Core Viewpoint - Nvidia has faced significant stock volatility in 2025, with a year-to-date decline of 15%, primarily due to concerns over potential demand reduction for its data center chips amid tariff implications [1][9] Group 1: Tariff Impact and Customer Spending - Although semiconductors are exempt from aggressive tariffs, Nvidia's customers may still experience increased costs, potentially leading to reduced capital expenditures [2] - Major customers like Meta, Microsoft, Alphabet, and Amazon have provided positive updates on their AI spending plans for 2025, indicating continued demand for Nvidia's chips [2][12] - Meta raised its 2025 capex forecast to $64 billion to $72 billion, Microsoft plans to spend around $80 billion, Alphabet maintains a $75 billion forecast, and Amazon is set to spend approximately $105 billion [12] Group 2: Nvidia's Technological Advancements - Nvidia's H100 GPU was the leading AI data center chip in 2023 and most of 2024, but has been succeeded by the more advanced Blackwell and Blackwell Ultra architectures, with the latter offering up to 50 times faster AI inference in specific configurations [4][6] - The upcoming Rubin GPUs, expected in 2026, are projected to deliver 3.3 times more compute performance, further enhancing Nvidia's position in the AI market [7] Group 3: Market Position and Future Growth - Nvidia generated $115.2 billion in data center revenue for fiscal 2025, marking a 142% increase from the previous year, with predictions of data center spending exceeding $1 trillion annually by 2028 [14] - Demand for Nvidia's chips currently exceeds supply, making it difficult for companies to cancel orders without risking a competitive disadvantage in AI [16] - Nvidia's stock is viewed as a buying opportunity, trading at a P/E ratio of 39, significantly lower than its 10-year average above 50 [11]

赶在五一假期前夕，DeepSeek给我们送出一份惊喜大礼。 延续一贯的开源节奏，DeepSeek在Hugging Face正式发布DeepSeek-Prover-V2，并同步上线模型卡及示例代码。此次共推出两个版本： *核心贡献者†在DeepSeek-AI实习期间完成的工作 据官方论文披露，DeepSeek-Prover-V2的训练核心是"递归+强化学习"的组合：即先由DeepSeek-V3拆解复杂定理，生成一系列子目标和推理思路；再通过 GRPO算法，从多种候选方案中自动学习如何选出最优解。 模型特别引入了两种互补的"解题风格"： DeepSeek-Prover-V2-7B：基于上一代V1.5模型，支持最长32K上下文输入； DeepSeek-Prover-V2-671B：在DeepSeek-V3-Base基础上训练，推理性能最强。 训练过程分为两阶段，在第一阶段，研究人员主要训练快速模式，采用"专家迭代"方法：模型先尝试解决难题，成功的答案再作为新数据反哺模型，不断 打磨自己的能力。 待快速模式趋于稳定后，研究人员进入第二阶段，开始训练更复杂的逻辑推理能力。他们将DeepSeek-V3的数学知识迁移到新模 ...

Alibaba released the next generation of its open-sourced large language models, Qwen3, on Tuesday — and experts are calling it yet another breakthrough in China's booming open-source artificial intelligence space. In a blog post, the Chinese tech giant said Qwen3 promises improvements in reasoning, instruction following, tool usage and multilingual tasks, rivaling other top-tier models such as DeepSeek's R1 in several industry benchmarks. Qwen3 is Alibaba's debut into so-called "hybrid reasoning models," wh ...

技术突破：从理论根基到系统能力的全栈创新 全球人工智能领域再传DataCanvas强音!九章云极DataCanvas公司科研团队的两项原创成果《A Solvable Attention for Neural Scaling Laws》与《DyCAST: Learning Dynamic Causal Structure from Time Series》被人工智能三大顶级会议之一 ICLR(International Conference on Learning Representations)正式收录。这两项成果分别从神经网络基础理解与动态因果 系统建模两大方向取得进展，标志着九章云极DataCanvas团队在AI底层技术创新与国际学术影响力上实现跨越式提 升。 顶会严选：印证DataCanvas AI科研实力 ICLR与NeurIPS、ICML是人工智能领域公认的全球三大顶级学术会议之一，由深度学习先驱Yoshua Bengio、Yann LeCun等人于2013年发起成立。ICLR凭借其对深度学习核心问题的持续深耕、严苛的学术标准与开放协作的社区文 化，已成为全球AI学者发布里程碑成果的首选平台，在谷歌 ...

