Anthropic has officially confirmed that it is building a silicon engineering team to develop proprietary accelerators for Claude. This decision marks a strategic shift for one of the leading AI companies: instead of relying solely on NVIDIA and AMD GPUs or Google Cloud and Amazon Web Services (AWS) accelerators, it aims to design some of its hardware specifically tailored to its models’ needs.
The key points of Anthropic’s custom chip in 20 seconds
- Anthropic has confirmed it is forming a team to design custom chips for Claude.
- The company seeks to optimize hardware, software, and AI models in tandem.
- This project will not replace NVIDIA GPUs or Google and AWS chips in the short term.
- The strategy follows the trend initiated by OpenAI, Microsoft, Meta, and other major industry players.
In recent months, various reports have pointed to this move, but until now, the company had not publicly confirmed its plans. The publication of specialized job openings and statements collected by Business Insider clarify the situation: Anthropic wants to directly participate in designing the hardware on which future generations of Claude will run.
The goal is no longer just to buy accelerators, but to design them
The growth of artificial intelligence is forcing companies to rethink their dependence on a limited group of chip manufacturers. NVIDIA continues to dominate the AI accelerator market, while AMD tries to gain market share with its Instinct family. At the same time, major cloud providers have developed their own processors, such as Google’s TPU and AWS’s Trainium and Inferentia.
Anthropic aims to position itself differently. Its goal isn’t to become a general-purpose chip manufacturer but to develop accelerators specifically tuned to Claude’s requirements.
Job listings show that the company seeks engineers with expertise in all phases of integrated circuit development: front-end design, pre-silicon verification, physical design, analog integration, advanced packaging, signal integrity, power management, and foundry relations.
This indicates that Anthropic wishes to be involved in almost every aspect of chip development, although it will likely continue to rely on external partners for manufacturing and certain technological blocks.
Hardware and models developed jointly
One of the most significant changes is the focus on joint development of hardware and software.
Instead of adapting Claude to existing hardware on the market, Anthropic wants some of the hardware to evolve based on the specific characteristics of its models. This would allow optimization of aspects such as:
- Inference speed.
- Energy consumption per query.
- Memory utilization.
- Performance under specific Claude workloads.
- Economic efficiency in large-scale deployments.
This approach is reminiscent of other tech companies that have chosen to design silicon tailored for their workloads rather than relying solely on general-purpose processors.
NVIDIA will remain essential for years to come
Despite the announcement, Anthropic will not break its dependence on current suppliers.
The company maintains close collaboration with AWS, where it already deploys large-scale systems based on Trainium2, and also works with Google Cloud infrastructure, where Claude leverages Tensor Processing Units (TPUs) for certain tasks.
Furthermore, NVIDIA continues to be a cornerstone of foundational model training across the industry.
Therefore, Anthropic’s future involves a multi-chip strategy, where different architectures will coexist based on the workload type.
A move aligned with industry trends
This announcement confirms a trend that is accelerating in 2026.
An increasing number of AI model developers are trying to control not just the software but also the hardware on which their systems run.
OpenAI is already working on its own accelerator with Broadcom. Meta has been developing its Meta Training and Inference Accelerator (MTIA) for years. Microsoft continues expanding its Maia family, Google evolves its TPUs generation after generation, and Amazon keeps reinforcing its Trainium and Inferentia chips.
In this context, Anthropic’s decision makes sense. The costs of training and inference continue to rise, and access to cutting-edge accelerators remains one of the main bottlenecks industry-wide.
Although the company has not announced specific dates or technical details, its strategic message is clear: Claude’s future will not rely solely on third-party hardware evolution.
If the project advances to production, Anthropic will be able to jointly optimize models, compilers, kernels, memory, and silicon architecture to improve performance and energy efficiency. While NVIDIA and AMD will not become obsolete, this shift bolsters a growing industry trend: major AI firms aim to control the entire tech stack, from models to the chips that run them.
Frequently Asked Questions
Will Anthropic stop using NVIDIA GPUs?
No. The company will continue using NVIDIA hardware, along with AWS’s Trainium accelerators and Google’s TPUs, while developing its own silicon.
What will Anthropic’s custom chip be used for?
Its goal is to optimize Claude’s operation by improving performance, energy efficiency, and cost-effectiveness through hardware specifically designed for its models.
Will Anthropic manufacture its own chips?
Not necessarily. Designing the chip and manufacturing it are separate processes. Production will likely be handled by a specialized foundry, though the company has not yet announced its partner.
Is this a common strategy in the AI industry?
Increasingly so. OpenAI, Meta, Microsoft, Google, and Amazon are all investing in proprietary accelerators to reduce costs, boost performance, and lessen reliance on external suppliers.
