Title: 又一巨头，下场造芯 | BestBlogs.dev
URL Source: https://www.bestblogs.dev/article/a0fd39a399?amp%3Butm_medium=feed&%3Butm_campaign=resources&%3Bentry=rss_article_item
Published Time: 2026-08-06 08:59:00
Markdown Content: Skip to main contentAudio 2 
Search Ctrl+K
Change language Switch ThemeSign In
[](https://www.bestblogs.dev/en/explore/brief "Daily Brief")[](https://www.bestblogs.dev/en/explore/newsletter "Weekly Picks")[](https://www.bestblogs.dev/en/explore/topics "Topics")
[](https://www.bestblogs.dev/en/settings "Settings")[](https://www.bestblogs.dev/en/docs "Help Center")
Narrow Mode
82
又一巨头，下场造芯
Anthropic is building a custom chip team and in talks with Samsung over a 2nm process. This article unpacks the motivations, industry trends, and potential impact on Nvidia. 半 半导体行业观察Follow·
Today·2221 words (about 9 min)
AI Summary & Key Points
Summary
This article reports that Anthropic has publicly acknowledged for the first time its in-house chip development plans, is assembling a custom silicon team, and is in discussions with Samsung on 2nm process technology and advanced packaging. The piece traces the timeline: from Reuters' initial report in April revealing the intent, to The Information's July report on talks with Samsung, to the recent hire of Clive Chan, a former member of OpenAI's chip team. Samsung, SK Hynix, and Micron all participated as strategic investors in Anthropic's $65 billion funding round.
The article analyzes the core drivers behind the move: Nvidia GPU supply is already booked out through 2027, procurement costs are climbing, and demand for Claude continues to grow—making pure external purchasing insufficient to balance expansion speed with cost control. It also notes that Anthropic's compute strategy emphasizes diversification, with partnerships across AWS, Google, Nvidia, and AMD. On the training side, it uses Nvidia GPUs, Google TPUs, and Amazon Trainium simultaneously, and plans to introduce a next-generation TPU co-developed by Alphabet and Broadcom in 2027.
The article places Anthropic within the broader industry trend: OpenAI unveiled its inference chip Jalapeño, developed with Broadcom, in June; Google's TPU has iterated to its seventh generation; Amazon has Inferentia and Trainium; and Meta continues to push forward with MTIA. The conclusion is that a hybrid compute landscape—custom chips coexisting with Nvidia's general-purpose GPUs—will become the norm. Nvidia's dominance in the training market will remain intact in the short term, but large-scale inference will face increasing pressure as workloads shift away.
Main Points
* 1. Anthropic has publicly acknowledged its in-house chip plans for the first time and is in talks with Samsung on a 2nm process.
Job postings show Anthropic is assembling a custom chip team and in discussions with Samsung on 2nm process technology and advanced packaging. Samsung is also a strategic investor in its $65 billion funding round.
* 2. The core driver behind building custom chips is the supply-demand imbalance and cost pressure around Nvidia GPUs.
Nvidia GPU supply is already booked out through 2027, with rising procurement difficulty and costs. Pure external purchasing can no longer support Anthropic's expansion speed and cost-control needs.
* 3. Anthropic's compute strategy prioritizes diversified supply, with in-house chips serving as a complement rather than a replacement.
Anthropic has signed agreements with AWS, Google, Nvidia, and AMD, using three types of chips simultaneously for training and planning to introduce next-generation TPUs in 2027. In-house chips are meant to lower marginal costs and improve fit with Claude, not to replace external suppliers.
* 4. Custom chips have become a common choice among leading AI companies, with inference as the preferred initial use case.
OpenAI, Google, Amazon, and Meta have all built out in-house chip programs. The industry pattern is that inference workloads land first, since their patterns are more fixed and predictable, while training chips carry higher design barriers and risk.
* 5. Nvidia is irreplaceable in the short term, but the large-scale inference market will face workload diversion.
GPUs, the CUDA ecosystem, and cluster design capabilities remain the top choice for frontier training, but predictable, high-volume inference traffic is gradually migrating to custom chips, which will erode Nvidia's share of the inference market.
