Anthropic joins OpenAI, Google DeepMind, and Meta in a wave of custom chip development driven by a global chip shortage, rising inference costs, and the performance gains possible when hardware and models are designed together.

Why Anthropic Is Building Its Own Chips

Anthropic's decision to develop custom silicon is motivated by several converging pressures:

• Chip shortage and supply constraints. A global shortage of AI chips makes it difficult for Anthropic to secure the computing power needed to train and deploy increasingly advanced models .
• Soaring inference costs. Serving billions of tokens daily at an approximately $30 billion revenue run-rate makes custom silicon a powerful lever for cost reduction. Analysts estimate that co-designed chips could cut Claude inference costs by as much as half .
• Model-hardware co-optimization. Designing chips explicitly for Claude's architecture allows Anthropic to deliver better performance and efficiency than off-the-shelf GPUs can achieve .
• Multi-chip strategy. The custom chips will supplement, not replace, existing partnerships with Nvidia, AMD, AWS, and Google Cloud, as part of a broader diversification away from sole reliance on Nvidia GPUs .

Anthropic has not disclosed a target date for tape-out or production, which is normal for a program at this stage .

How Other Major AI Companies Are Building Custom Chips

OpenAI and the Jalapeño Chip

OpenAI unveiled its first custom AI chip, codenamed Jalapeño, on June 24, 2026 . Designed in collaboration with Broadcom and manufactured by TSMC, Jalapeño is an inference-specialized processor . OpenAI began the design effort in early 2025 and initially explored building its own foundries before pivoting to a design-partner model . The company has also announced a broader strategic collaboration with Broadcom worth up to $10 billion to deploy accelerator and networking systems through 2029 .

Google DeepMind and Eight Generations of TPUs

Google has the longest-running in-house AI chip program, with over a decade of custom silicon development. Its Tensor Processing Unit (TPU) family now spans eight generations .

The Ironwood TPU (7th gen), launched in late 2025, can connect 9,216 chips in a single pod for training and inference . In April 2026, Google announced its 8th-generation TPUs with two purpose-built architectures: TPU 8t for training and TPU 8i for inference . Google has also reportedly tapped AMD to help design one of its 10th-generation TPUs .

Notably, Google also supplies TPUs to Anthropic under a massive deal worth tens of billions of dollars, providing up to one million chips .

Meta and the MTIA Family

In March 2026, Meta unveiled a roadmap of four new in-house chips under its Meta Training and Inference Accelerator (MTIA) family . The MTIA 300 was already deployed; the MTIA 400, MTIA 450, and MTIA 500 are scheduled for release roughly every six months through 2027 .

Meta is co-developing these chips with Broadcom and plans to start manufacturing one of them in September 2026 as part of a push to double its overall computing capacity . The MTIA family covers both inference and, for the first time, custom training chips — previously Meta relied on Nvidia GPUs for model training .

Summary of Custom AI Chip Programs

What This Means for the AI Hardware Landscape

The shift toward custom silicon represents a fundamental change in how AI companies approach infrastructure. Rather than competing solely on model architecture and training data, the industry's leaders are now building hardware as a competitive advantage.

For Anthropic, the custom chip program is still in its earliest stages — no target date for production has been set, and no architectural details or manufacturing agreements have been announced . But the move signals that Anthropic views chip design as essential to its long-term strategy, especially as inference costs grow with the scale of deployments.

For the industry as a whole, the diversification away from Nvidia's GPUs could have far-reaching implications. If custom chips deliver on their promise of lower costs and higher performance, the economics of AI inference could shift dramatically, making AI applications more affordable and accessible.

Anthropic joins OpenAI, Google DeepMind, and Meta in a wave of custom chip development driven by a global chip shortage, rising inference costs, and the performance gains possible when hardware and models are designed together.

Why Anthropic Is Building Its Own Chips

Anthropic's decision to develop custom silicon is motivated by several converging pressures:

• Chip shortage and supply constraints. A global shortage of AI chips makes it difficult for Anthropic to secure the computing power needed to train and deploy increasingly advanced models .
• Soaring inference costs. Serving billions of tokens daily at an approximately $30 billion revenue run-rate makes custom silicon a powerful lever for cost reduction. Analysts estimate that co-designed chips could cut Claude inference costs by as much as half .
• Model-hardware co-optimization. Designing chips explicitly for Claude's architecture allows Anthropic to deliver better performance and efficiency than off-the-shelf GPUs can achieve .
• Multi-chip strategy. The custom chips will supplement, not replace, existing partnerships with Nvidia, AMD, AWS, and Google Cloud, as part of a broader diversification away from sole reliance on Nvidia GPUs .

Anthropic has not disclosed a target date for tape-out or production, which is normal for a program at this stage .

How Other Major AI Companies Are Building Custom Chips

OpenAI and the Jalapeño Chip

OpenAI unveiled its first custom AI chip, codenamed Jalapeño, on June 24, 2026 . Designed in collaboration with Broadcom and manufactured by TSMC, Jalapeño is an inference-specialized processor . OpenAI began the design effort in early 2025 and initially explored building its own foundries before pivoting to a design-partner model . The company has also announced a broader strategic collaboration with Broadcom worth up to $10 billion to deploy accelerator and networking systems through 2029 .

Google DeepMind and Eight Generations of TPUs

Google has the longest-running in-house AI chip program, with over a decade of custom silicon development. Its Tensor Processing Unit (TPU) family now spans eight generations .

The Ironwood TPU (7th gen), launched in late 2025, can connect 9,216 chips in a single pod for training and inference . In April 2026, Google announced its 8th-generation TPUs with two purpose-built architectures: TPU 8t for training and TPU 8i for inference . Google has also reportedly tapped AMD to help design one of its 10th-generation TPUs .

Notably, Google also supplies TPUs to Anthropic under a massive deal worth tens of billions of dollars, providing up to one million chips .

Meta and the MTIA Family

In March 2026, Meta unveiled a roadmap of four new in-house chips under its Meta Training and Inference Accelerator (MTIA) family . The MTIA 300 was already deployed; the MTIA 400, MTIA 450, and MTIA 500 are scheduled for release roughly every six months through 2027 .

Meta is co-developing these chips with Broadcom and plans to start manufacturing one of them in September 2026 as part of a push to double its overall computing capacity . The MTIA family covers both inference and, for the first time, custom training chips — previously Meta relied on Nvidia GPUs for model training .

Summary of Custom AI Chip Programs

What This Means for the AI Hardware Landscape

The shift toward custom silicon represents a fundamental change in how AI companies approach infrastructure. Rather than competing solely on model architecture and training data, the industry's leaders are now building hardware as a competitive advantage.

For Anthropic, the custom chip program is still in its earliest stages — no target date for production has been set, and no architectural details or manufacturing agreements have been announced . But the move signals that Anthropic views chip design as essential to its long-term strategy, especially as inference costs grow with the scale of deployments.

For the industry as a whole, the diversification away from Nvidia's GPUs could have far-reaching implications. If custom chips deliver on their promise of lower costs and higher performance, the economics of AI inference could shift dramatically, making AI applications more affordable and accessible.