AI日报 | 2026年10月21日下午版：Anthropic招股书曝光半年亏807亿美元，DeepSeek开源推理框架DeepInfer

发布时间：2026-10-21 | 分类：AI日报 | 阅读：约 8 分钟

今日下午核心要闻
Anthropic IPO 招股书曝光：上半年净亏损 807 亿美元，营收 163 亿美元同比增长近 14 倍
DeepSeek 开源内部推理框架 DeepInfer，支撑 V4.1 Flash 单集群日服务 155 万亿 Token
Spotify 移除 5000 万首低质 AI 生成音乐，宣布优先展示人类原创作品
谷歌开源 TorchTPU，打破 TPU 生态封闭格局，支持 PyTorch 原生调用
智元人形机器人广交会摆摊 18 天完成 827 单，具身智能商业化加速
联想等企业发起 AI OPC 计划，推动人工智能产业协同与生态网络建设
IDC 预计 2026 年中国 AI 座舱市场规模达 16.6 亿元，车载 AI 渗透提速

一、Anthropic IPO 招股书曝光：半年亏损 807 亿美元，算力账单冲向 7500 亿

AnthropicIPO财务

据《华尔街日报》报道，AI 公司 Anthropic 的 IPO 招股书草案曝光。文件显示，公司 2026 上半财年净亏损高达 807 亿美元，较上年同期 130 亿美元大幅扩大。不过，其中约 720 亿美元来自可转换票据等融资工具的会计公允价值变动，并非经营性现金流出。

营收方面，Anthropic 上半年营收从去年同期 11 亿美元飙升至 163 亿美元，Q3 营收进一步环比增长 45% 至 168 亿美元，并在调整后口径下实现季度盈利。更令市场关注的是其算力账单：上半年算力与基础设施支出接近 90 亿美元，已超过 2025 全年 73 亿美元的水平。公司已签署超 6000 亿美元的算力采购承诺，若当前谈判全部达成，潜在支出将升至 7500 亿美元。

编辑点评：Anthropic 的招股书揭示了 AI 大模型竞赛的真实成本——营收增长惊人但算力开支更为惊人。约 1.5 万亿美元的最新估值较此前传闻的 2 万亿有所下调，显示投资者正更审慎评估 AI 公司的投入产出比。

二、DeepSeek 开源推理框架 DeepInfer，GitHub 一天涨星 1.3 万

DeepSeek开源推理框架

10月20日，DeepSeek 宣布将内部推理框架 DeepInfer 全面开源，采用 Apache 2.0 协议。该框架是支撑 DeepSeek V4.1 Flash 线上百亿级调用的核心基础设施，实现了单集群日服务 155 万亿 Token 的高吞吐、低延迟推理能力，同时兼容英伟达 GPU 和华为昇腾芯片。

accompanying 开源的还有生产级分词组件 deepseek-recipe。消息发布后，DeepInfer 在 GitHub 一天内获得 1.3 万颗星，登顶 Hacker News 热榜。DeepSeek 表示，未来新模型的推理代码及优化也将在此框架持续更新。

行业意义：DeepSeek 将"印钞机"级别的推理基础设施开源，大幅降低了大模型推理部署的门槛。这也是中国 AI 公司"真开源"的又一标杆——不仅开源模型权重，更开源生产级服务框架。

三、Spotify 移除 5000 万首低质 AI 音乐，优先展示人类作品

SpotifyAI音乐内容治理

流媒体音乐巨头 Spotify 宣布从平台移除超过 5000 万首低质 AI 生成音乐曲目，并调整推荐算法，优先展示人类创作的音乐作品。此举是对 AI 生成内容泛滥导致平台质量下降的直接回应。

Spotify CEO 表示，虽然公司支持 AI 作为创作辅助工具，但大量低质 AI 音乐严重影响了用户体验和创作者生态。新政策将对 AI 生成内容实施更严格的标识和质量审核机制。

四、谷歌开源 TorchTPU，打破 TPU 生态封闭格局

Google开源TPU

谷歌正式开源 TorchTPU 项目，使 PyTorch 开发者能够原生调用 TPU 进行模型训练和推理，无需再依赖专有 XLA 编译流程。这一举措被视为谷歌打破 TPU 生态封闭格局、与英伟达 CUDA 生态争夺开发者的关键一步。

TorchTPU 支持 PyTorch 2.x 的全部核心 API，开发者仅需修改少量代码即可将训练任务从 GPU 迁移到 TPU。业内分析认为，这将显著降低 TPU 的使用门槛，加速谷歌云在 AI 训练市场的份额增长。

五、智元人形机器人广交会摆摊 18 天完成 827 单

具身智能人形机器人

在第 138 届广交会上，智元机器人的人形机器人"远征 A2"在展位连续工作 18 天，完成了 827 单饮品制作和服务任务，轻松应对大客流。这是人形机器人在大型商业场景中的标志性落地案例。

