Anthropic is building custom AI chips for Claude alongside existing suppliers.Custom chips give AI firms more control over cost and performance.

 

Anthropic is forming an internal semiconductor team to design custom chips for its Claude AI models, marking its first public confirmation that it plans to develop its own silicon.

The company said it is recruiting engineers across hardware and software disciplines to develop chips alongside its AI models. The aim is to improve Claude’s performance and efficiency while supporting large-scale deployments.

Anthropic has not disclosed when its first chip will be completed or whether it plans to manufacture the hardware itself. Reuters reported in April that the company was considering developing its own AI processors, while The Information later reported that Anthropic had held talks with Samsung Electronics as a potential manufacturing partner.

The project will not replace Anthropic’s existing hardware suppliers. The company said custom silicon will form part of a broader multi-chip strategy that continues to use processors and computing infrastructure from Amazon Web Services, Google, Nvidia, and AMD.

Anthropic is already running Claude across custom accelerator architectures developed by other companies. AWS says almost one million of its Trainium2 processors are being used to train and serve Claude through Project Rainier, an AI infrastructure project developed with Anthropic.

Reuters reported when the project launched that the cluster initially contained nearly 500,000 Trainium2 chips, with Anthropic expected to use more than one million across AWS by the end of 2025. The infrastructure is used for both developing and deploying Claude.

Anthropic’s own silicon would add another processor architecture to that existing mix. The company has said AWS, Google, Nvidia, and AMD will remain part of its computing strategy.

Anthropic’s recruitment efforts provide further detail about the scope of the project. A recent job posting for its custom silicon team sought engineers with experience taking semiconductor designs through development, verification, and production.

The position carried a listed salary range of $320,000 to $485,000 and called for candidates who had previously contributed directly to completed semiconductor designs. The posting also indicated that engineers would be expected to make design and scheduling decisions within a relatively small team.

Developing a high-end AI processor requires substantial investment before manufacturing begins. Industry sources cited by Reuters estimate that designing an advanced AI chip can cost about $500 million, reflecting engineering costs as well as the testing and verification needed before a design can enter production.

Anthropic joins Google, AWS, Microsoft, Meta, and OpenAI among companies developing custom AI processors for their own infrastructure or services.

Inference drives custom chip design

Inference has become a major target for recent custom silicon projects.

Microsoft designed its Maia 200 accelerator specifically around large-scale AI inference. The company says the processor provides 30% better performance per dollar than the latest-generation hardware already running in its fleet and is being used for workloads including OpenAI models and Microsoft’s own AI services.

OpenAI has taken a similar approach with its first custom accelerator, developed with Broadcom. Reuters reported in June that the processor is intended primarily for inference and will be used internally rather than sold as a commercial chip.

Reuters also reported that reducing infrastructure costs and adding another source of compute alongside Nvidia GPUs were among the objectives of OpenAI’s chip programme.

AMD’s acquisition of AI chip startup Taalas in August adds another inference-focused example. Taalas has been developing technology aimed at reducing compute and memory bottlenecks when AI models are running.

The focus is not limited to processing power. Custom chips also allow companies to design memory, networking, and software around particular AI workloads.

Microsoft’s Maia 200 includes 216GB of HBM3e high-bandwidth memory and 7 terabytes per second of memory bandwidth. The company says the architecture was designed to keep large models supplied with data and increase token throughput during inference.

Google’s seventh-generation Ironwood TPU provides 192GB of high-bandwidth memory per chip and about 7.37 terabytes per second of memory bandwidth, according to Google Cloud documentation. The previous Trillium generation offered 32GB and 1.64 terabytes per second.

Google notes that high-bandwidth memory can remain a constraint for some memory-intensive AI workloads. That makes memory bandwidth another design consideration alongside raw compute performance.

Microsoft and Google are also designing compilers, networking, memory systems, and software around their custom accelerators. Anthropic has similarly said it wants engineers across the hardware and software stack to co-design processors and AI models.

Power and infrastructure costs

Power use is another design target as AI companies deploy larger numbers of accelerators inside data centres. Rather than measuring only peak processing performance, chip developers increasingly publish performance-per-watt figures for successive processor generations.

Google says Ironwood delivers about twice the performance per watt of its previous-generation Trillium TPU. The company also says the processor is nearly 30 times more power-efficient than its first Cloud TPU introduced in 2018.

AWS says Trainium3 can provide up to four times better performance per watt than Trainium2 and up to five times more output tokens per megawatt at similar latency.

Microsoft has also designed Maia 200 around both performance per dollar and performance per watt for large inference deployments. Its processor operates within a 750-watt system-on-chip power envelope, while the wider Maia infrastructure includes custom networking and cooling systems.

These programmes extend beyond the processor itself. Microsoft and Google are integrating memory, networking, software, and data-centre infrastructure around their custom accelerators.

Custom processors can be designed around specific training or inference workloads, rather than relying entirely on externally supplied accelerators.

That does not mean AI companies are replacing Nvidia or AMD altogether. Several proprietary chip programmes are intended to run alongside externally supplied GPUs.

Meta provides one example. Reuters reported in July that the company plans to put its next-generation internal AI processor into production in September as part of its Meta Training and Inference Accelerator, or MTIA, programme.

The processor is intended to lower infrastructure costs and give Meta greater control over its computing requirements, but it is designed to complement Nvidia and AMD GPUs already used across the company’s data centres.

Meta has also tied the programme to a much larger expansion of its computing infrastructure. Reuters reported that the company expects AI infrastructure spending of as much as $145 billion in 2026 and is working toward 14 gigawatts of computing capacity in 2027.

OpenAI is likewise continuing to use Nvidia hardware while developing its Broadcom-designed processor, while Anthropic has said its own chips will sit within a broader multi-chip strategy.

More AI companies build their own silicon

Anthropic’s plans place it alongside other AI companies developing proprietary processors. OpenAI unveiled its purpose-built chip developed with Broadcom in June, while Meta is preparing another generation of its internal AI processor for production.

Google has been developing its own processors for longer than most AI companies now entering the market. Ironwood is the seventh generation of its Tensor Processing Unit architecture and can be deployed in configurations containing thousands of connected chips.

AWS has developed its Trainium and Inferentia families of purpose-built AI chips, while Microsoft has expanded its Maia accelerator programme inside Azure. Those platforms continue to coexist with hardware supplied by Nvidia, AMD, and other chipmakers.

French AI company Mistral has also said it is considering developing its own silicon.

The programmes vary in scale and purpose, but many focus on the same infrastructure constraints: compute availability, inference cost, memory bandwidth, power efficiency, and tighter integration between hardware and software.

Meta has also said adopting successive generations of externally supplied GPUs at its scale requires substantial engineering work, while Microsoft and Google have built custom networking and software around their own accelerators.

Anthropic is entering that market with a computing strategy that already spans several chip suppliers and cloud platforms. Its use of AWS Trainium means Claude is already being trained and served on purpose-built AI processors alongside conventional GPU infrastructure.

The company has not yet provided technical specifications, a manufacturing partner, a launch date, or details on whether its first internally designed processor will focus on training, inference, or both.

 

 

 

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.