Anthropic to design its own chips for its Claude models

Anthropic is building an internal team to design chips for its Claude models. The company wants to develop hardware and software simultaneously to make its AI systems faster and more efficient, as it told news agency Reuters, among others. The move reduces dependence on external chipmakers, but does not replace those suppliers entirely.

Anthropic is pursuing a so-called multi-chip strategy: alongside its self-designed processors, the company will continue to use chips from Amazon Web Services, Google, Nvidia and AMD. This is therefore a complement to the existing supplier network, not a full transition to proprietary silicon.

The move follows a period of strong revenue growth. At the start of 2025, Claude's annualised revenue stood at around $1 billion; by August 2025 that figure had already exceeded $5 billion. Estimates of more than $30 billion are cited for 2026, though such projections should be treated with appropriate caution. That scale makes the business case for an in-house chip programme more tangible.

Why Anthropic now wants to build its own chips

The motivation behind the initiative is twofold. On one hand, the persistent shortage in the AI chip market is a factor: large models require enormous amounts of computing power, and the availability of suitable processors is a recurring bottleneck for AI companies. On the other hand, co-designing hardware and software creates room to align Claude models more precisely with the underlying architecture, which in theory leads to lower latency and better energy efficiency.

Developing an advanced AI chip in-house is costly. Estimates point to approximately $500 million for the design alone, excluding manufacturing costs. That figure is achievable for a company at Anthropic's revenue level, but it also illustrates why smaller AI labs are unlikely to take this route anytime soon.

Anthropic has already hired a senior silicon engineer who previously served as the second chip engineer at OpenAI and also gained experience at Tesla. The team is continuing to recruit at the intersection of hardware and software engineering.

Samsung as a potential manufacturing partner

For the actual production of the chips, Anthropic is holding exploratory talks with Samsung. Those discussions cover topics including a 2-nanometre manufacturing process and advanced packaging techniques, though the talks are still at an early stage. Samsung recently also invested in Anthropic, further deepening the relationship between the two companies.

It is not yet known when a first self-designed chip would be ready or when it would enter production. Chip development typically involves long lead times: from design to working chip, multiple years are common, depending on the complexity and the chosen manufacturing process.

Anthropic follows a pattern established by large tech companies

Anthropic is not the first company to take this path. Google has been developing its own Tensor Processing Units (TPUs) for training and running AI models for years. Meta is building the MTIA accelerator, Amazon has Trainium and Inferentia in-house, and OpenAI is working with Broadcom on its own chip under the name Jalapeño.

What sets Anthropic apart from some of these examples is that it is primarily an AI research and product company without the broader chip infrastructure that an Amazon or Google has already built up. Establishing an in-house chip design team therefore requires a considerable organisational and financial investment that is separate from its core activity: building and improving language models.

For the European and Dutch AI scene, this news is a signal that the vertical integration of AI systems, from silicon to model, is becoming an increasingly strategic theme among major players. European AI companies and policymakers focusing on sovereign computing capacity or their own infrastructure may recognise a comparable trade-off here: anyone who wants to maintain control over AI performance and costs in the long run will sooner or later face the question of how deep that control needs to reach into the stack.