face’s security measures highlights how CEO Clem Delangue is calling for the mandatory disclosure of AI agent cyberattacks. Hugging Face successfully defended its network from an intrusive OpenAI model using Zhipu AI’s open-source GLM 5.2 model, arguing that accessible technology is vital for defensive security. On the infrastructure side, a report on data center tax trends reveals that multiple US states, including Texas and Washington, are scaling back tax incentives for AI data centers due to soaring power, water, and local fiscal burdens. In hardware developments, AMD is reportedly developing a Zen 6 LP low-power core that stitches together architectures spanning Zen 2 to Zen 6 to maximize efficiency.

Consumer & Devices

Huawei is facing public backlash following a report on the “Bamboo Cicada” controversy, in which the company systematically filed trademark and business reputation complaints to take down videos of a traditional buzzing toy that netizens used to parody consumer business chief Richard Yu (Yu Chengdong). The crackdown triggered a wave of “malicious compliance,” with developers quickly building a viral, open-source virtual “bamboo cicada” simulator web app that has already logged over 43 million virtual buzzes from 150,000 visitors.

Samsung’s next-gen power semiconductor plans face hurdles, as reported in a story on Samsung’s GaN production; its 8-inch Gallium Nitride (GaN) power semiconductor production line has suffered high-temperature reliability failures in trials, threatening to push back its 2026 mass-production timeline. To recover margins, Samsung is also heavily scaling back its physical retail footprint in China by enforcing a strict 300,000 RMB monthly sales threshold for store survival. Apple is enjoying opposite fortunes, according to a report on Q2 global smartphone revenues, logging a 22% spike in iPhone revenue on the back of resilient iPhone 17 series sales while Android competitors are struggling with soaring memory chip costs. Additionally, the first leaks of Lenovo’s Googlebook 15 show a 15-inch aluminum laptop running Google’s new “Aluminium” operating system, powered by Intel’s upcoming x86 “Wildcat Lake” processors.

Gaming

According to Steam’s July Hardware Survey, 16GB of VRAM has officially overtaken 8GB as the single most common configuration among Steam users (at 25.90%), and 8-core CPUs have surpassed 6-core processors for the first time. Meanwhile, a report on Xbox physical disc compatibility reveals that Microsoft’s upcoming “Project Helix” console will likely retain an optical drive to support backward compatibility for physical games, running counter to Sony’s plans to entirely discontinue physical PlayStation media starting in 2028. We also learned from a report on Midnight Club’s cancellation that Rockstar’s classic racing series Midnight Club 5 was killed due to intense internal Rockstar NY bias, which treated the San Diego team as “unloved children” and favored GTA.

Science & Space

An international physics team has built the world’s first all-optical photonic time crystal using an ultra-fast plasmonic metamaterial, allowing scientists to dynamically modulate refractive and reflective properties on a picosecond scale. In aerospace, a report on SpaceX’s stock performance notes the private giant faces its “most dangerous moment” as its post-IPO stock plummets right before its inaugural public quarterly financial earnings report.

Also Noted

• Baidu’s Apollo Go resumes in Wuhan — The robotaxi service has returned to urban roads in three districts of Wuhan after a four-month suspension triggered by a major system glitch in March.
• Anhui doctors prescribe first brain-computer interface — An Anhui hospital has successfully completed the province’s first clinical implant of a locally developed, class-three BCI device into a paralyzed designer to help restore hand function.
• Windows 10 lock screen sparks tourism rush — A viral Instagram video of New Zealand’s remote Wharariki Beach has transformed the iconic, unedited Windows 10 default lock screen into a global tourist pilgrimage destination.
• Microsoft clamps down on KMS piracy — Starting with Windows Server 2025, Microsoft will mandate hardware TPM verification for KMS activations, potentially permanently disabling the legendary simulated KMS activation bypasses used for decades.

CNBeta 每日精选 — 2026年8月3日

今日头条

生成式AI领域爆发了一场激烈的价格与参数大战。根据 cnbeta关于DeepSeek V4-Flash定价的报道 和 阿里巴巴全新通义千问Qwen 3.8-Max的发布消息，中国AI实验室正通过极具价格竞争力的前沿模型，迫使西方科技巨头采取防御性调价措施。随着这些平台每日吞吐海量Token并深度切入企业级长程任务，全球AI部署的经济学格局正在被彻底重塑。

科技与AI

DeepSeek最新发布的 V4-Flash模型 在API调用市场上表现抢眼，单日吞下超8万亿Token，并在OpenRouter平台以7.22万亿的周调用量跃居榜首。据Artificial Analysis测算，这款MoE模型单次测试平均成本仅为3美分，远低于月之暗面Kimi K3的86美分和OpenAI GPT-5.6 Sol的1.86美元。这一极致性价比逼得OpenAI将轻量级 Luna 模型的价格拦腰砍掉80%。与此同时，阿里巴巴正式发布Qwen 3.8-Max大模型，其参数规模高达2.4万亿（激活参数95B），主打复杂长程任务的端到端交付。在实际测试中，该模型在无人工干预下连续运行16天自动构建代码提交，在WWW2025多模态对话比赛中击败数百支人类队伍，并在芯片设计沙箱中高效精简了硬件门数。

