Leading artificial intelligence companies are racing to develop their own chips to secure cheaper and more efficient computing power. Model operators like Anthropic, Meta, and OpenAI—the end consumers of the AI computing ecosystem—are now moving to produce the "raw materials" of chips themselves, having previously relied primarily on Nvidia. While these companies have focused on securing data centers, the midstream segment of the ecosystem, they have concluded that controlling AI chips is ultimately necessary to lead the market. Some observers suggest this could become a long-term move away from Nvidia dependence.
Reuters reported on April 9, citing sources, that Anthropic, developer of the AI model "Claude," is exploring the development of its own chips. While noting this is still in early stages, Reuters explained that Anthropic aims to address the global semiconductor shortage and build more advanced AI systems.
Anthropic currently develops Claude by leasing data centers equipped with Nvidia graphics processing units (GPUs) and Amazon and Google chips. Although the company secured 3.5 gigawatts of computing capacity equipped with Google's Tensor Processing Units (TPUs) through a partnership with Google and Broadcom announced on April 7, it appears Anthropic has been internally planning its own chips.
OpenAI, which competes fiercely with Anthropic in AI models, revealed its ambitions for custom chip development early on. OpenAI announced in October last year that it would partner with Broadcom to develop custom chips, which will be installed in OpenAI's data centers according to the company's specifications.
Hyperscalers—large-scale data center operators—have been developing their own chips for years to gain market advantage. Google is considered the clear frontrunner. Anticipating an explosion in computing demand, Google began developing custom chips over a decade ago, first deploying TPUs internally in 2015 and releasing its first-generation cloud TPU in 2018.
Amazon and Microsoft have also unveiled their own chips to catch up with Google. Amazon has been developing "Inferentia" for inference and "Trainium" for training since 2018. Microsoft unveiled its first AI chip, "Maia 100," in November 2023, followed by "Maia 200" with improved inference capabilities earlier this year. Meta went a step further last month, declaring it would release chips independent of Nvidia every six months.
Amazon CEO Andy Jassy said in his shareholder letter released the same day, "Our estimated annual revenue from our own AI chips has exceeded $20 billion (29.6 trillion won). If we had sold chips this year, annual revenue would have reached $50 billion." He added, "Because chip demand is so high, there is a very high possibility of large-scale sales to third parties in the future." This suggests Amazon could sell chips separately, rather than only providing them through Amazon Web Services (AWS) data center leases as it does currently.
The reason these companies want their own chips is to break free from Nvidia dependence. With Nvidia controlling 90% of the data center chip market—the core of AI—AI developers are scrambling to secure Nvidia GPUs. Companies believe they must have their own chips to escape a structure where GPU supply volume and timing dictate their business outcomes.
Having custom chips allows companies to tailor data center designs, including cooling and power infrastructure, to extract more computing power from the same power consumption and floor space.
The challenge is that companies without design experience need enormous time and investment to develop their own chips. The Information noted that "more and more companies are pursuing custom AI chip development to reduce their dependence on Nvidia," while pointing out that "developing in-house chips takes much longer than signing additional cloud contracts." Reuters reported that designing advanced AI chips requires $500 million for personnel and manufacturing processes.
Due to these constraints, big tech companies are pursuing data center expansion strategies alongside custom chip development. Meta announced the same day a $21 billion data center construction contract with AI infrastructure company CoreWeave. Including existing contracts, Meta has committed $35.2 billion (52 trillion won) to secure data center capacity through 2032. While Meta is building its own data centers, the company apparently believes this is insufficient to outpace competitors and is seeking external support. Partnering with CoreWeave also offers the advantage of early access to Nvidia's latest AI accelerator, "Vera Rubin," enabling Meta to get ahead in the development race. Google also signed a multi-year supply agreement with Intel for the latest central processing units (CPUs) the same day, reflecting the growing importance of CPUs in the agentic AI market, where AI reasons independently and executes tasks.
