In a notable move, Anthropic is expanding its workforce by seeking individuals with direct experience in semiconductor design. The positions offer a salary range between $320,000 and $485,000 and are presented straightforwardly, emphasizing the need for candidates who have successfully shipped silicon and can operate independently without large organizational support. On Wednesday, Anthropic confirmed its intent to establish an in-house custom silicon team aimed at designing chips specifically for its AI model, Claude.
This development marks a significant addition to a trend among major tech players. Google has its Tensor Processing Units (TPUs), Amazon offers Trainium and Inferentia, OpenAI utilizes Broadcom-designed inference processors, and both Meta and Microsoft are engaged in their own silicon initiatives. With Anthropic’s recruitment announcement, it’s clear that even the last major lab without custom silicon is stepping towards developing its own capabilities, a shift away from reliance on Nvidia.
Anthropic has articulated a focus on co-design, where the chip and AI model are developed simultaneously to optimize performance. A company spokesperson emphasized this approach is crucial to ensuring "Claude runs faster and more efficiently at the scale users require." This philosophy, previously adopted by Apple for its M-series and by Google for TPUs, is rooted in the principle that hardware tailored for a specific workload can operate more efficiently by excluding unnecessary features.
Despite this effort to build custom chips, Anthropic has clarified that it will maintain a multi-chip strategy, leveraging infrastructure from AWS, Google, Nvidia, and AMD. There was no indication of when these chips might be ready or whether the company will initiate manufacturing on its own. Although there have been reports about potential collaboration with Samsung for a custom chip, these have not been confirmed as a manufacturing agreement.
In the meantime, Anthropic is not waiting for its in-house team to begin utilizing custom chips. Earlier this year, the company expanded its existing partnership with Google and Broadcom to secure approximately 3.5 gigawatts of next-generation TPU capacity, scheduled to go live in 2027. This is in addition to the gigawatt of capacity anticipated to arrive in 2026 through a Google Cloud agreement made in the previous October. Broadcom is pivotal to this arrangement, as it not only provides the custom TPUs but also has obligations for networking components within Google's next-gen AI infrastructure, set to run through 2031. To put it in perspective, 3.5 gigawatts is equivalent to the energy consumption of a medium-sized city, exclusively dedicated to running Anthropic’s AI models. Krishna Rao, the CFO of Anthropic, described this commitment as the company's most significant investment in computing resources to date, reflecting the rapid expansion of its customer base. Anthropic's revenue has seen impressive growth, rising from about $9 billion at the end of 2025 to over $30 billion now, with more than a thousand clients each spending upwards of $1 million annually. Such revenue surges make investing in chip development a realistic consideration for the company, even as it has been procuring custom silicon from Broadcom and Google for the past year.
In the competitive landscape of AI, the necessity for in-house chip development stems from basic economic calculations. Anthropic, with its $30 billion run-rate revenue, processes vast amounts of data daily, and the costs associated with this are largely dictated by third-party hardware. Reducing these costs could yield significant savings that benefit every query processed going forward. General-purpose accelerators, while versatile, carry inefficiencies that aren't utilized by a specific lab's models. A chip specifically designed for Anthropic's model family could optimize performance by eliminating unnecessary overhead. However, developing advanced chip technologies entails substantial investment and time commitments, necessitating that the model architecture remains consistent enough to align with hardware advancements. This is why the labs undertaking this path typically have sufficient revenue to spread the costs and stable models to target effectively.
