Anthropic will design its own hardware to power Claude

Anthropic 将设计自研硬件以驱动 Claude

Anthropic is hiring a “custom silicon team” to design chips on which to run its models, the company has revealed. Yesterday, Business Insider noticed a job listing for a senior engineer with experience shipping semiconductor designs. (You can see listings for a silicon engineer and a technical program manager, silicon on Anthropic’s job board right now.)

Anthropic 公司近日透露，他们正在组建一支“定制芯片团队”，旨在设计用于运行其模型的芯片。昨天，Business Insider 注意到该公司发布了一则招聘启事，寻找具有半导体设计交付经验的高级工程师。（目前在 Anthropic 的招聘页面上，可以看到芯片工程师和芯片技术项目经理的职位空缺。）

A spokesperson for Anthropic then confirmed the plans to both Business Insider and TechCrunch. The spokesperson clarified that Anthropic will still take a “multi-chip approach,” with plans to use hardware from other companies alongside its own designs as it continues to scale up. This is confirmation of a rumor that has been circulating for a little bit; The Information previously reported that Anthropic was considering working with Samsung as a hardware manufacturing partner.

随后，Anthropic 的发言人向 Business Insider 和 TechCrunch 证实了这一计划。该发言人澄清说，Anthropic 仍将采取“多芯片策略”，计划在继续扩大规模的同时，将其他公司的硬件与其自研设计结合使用。这证实了此前流传的一则传闻；《The Information》曾报道称，Anthropic 正在考虑与三星合作，将其作为硬件制造合作伙伴。

Anthropic is not alone in walking this path. Its competitor, OpenAI, recently announced a new custom chip called Jalapeño designed for large language model inference in data centers. OpenAI partnered with Broadcom to develop the chip. Google has been running its models on its own hardware for a while; Meta has also designed and deployed its own chips, and Mistral is reportedly looking into doing the same.

走这条路的不止 Anthropic 一家。其竞争对手 OpenAI 最近宣布推出一款名为“Jalapeño”的新型定制芯片，专为数据中心的大语言模型推理而设计。OpenAI 与博通（Broadcom）合作开发了该芯片。谷歌长期以来一直使用自研硬件运行其模型；Meta 也设计并部署了自己的芯片，据报道 Mistral 也在考虑采取同样的做法。

There are a few reasons AI providers are doing this. First, much of the industry is heavily reliant on Nvidia for the hardware the companies’ models run on, and Nvidia’s continued leverage there is a potential strategic vulnerability, especially as AI companies operate in an environment where compute infrastructure is highly competitive as demand continues to outstrip current capacity.

人工智能提供商这样做有几个原因。首先，整个行业在运行模型所需的硬件上严重依赖英伟达（Nvidia），而英伟达在该领域的持续影响力构成了一种潜在的战略脆弱性，尤其是在人工智能公司所处的环境中，计算基础设施竞争激烈，且需求持续超过现有产能。

Second, designing chips for specific models and vice versa could lead to better performance. So, for example, if OpenAI can reap the rewards of that vertical integration, you can bet Anthropic and other frontier model providers will want that advantage as well. To that point, Anthropic says its teams will co-design new hardware and models side by side. It has co-designed certain hardware with partners before, but the plan is now to bring more silicon expertise inside Anthropic itself.

其次，为特定模型设计芯片（反之亦然）可以带来更好的性能。因此，举例来说，如果 OpenAI 能从这种垂直整合中获益，那么可以肯定的是，Anthropic 和其他前沿模型提供商也会希望获得同样的优势。就这一点而言，Anthropic 表示其团队将同步协同设计新硬件和模型。此前，他们曾与合作伙伴共同设计过某些硬件，但现在的计划是将更多的芯片专业知识引入 Anthropic 内部。

Anthropic may also hope this could help its frontier models get some extra competitive edge as software developers and other users begin exploring running cheaper, smaller, or open-weight models on their own hardware or on edge devices. However, since Anthropic is still in the process of hiring key team members, it will be some time before either the company or its users see any benefits.

Anthropic 可能还希望，随着软件开发人员和其他用户开始探索在自己的硬件或边缘设备上运行更便宜、更小巧或开放权重的模型，此举能帮助其前沿模型获得额外的竞争优势。然而，由于 Anthropic 仍处于招聘关键团队成员的阶段，无论是公司还是用户，要看到实际成果还需要一段时间。