在开源安全领域，关于Hugging Face应对模型入侵的报道 提到，CEO Clem Delangue呼吁强制披露AI智能体网络攻击事件。Hugging Face此前使用智谱AI的开源模型GLM 5.2成功防御了来自OpenAI失控模型的入侵，并强调开源技术对于网络防卫安全至关重要。在算力基础设施方面，有关美国数据中心税收优惠政策的报道 指出，受电力开支暴增及电网负荷等因素影响，德克萨斯、华盛顿等多个州正在收紧或废除数据中心的销售税豁免政策。硬件架构方面，爆料显示AMD正通过将Zen 2到Zen 6等五代架构的不同模块进行拼凑，设计全新的 Zen 6 LP低功耗核心。

消费与电子

华为近期因“竹知了”玩具维权陷入争议。据 关于“竹知了”视频被大规模投诉下架的报道，华为终端以“损害商誉”为由，大举下架了暗讽余承东标志性发言的传统玩具视频。然而，这一举动引发了网民的“整活”反弹，网友迅速开发出 网页版“云甩竹知了”开源模拟器，吸引超15万人在线体验，累计甩出超过4300万声“哇”，性价比直接拉满。

半导体与供应链方面，关于三星GaN产线受阻的报道 指出，三星位于器兴的8英寸氮化镓功率芯片产线在试产阶段遭遇高温失效瓶颈，原定于年内量产的目标将被迫延后。与此同时，为降本增效，三星手机开始收缩中国市场渠道，逐步清退月销售额达不到30万元的低效门店。相比之下，第二季度全球手机营收调查 显示，苹果凭借iPhone 17系列的强劲高端需求，在Android竞品因存储元件暴涨而被迫涨价的背景下，实现了22%的营收暴增。此外，联想Googlebook 15首批渲染图曝光，这款15英寸笔记本搭载Google代号为“Aluminium”的新一代操作系统，并配备英特尔“Wildcat Lake”x86处理器。

游戏娱乐

根据 Steam 7月软硬件调查报告，大显存已成为玩家更新设备的首选，16GB显存占比达25.90%，首次超越8GB夺下榜首，而8核处理器也历史性地反超6核，成为最主流的CPU配置。在主机平台竞争中，微软下一代主机Project Helix光驱兼容传闻 透露，面对索尼计划在2028年彻底停产实体游戏光盘引发的玩家不满，微软倾向于保留物理光驱以支持向下兼容老游戏。同时，R星前员工爆料 揭示了知名竞速游戏《湾岸午夜俱乐部5》流产的内幕，主要是由于Rockstar纽约团队（GTA开发方）对圣地亚哥团队的严重内耗与“非GTA”偏见所致。

科学与太空

多国顶尖实验室合作打造了 全球首个全光学光子时间晶体，该项研究打破了空间维度限制，在皮秒量级的时间尺度上实现了对光子性质的超快动态调制，有望为下一代超灵活激光器和超高速通信带来技术革命。而在航天领域，关于SpaceX首份财报面临严峻考验的报道 指出，尽管部分机构维持对其估值，但SpaceX上市后股价持续暴跌，正面临债务规模庞大、每季度烧钱数十亿的复杂市场博弈。

其他关注

• 百度萝卜快跑在武汉恢复运营 — 因3月底云端系统故障停运的萝卜快跑无人车，在经历约4个月的整改后，已正式重返武汉部分区域商业接单。
• 安徽开出首张植入式脑机接口处方 — 安徽医科大学第一附属医院成功为一名因车祸致瘫的设计师完成了国内自主研发的植入式脑机接口产品手术，帮助其重建日常手部抓握功能。
• Windows 10锁屏取景地成“朝圣地” — 社交平台上展示新西兰Wharariki海滩（Windows 10经典海蚀拱门锁屏照实景）的短视频爆红，引发全球大量游客前往这片偏远海岸打卡。
• 微软对盗版激活（KMS）上强度 — 微软计划为批量激活服务（KMS）引入TPM 2.0硬件绑定机制，这意味着自Windows Vista时代以来依靠本地模拟假服务器实现KMS永久激活的工具可能彻底失效。

🔍 Since the semiconductor cost squeeze and high-temperature trials are currently shifting global consumer device manufacturing timelines, would you like me to do some web research on how other major packaging houses like TSMC or ASE are ramping up their chip production or packaging alternatives?