Sign in to highlight text and take notes as you read. Sign in now
半导体行业观察 2026-08-06 08:59 安徽
半导体行业观察：Anthropic正在组建一支芯片设计团队，为其人工智能系统开发定制芯片。Business Insider最先报道了这一消息，Anthropic随后向TechCrunch证实了此事。
公众号记得加星标⭐️，第一时间看推送不会错过。
Anthropic正在组建一支芯片设计团队，为其人工智能系统开发定制芯片。Business Insider最先报道了这一消息，Anthropic随后向TechCrunch证实了此事。招聘信息显示，公司正在为"定制芯片团队"（custom silicon team）招募具有芯片设计经验的工程师，这也是Anthropic首次公开承认其自研硬件的野心。
Claude的开发商表示，计划与硬件团队协同设计芯片与模型，以提升技术运行的速度和效率。这并非Anthropic首次被曝出自研芯片的计划。今年4月，路透社曾报道称，面对芯片短缺的压力，Anthropic正在考虑自行生产AI芯片，彼时这一想法尚处于设想阶段。到7月初，The Information的报道显示事态已推进：Anthropic与三星就芯片制造展开了接触，双方讨论的方向包括三星的2纳米制程工艺和先进封装技术。有报道还提到，Anthropic近期招揽了Clive Chan——此人曾是OpenAI自研芯片团队的早期成员之一。三星本身也是Anthropic近期650亿美元融资轮的战略投资方之一，同轮投资方还包括SK海力士和美光。
不过截至目前，芯片的具体用途、如何融入服务器架构、性能目标等关键细节仍未确定，报道称相关讨论仍处于早期阶段。对于三星合作的传闻，Anthropic方面没有正面确认，只是向媒体表示，包含谷歌、亚马逊和英伟达芯片在内的多元化硬件体系，仍将是公司算力策略的核心。
驱动Anthropic下场造芯的核心原因是供需矛盾。Claude的市场需求持续攀升，而英伟达GPU产能被普遍认为已排产至2027年，采购难度和成本双双上升。目前Anthropic已与AWS、谷歌、英伟达和AMD签署协议以获取AI计算硬件，训练层面则同时使用英伟达GPU、谷歌TPU和亚马逊Trainium三类芯片，且已宣布从2027年起在训练基础设施中引入由Alphabet与博通联合开发的新一代TPU。这一系列布局的共同逻辑是分散供应来源，避免受制于单一供应商，而非彻底转向某一方案。仅靠外部采购，显然难以支撑其扩张速度和成本控制的双重需求，自研芯片被视为进一步压低边际成本、提升硬件与Claude模型适配度的手段。
Anthropic并非孤例。今年6月，OpenAI发布了与博通合作打造的Jalapeño芯片，专为推理工作负载设计，是OpenAI在多年完全依赖英伟达之后首次涉足定制芯片，官方称其在早期测试中相比主流通用GPU具备更优的能效表现。谷歌DeepMind长期依赖Alphabet自研的TPU芯片驱动其AI模型，目前已迭代至第七代；亚马逊2019年推出Inferentia、2021年推出Trainium；Meta则一直在开发用于AI工作负载的MTIA加速器，主要用于推荐系统模型，并逐步扩展到生成式AI推理场景。行业分析普遍认为，这一模式已形成规律：当AI公司的营收和算力规模达到一定门槛后，自研芯片的经济账开始成立——即便前期芯片设计投入高达数亿美元，几年内也能通过效率提升收回成本。
对英伟达而言，短期内被完全替代的可能性不大，其GPU、CUDA生态、网络方案、软件工具链和集群设计能力仍难以复制，尤其在前沿模型训练和快速迭代的研究场景中依然是首选。但真正的压力来自于可预测、高体量的推理流量逐步被迁移到定制芯片上——如果Anthropic、OpenAI、谷歌、亚马逊、Meta等相继把稳定的推理负载转向自研硬件，英伟达在前沿训练市场的主导地位或许不受影响，但在大规模推理服务这一环节将面临更多分流压力。多份行业预测也指向同一方向：定制芯片与英伟达通用GPU并存的混合算力市场，将成为未来一段时间的主流格局，而非非此即彼的替代关系。
从现有信息看，Anthropic的芯片项目仍处于探索阶段，尚未确定芯片是面向训练、推理，还是两者兼顾，也未明确其将如何嵌入服务器架构、目标性能指标等关键参数。相比之下，OpenAI的Jalapeño定位更为清晰——一款专为推理工作负载设计、已进入测试阶段并计划于今年底前部署的产品。若Anthropic与三星的合作最终落地，讨论中提及的2纳米制程和先进封装技术，指向的更可能是一款兼顾能效与算力密度的推理芯片，这也与目前行业的普遍选择相符：多数公司的自研芯片首先落地在负载模式相对固定、可预测性更强的推理环节，训练芯片的自研门槛和风险仍然更高。
（来源：内容来自半导体行业观察综合）
*免责声明：本文由作者原创。文章内容系作者个人观点，半导体行业观察转载仅为了传达一种不同的观点，不代表半导体行业观察对该观点赞同或支持，如果有任何异议，欢迎联系半导体行业观察。
_END_ 今天是《半导体行业观察》为您分享的第 4490 内容，欢迎关注。 推荐阅读
加星标⭐️第一时间看推送
求点赞
求分享
求推荐
Key Quotes
> Once an AI company's revenue and compute scale cross a certain threshold, the economics of building custom chips start to work—even if upfront chip design investment runs into hundreds of millions of dollars, the efficiency gains can recoup the cost within a few years.
> A hybrid compute market where custom chips coexist with Nvidia's general-purpose GPUs will become the dominant pattern for the foreseeable future, rather than an either-or replacement.
Tags
AI Hardware & Chips
AI Commercialization
AI Infrastructure
Compute
Nvidia
Related Articles
* Even with self-evolution, current AI cannot replace human judgment | Yuandong Tian, arguing that human judgment remains irreplaceable, and that a sense of direction (taste) and deep domain expertise are the ultimate moat.") * Rethinking Development Infrastructure: When Agents Become First-Class Citizens * Tech Enthusiast Weekly (Issue 399): Notes on Visiting China's AI Giants * Advanced Packaging: Entering a New Battlefield, exploring its advantages in addressing cost and size bottlenecks, TSMC and domestic vendors' layout, and the process challenges faced.") * Nvidia Buys 5 Million Optical Modules a Year, Then Dismantles Them Itself * AI and Careers Over the Last Three Years: Turbulence, Patterns, and Professional Portraits * The next bottleneck for AI may not be GPU and storage, but semiconductor equipment. * Dowson Tong and Yao Shunyu in Conversation: Tencent AI's Second Half * Broadcom Told the Truth, But No One Noticed * The Missing Link in Co-Packaged Optics, emphasizing that merely improving photonic device performance cannot achieve scale; instead, package PDK must enable multi-domain system-level integration—covering optoelectrical, thermal, mechanical, and manufacturing aspects—and ensure manufacturability.")
Make your daily reading actually fit you.A daily brief built from the sources you follow. Get started free HomeDiscoverSettings