Anthropic eyes custom AI chip, talks with Samsung

Key Takeaways

• Anthropic is exploring custom chip development with Samsung, though plans remain early-stage
• The move follows OpenAI's Jalapeño chip launch and similar efforts by Amazon, Microsoft, and Meta
• Anthropic already secured 3.5 GW of Google TPU capacity starting 2027, suggesting the Samsung chip would complement rather than replace existing arrangements

Anthropic is holding early discussions with Samsung about developing a custom AI chip, according to The Information. The talks come as AI companies scramble to secure dedicated silicon and reduce their dependence on NVIDIA, which controls more than 80% of the AI training chip market.

The report indicates Anthropic has not decided what the chip would be used for, how it would integrate into its servers, or its target performance specifications. That vagueness matters. It suggests Anthropic is still evaluating whether to optimize for training, inference, or a hybrid approach.

Why Anthropic needs its own silicon

Anthropic currently runs Claude on a mix of hardware. Google's Tensor Processing Units handle some workloads. Amazon's custom chips power other parts of the infrastructure, a natural fit given Amazon's $4 billion investment in the company. But relying on partners for critical infrastructure creates dependency and limits optimization.

In April, Anthropic signed a long-term deal with Google and Broadcom for roughly 3.5 gigawatts of AI computing capacity via Google's TPUs, starting in 2027. That agreement secures supply but does not give Anthropic the ability to tune hardware specifically for Claude's architecture. Custom silicon would.

The economics are stark. Training frontier AI models demands massive compute clusters, often 40,000 or more high-end GPUs. But inference, the process of serving responses to users, consumes even more resources over time because every chatbot query, every API call, burns additional compute cycles. Control over inference hardware directly affects margins.

The race away from NVIDIA

OpenAI moved first among the pure-play AI labs. Last month it introduced Jalapeño, its first in-house inference chip, built with Broadcom and manufactured by TSMC. The chip will run inside OpenAI's own infrastructure rather than being sold commercially.

The hyperscalers have been at this longer. Amazon developed Trainium for training and Inferentia for inference, both available to AWS customers. Microsoft built the Maia accelerator for Azure. Meta continues expanding its MTIA chips for recommendation systems and generative AI workloads. Google has iterated on TPUs for over a decade.

Each company faces the same calculation. Designing a cutting-edge AI chip costs roughly $500 million, according to Reuters, factoring in specialist engineering talent and exhaustive manufacturing validation. But at sufficient scale, owning the silicon stack pays back in lower operating costs and hardware precisely matched to model architectures.

Why Samsung?

Samsung is one of three companies in the world capable of manufacturing advanced chips at scale, alongside TSMC and Intel. TSMC dominates AI chip production, it makes NVIDIA's GPUs, AMD's accelerators, and now OpenAI's Jalapeño. Partnering with Samsung could give Anthropic manufacturing capacity without competing for TSMC allocation against every other AI player.

Samsung also brings memory expertise. High-bandwidth memory (HBM) is a critical bottleneck for AI chips, and Samsung is one of the top three HBM producers. A partnership could tightly integrate processor and memory development.

That said, Samsung's foundry business has struggled with yields on advanced process nodes compared to TSMC. Any Anthropic chip would need to work within those constraints or wait for Samsung to close the gap.

What this means for the chip shortage

The AI chip shortage is not ending soon. "We're going to have a supply shortage for chips for years," Sam Altman has said. Custom chip development does not solve this problem immediately. Chips take three to five years from initial design to volume production.

But the trend is clear. Every major AI company is building optionality. Google, Amazon, Microsoft, Meta, and now OpenAI and Anthropic are all investing in proprietary silicon. NVIDIA will remain dominant for years, but its pricing power may erode as alternatives mature.

Frequently Asked Questions

Is Anthropic building its own AI chip?

Anthropic is in early-stage talks with Samsung about developing a custom AI chip, but has not finalized specifications, use case, or timeline.

What chips does Anthropic currently use for Claude?

Anthropic uses a mix of Google TPUs and Amazon custom chips to train and run Claude, reflecting investments from both companies.

Why are AI companies building custom chips instead of using NVIDIA?

Custom chips can be optimized for specific model architectures, reduce per-query inference costs, and decrease dependence on NVIDIA's supply constraints and pricing power.

How much does it cost to design an AI chip?

According to Reuters, designing a cutting-edge AI chip costs around $500 million, accounting for specialized engineering and manufacturing validation.

Source: Tech-Economic Times / ET