智元机器人表示，远征 A2 已实现从感知、决策到执行的全链路自主运行，能够在复杂人流环境中稳定工作。广交会期间的实测数据验证了具身智能在零售、餐饮等服务场景的商业可行性。

六、联想等发起 AI OPC 计划，推动产业协同

产业合作AI生态

联想集团联合多家企业共同发起 AI OPC（Open Partnership for Computing）计划，旨在推动人工智能产业协同与生态网络建设。该计划聚焦 AI 基础设施标准化、应用互操作性和供应链协同三大方向。

参与企业涵盖芯片制造商、云服务商、终端设备厂商和行业解决方案提供商，计划通过开放接口和统一标准，降低 AI 部署成本，加速 AI 技术在制造、医疗、教育等垂直行业的规模化落地。

七、IDC：2026 年中国 AI 座舱市场规模将达 16.6 亿元

IDC智能座舱车载AI

市场研究机构 IDC 发布最新报告，预计 2026 年中国 AI 座舱市场规模将达到 16.6 亿元。报告指出，随着大模型端侧部署技术成熟，车载 AI 正从简单的语音助手向多模态交互、主动服务和场景感知方向快速演进。

IDC 分析师表示，FP4（4位浮点）量化技术正成为端侧 AI 的关键选择，能够在有限算力下实现大模型的高效推理，这将加速 AI 座舱从高端车型向中端市场的渗透。

今日 AI 数据一览

指标	数据	变化
Anthropic 上半年净亏损	807 亿美元	同比扩大 520%
Anthropic 上半年营收	163 亿美元	同比增长 14 倍
Anthropic 算力支出承诺	最高 7500 亿美元	未来 7-10 年
DeepInfer 单日服务量	155 万亿 Token	单集群
Spotify 移除 AI 音乐	5000 万首	低质内容清理
中国 AI 座舱市场规模（2026E）	16.6 亿元	IDC 预测
智元机器人广交会服务单量	827 单 / 18 天	人形机器人实测

信息来源

AI工具猫编辑部：今日下午版聚焦 AI 产业的"成本与开源"双重主题。Anthropic 招股书揭示了大模型竞赛的惊人成本——半年亏损 807 亿、算力承诺 7500 亿；而 DeepSeek 和谷歌分别以 DeepInfer 和 TorchTPU 的开源，推动推理基础设施的民主化。一边是天价投入，一边是开放共享，AI 产业正走向更成熟的发展阶段。

AI Daily | October 21, 2026 Afternoon Edition: Anthropic Prospectus Reveals Half-Year Loss of $80.7 Billion, DeepSeek Open-Sources Inference Framework DeepInfer

Published: 2026-10-21 | Category: AI Daily | Read: about 8 min

Top Stories This Afternoon
Anthropic's IPO prospectus revealed: an $80.7 billion net loss in the first half, with revenue of $16.3 billion, up nearly 14x year on year
DeepSeek open-sources its internal inference framework DeepInfer, supporting V4.1 Flash at 155 trillion tokens per day on a single cluster
Spotify removes 50 million low-quality AI-generated tracks and announces priority for human-created works
Google open-sources TorchTPU, breaking the closed TPU ecosystem and supporting native PyTorch calls
AgiBot humanoid robots sell for 18 days at the Canton Fair, completing 827 orders as embodied intelligence commercialization accelerates
Lenovo and other companies launch the AI OPC initiative to advance AI industry collaboration and ecosystem network building
IDC forecasts China's AI cockpit market will reach 1.66 billion yuan in 2026, as in-car AI penetration accelerates

1. Anthropic IPO Prospectus Revealed: $80.7 Billion Half-Year Loss, Compute Bill Heading Toward $750 Billion

AnthropicIPOFinancials

According to The Wall Street Journal, the draft IPO prospectus of AI company Anthropic has been revealed. The document shows the company posted a net loss of $80.7 billion in the first half of fiscal 2026, a sharp widening from $13 billion a year earlier. However, about $72 billion of that came from accounting fair-value changes on financing instruments such as convertible notes, not operating cash outflows.

On revenue, Anthropic's first-half revenue soared from $1.1 billion a year earlier to $16.3 billion, with Q3 revenue up another 45% quarter on quarter to $16.8 billion, and the company achieved a quarterly profit on an adjusted basis. More striking is its compute bill: first-half compute and infrastructure spending approached $9 billion, already exceeding the $7.3 billion for all of 2025. The company has signed compute purchase commitments of more than $600 billion, and if all current negotiations are concluded, potential spending would rise to $750 billion.

Editor's note: Anthropic's prospectus reveals the real cost of the large-model race — astonishing revenue growth, but even more astonishing compute spending. The latest valuation of about $1.5 trillion is down from the previously rumored $2 trillion, showing investors are assessing AI companies' input-output ratios more cautiously.

2. DeepSeek Open-Sources Inference Framework DeepInfer, 13,000 GitHub Stars in a Day

DeepSeekOpen SourceInference Framework

On October 20, DeepSeek announced the full open-sourcing of its internal inference framework DeepInfer under the Apache 2.0 license. The framework is the core infrastructure supporting DeepSeek V4.1 Flash's hundreds of billions of online calls, delivering high-throughput, low-latency inference at 155 trillion tokens per day on a single cluster, while compatible with NVIDIA GPUs and Huawei Ascend chips.

