Anthropic co-designing custom AI inference chips to bypass costly Nvidia GPUs — Samsung reported as manufacturing partner for Claude maker
Anthropic has announced its building a team to co-design custom ASIC chips for AI inferencing workloads. This will give it greater control over its compute buildout and allow it to make its AI models specifically more efficient for its designs.
Anthropic, the maker of the popular AI language model Claude, is reportedly co-designing custom AI inference chips with an unspecified partner. The tech giant has started hiring engineers to develop the chips and aims to complete the project quickly, as highlighted in a job listing seeking a candidate who can work to a strict timeline.
This move by Anthropic is part of a broader trend where major AI companies, including Google, Meta, Microsoft, Amazon, and OpenAI, are investing in their own custom hardware to handle AI inferencing workloads more efficiently and cost-effectively.
The global shortage of chip supplies and the increasing focus on model optimizations have made custom silicon a more attractive option for AI companies. By designing their own chips, these companies can potentially reduce their total cost of ownership by up to 65%, and improve power efficiency compared to expensive Nvidia GPUs. Several companies, such as Google, Amazon, and Meta, have already invested in developing their own custom AI chips, including the Tensor Processing Unit (TPU), Trainium, and Inferentia, respectively.
While co-designing and manufacturing custom AI chips can be a costly and complex process, it is becoming a strategic move for the biggest players in the AI industry. Nvidia, however, is one of the main beneficiaries of this trend, as many AI companies still rely on its GPUs for inferencing workloads. Meanwhile, companies like TSMC, Broadcom, and Marvell stand to gain from the custom ASIC design and build market, as they provide the necessary manufacturing, packaging, and interconnect technologies for these custom chips.
Samsung is reportedly a potential manufacturing partner for Anthropic, bringing its expertise in memory production and advanced packaging technologies to the table. However, the true winner in this AI chip race is likely to be TSMC, the Taiwanese company that produces most of the world's cutting-edge silicon and is involved in the production of the majority of the custom AI chips discussed.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written; read the original for the full account.
