Anthropic announced ambitions to design and manufacture AI accelerators in-house, joining Nvidia as a chip developer.

Artificial intelligence has created the fastest infrastructure buildout the technology industry has ever seen. Every new AI model requires more computing power than the last, forcing cloud providers and AI developers into a race to secure enough chips to keep expanding. NVIDIA and Advanced Micro Devices have ramped up production at a historic pace, yet demand continues to outrun supply. That imbalance is now changing the competitive landscape. Instead of waiting for more chips to become available, the world's largest AI companies are deciding to design their own.

Anthropic confirmed Wednesday that it is building an in-house chip design team to create custom silicon for its Claude family of AI models, validating an exclusive Reuters report from April. According to the report, Anthropic will continue hiring engineers capable of co-designing hardware and software to improve Claude's speed and efficiency while maintaining relationships with its existing hardware suppliers. Anthropic stressed that this is part of a multi-chip strategy, not a replacement for partners such as NVIDIA, AMD, Amazon's AWS, and Google Cloud. Instead, the company wants greater flexibility as demand for AI computing continues climbing.

The move reflects just how constrained AI hardware remains. NVIDIA's Blackwell GPUs remain difficult to obtain as cloud providers continue expanding AI infrastructure. Meanwhile, shipments of NVIDIA's next-generation Vera Rubin platform are already ramping up despite carrying price tags of roughly $55,000 per GPU. A fully populated Vera Rubin NVL72 rack containing 72 GPUs is estimated to cost between $7.8 million and $9.1 million before networking and supporting infrastructure are included. A hyperscale AI data center can deploy anywhere from 5,000 to 50,000 racks. With a single server rack now costing up to $9 million, the world's AI leaders are ditching the waiting list to build their own silicon.

Anthropic is hardly alone. OpenAI recently unveiled its first custom AI chip developed alongside Broadcom, while Meta Platforms continues developing its MTIA accelerators. Amazon already designs Trainium and Inferentia chips, and Google has spent years building its Tensor Processing Units. Even outside technology, custom silicon is becoming more common. Automakers including Hyundai, Mercedes-Benz, and NIO are designing specialized chips for autonomous driving and AI-powered vehicles. Rather than relying entirely on third-party suppliers, companies increasingly want hardware optimized for their own software.

These companies are not trying to replace NVIDIA overnight. Designing a cutting-edge AI processor can cost roughly $500 million before manufacturing even begins. Instead, custom chips let AI developers optimize certain workloads, lower operating costs, and diversify supply chains that remain under pressure. If NVIDIA and AMD had excess manufacturing capacity, there would be far less incentive for customers to spend hundreds of millions of dollars designing their own processors. Instead, AI developers are concluding they need every source of computing they can find.

Custom silicon could eventually reduce some purchases of merchant GPUs. Yet Anthropic explicitly said it will continue buying hardware from NVIDIA, AMD, Amazon, and Google as part of its diversified strategy. Anthropic isn't abandoning NVIDIA—it is acknowledging that one supplier, no matter how dominant, simply can't satisfy all of AI's unprecedented demand. That's a bullish signal not just for NVIDIA, but for the entire AI semiconductor ecosystem.