Released alongside was the production-grade tokenizer component deepseek-recipe. After the announcement, DeepInfer gained 13,000 stars on GitHub in a single day and topped the Hacker News front page. DeepSeek said the inference code and optimizations for future new models will also be continuously updated in this framework.

Industry significance: DeepSeek has open-sourced "money-printing-machine"-level inference infrastructure, greatly lowering the barrier to deploying large-model inference. It is another benchmark for Chinese AI companies' "true open source" — open-sourcing not only model weights but also production-grade serving frameworks.

3. Spotify Removes 50 Million Low-Quality AI Tracks, Prioritizes Human Works

SpotifyAI MusicContent Governance

Streaming music giant Spotify announced it has removed more than 50 million low-quality AI-generated tracks from the platform and adjusted its recommendation algorithm to prioritize music created by humans. The move is a direct response to declining platform quality caused by the flood of AI-generated content.

Spotify's CEO said that while the company supports AI as a creative assistant tool, the large volume of low-quality AI music has seriously harmed user experience and the creator ecosystem. The new policy will impose stricter labeling and quality review mechanisms on AI-generated content.

4. Google Open-Sources TorchTPU, Breaking the Closed TPU Ecosystem

GoogleOpen SourceTPU

Google has officially open-sourced the TorchTPU project, enabling PyTorch developers to natively call TPUs for model training and inference without relying on the proprietary XLA compilation flow. The move is seen as a key step for Google to break the closed TPU ecosystem and compete with NVIDIA's CUDA ecosystem for developers.

TorchTPU supports all core APIs of PyTorch 2.x, and developers only need to modify a small amount of code to migrate training tasks from GPU to TPU. Industry analysts believe this will significantly lower the barrier to using TPUs and accelerate the growth of Google Cloud's share in the AI training market.

5. AgiBot Humanoid Robot Works the Canton Fair for 18 Days, Completing 827 Orders

Embodied IntelligenceHumanoid Robots

At the 138th Canton Fair, the humanoid robot "Expedition A2" from AgiBot worked continuously for 18 days at its booth, completing 827 drink-making and service tasks and easily handling heavy foot traffic. It is a landmark deployment case of humanoid robots in a large commercial scenario.

AgiBot said Expedition A2 has achieved fully autonomous operation across the entire chain from perception and decision-making to execution, and can work stably in complex crowd environments. Real-world data from the Canton Fair verified the commercial feasibility of embodied intelligence in service scenarios such as retail and dining.

6. Lenovo and Others Launch AI OPC Initiative to Advance Industry Collaboration

Industry CollaborationAI Ecosystem

Lenovo Group, together with several companies, jointly launched the AI OPC (Open Partnership for Computing) initiative to advance AI industry collaboration and ecosystem network building. The initiative focuses on three directions: AI infrastructure standardization, application interoperability and supply chain collaboration.

Participating companies span chipmakers, cloud service providers, terminal device makers and industry solution providers. The initiative aims to lower AI deployment costs through open interfaces and unified standards, accelerating large-scale AI adoption in vertical industries such as manufacturing, healthcare and education.

7. IDC: China's AI Cockpit Market to Reach 1.66 Billion Yuan in 2026

IDCSmart CockpitIn-Car AI

Market research firm IDC released a new report forecasting that China's AI cockpit market will reach 1.66 billion yuan in 2026. The report notes that as on-device large-model deployment technology matures, in-car AI is rapidly evolving from a simple voice assistant toward multimodal interaction, proactive service and scenario awareness.

IDC analysts said FP4 (4-bit floating point) quantization is becoming a key choice for on-device AI, enabling efficient large-model inference under limited compute, which will accelerate AI cockpits' penetration from premium models to the mid-range market.

Today's AI Data at a Glance

Metric	Data	Change
Anthropic first-half net loss	$80.7 billion	Up 520% YoY
Anthropic first-half revenue	$16.3 billion	Up 14x YoY
Anthropic compute spending commitments	Up to $750 billion	Next 7-10 years
DeepInfer daily serving volume	155 trillion tokens	Single cluster
Spotify AI music removed	50 million tracks	Low-quality content cleanup
China AI cockpit market size (2026E)	1.66 billion yuan	IDC forecast
AgiBot Canton Fair service orders	827 orders / 18 days	Humanoid robot real-world test

Sources

AICat Editorial: Today's afternoon edition focuses on the dual themes of "cost and open source" in the AI industry. Anthropic's prospectus reveals the astonishing cost of the large-model race — an $80.7 billion half-year loss and $750 billion in compute commitments — while DeepSeek and Google are democratizing inference infrastructure with the open-source DeepInfer and TorchTPU. On one side sky-high investment, on the other open sharing — the AI industry is moving toward a more mature stage of development.