SCoT团队 投稿 量子位 | 公众号 QbitAI 不怕推理模型简单问题过度思考了，能 动态调整CoT的新推理范式SCoT来了！ SCoT，即自 结构化推理链 （Self-structured Chain of Thought ） 。 它通过 将推理过程分解为最小语义原子步骤 ，能动态生成适配不同复杂度问题的CoT结构，解决了现有方法在推理多样性和效率上的不足。 另外，为了激发推理能力，研究人员还提出了 AtomThink ，这是一个包含数据构造、训练、推理和评估的全过程框架， 用来提升多模态大 模型在复杂推理任务上的表现 。 实验中，SCoT使模型能根据问题复杂度自动调整推理链长度，复杂问题的推理步骤更长。 在多个数据集上，AtomThink框架显著提升了基线模型的准确率，数据利用效率和推理效率也表现出显著优势。 并且，原子能力评估揭示了多模态模型在不同推理能力上的分布特征，为理解多模态推理模式提供了新视角。 这项研究由来自中山大学、香港科技大学、上海交通大学、香港大学、华为诺亚方舟实验室的研究人员联合提出，以下是更多细节。 SCoT、AtomThink长啥样？ 当前，结构化和非结构化CoT面临一定的挑战 ...

作者：Cage、Yongxin、Siqi 编辑：Siqi DeepSeek R1 催化 了 reasoning model 的竞争：在过去的一个月里，头部 AI labs 已经发布了三个 SOTA reasoning models：OpenAI 的 o3-mini 和deep research， xAI 的 Grok 3 和 Anthropic 的 Claude 3.7 Sonnet。 随着头部 Al labs 先后释出自己的 reasoning model，新范式的第一轮竞赛暂时告一段落。 各家 reasoning model 各有长板，但都没有拉开大的领先优势：OpenAI 和 xAI 有着最强的 base model 和 竞赛解题能力，Anthropic 更关注真实世界的工程问题，Claude 3.7 Sonnet 的混合推理模型可能会成为 之后各家发布新模型的标准操作。 在这一波新模型密集发布后的间隙，我们对已有的 reasoning models 发布进行了总结梳理，除了平 行比较各些模型的实际能力和长板外，更重要的目标是识别出本轮发布中的关键信号。 整体上，我们还处于 RL Scaling 的早期 ...

Financial Data and Key Metrics Changes - Q4 revenue reached $39.3 billion, up 12% sequentially and 78% year on year, exceeding the outlook of $37.5 billion [7][8] - Fiscal 2025 revenue totaled $130.5 billion, an increase of 114% compared to the previous year [8] - GAAP gross margins were 73%, with non-GAAP gross margins at 73.5%, down sequentially as expected due to the initial deliveries of the Blackwell architecture [37] Business Line Data and Key Metrics Changes - Data center revenue for fiscal 2025 was $115.2 billion, more than doubling from the prior year, with Q4 data center revenue at a record $35.6 billion, up 16% sequentially and 93% year on year [8][9] - Consumer Internet revenue grew 3x year on year, driven by generative AI and deep learning use cases [19] - Automotive revenue reached a record $570 million, up 27% sequentially and 103% year on year, with full-year revenue increasing by 55% [34] Market Data and Key Metrics Changes - Sequential growth in data center revenue was strongest in the US, driven by the initial ramp of Blackwell [26] - Data center sales in China remained well below previous levels due to export controls, with expectations to maintain current percentages [27] Company Strategy and Development Direction - The company is focused on expediting the manufacturing of Blackwell systems to meet high customer demand, with expectations for gross margins to improve to the mid-seventies later in the year [39][65] - Blackwell architecture is designed to support the entire AI market, from pretraining to inference, ensuring adaptability in rapidly evolving markets [16][36] Management's Comments on Operating Environment and Future Outlook - Management expressed confidence in sustained strong demand for AI computing, driven by the transition to machine learning and AI-based software [67][70] - The company anticipates significant growth in enterprise AI applications, particularly in industrial sectors, which are expected to become a larger part of the consumption mix [110][116] Other Important Information - The company returned $8.1 billion to shareholders in Q4 through share repurchases and cash dividends [39] - Upcoming events include participation in the TD Cowen Healthcare Conference and the Morgan Stanley Technology, Media, and Telecom Conference [43] Q&A Session Summary Question: Future of inference-dedicated clusters - Management discussed the increasing blurring between training and inference, highlighting the need for architectures that can handle both efficiently [46][54] Question: Status of Blackwell ramp and NVLink 72 - Management confirmed successful ramping of Blackwell systems and expressed enthusiasm for the NVLink 72 platform, noting significant demand [57][60] Question: Confidence in sustaining strong demand - Management provided insights into capital investments in data centers and the ongoing vibrancy of AI start-ups, indicating a positive outlook for demand [67][70] Question: Dynamics of Blackwell Ultra launch - Management confirmed that Blackwell Ultra is on track for a second-half launch, with a smooth transition planned from the current generation [75][78] Question: Balance between custom ASICs and merchant GPUs - Management emphasized the general-purpose nature of their architecture compared to ASICs, highlighting the advantages in performance and software ecosystem [82][84] Question: Geographic demand dynamics - Management noted that while US demand surged, China remains a significant market, albeit at reduced levels due to export controls [94][96] Question: Growth of enterprise consumption - Management indicated that enterprise consumption is expected to grow significantly, driven by the need for AI in various industrial applications [110][116]