Reuters broke a story on April 9, 2026 that has been quietly reshaping how the AI industry thinks about its own supply chain. Anthropic, the company behind Claude, is exploring whether to design its own artificial intelligence chips. Such a move would reduce the company’s reliance on Google TPUs and Amazon’s custom silicon, the hardware Anthropic currently uses to train and run Claude.

The report, by Reuters journalists Max A. Cherney and Deepa Seetharaman, cites three people familiar with the matter. Importantly, the plans are at an early stage. Anthropic has not committed to a specific chip design, has not assembled a dedicated engineering team for the project, and may still decide to continue purchasing chips from external vendors rather than designing its own.

That last point matters for context: this is not an announcement of a product. It is a signal of where the pressure is building.

Why Anthropic Is Even Thinking About This

The financial backdrop makes the exploration understandable. Anthropic’s annualised revenue run rate crossed $30 billion in early April, up from roughly $9 billion at the end of 2025. More than 1,000 enterprise customers are now spending over $1 million per year with the company.

When you are at that scale, your compute bill is not a rounding error. Every token generated by Claude costs real money in chip cycles, cooling, power, and cloud fees. If Anthropic is running 80 percent of its revenue from enterprise customers, as recent reporting suggests, those customers are sending enormous volumes of inference traffic through Claude every day.

The arithmetic is simple: if you can design chips that run your own models more efficiently than general-purpose hardware or even existing TPUs, the cost savings compound very quickly at scale.

The $500 Million Problem

Designing a competitive AI chip is not a side project. Industry sources cited in the Reuters report estimate the development cost at roughly half a billion dollars, and that is before accounting for software co-design, fabrication ramp-up, testing, and ecosystem tooling. Some estimates put the all-in cost above $1 billion.

That is a substantial bet, especially when Anthropic already has a long-term compute deal with Google and Broadcom. The two companies announced a partnership in early April 2026 to provide Anthropic with approximately 3.5 gigawatts of next-generation TPU capacity coming online from 2027. Locking in that deal and then simultaneously exploring an alternative suggests Anthropic is not viewing these as mutually exclusive paths.

The more likely interpretation is that Anthropic is stress-testing its dependency on external providers. Even with contracted TPU access, Anthropic has no direct control over the chip architecture that runs its models. That is a strategic vulnerability for a company with $30 billion in revenue and ambitions that presumably extend well beyond 2027.

Everyone Is Doing This Now

Anthropic would not be the first AI lab to go down this road.

Meta has had a custom silicon programme for years, originally designed for recommendation algorithms and now expanding toward AI model inference. OpenAI has been building toward custom silicon, with reports of a dedicated chip team and procurement relationships with TSMC. Google designed its own TPUs specifically to run its AI workloads. Amazon’s Trainium and Inferentia chips power its own Bedrock offerings.

The pattern is consistent. As AI workloads become large enough to drive meaningful hardware economics, the major players start building chips that are optimised specifically for their software and their usage patterns. It is a vertical integration playbook that mirrors what Apple did with its M-series chips: give up hardware flexibility in exchange for software-hardware efficiency that no off-the-shelf product can match.

Anthropic entering this conversation, even at an exploratory stage, says something about how seriously the company is treating its own long-term trajectory.

What This Means for Business

The chip story is infrastructure news, but it has practical implications for businesses deploying AI today.

The steady cost reduction in AI inference that most enterprise customers have experienced over the past two years has been driven partly by hardware improvements and partly by software optimisation. As AI labs move toward custom silicon tuned specifically for their models, that cost reduction trajectory is likely to continue, and possibly accelerate.

For businesses currently worried about the cost of running AI agents at scale, the infrastructure news from this week is actually encouraging. The major AI labs are investing heavily to bring costs down. Today’s per-token pricing is not a ceiling. It is likely closer to the floor of a market that is still sorting out its economics.

The more immediate consideration is reliability. Companies building production workflows on top of Claude, ChatGPT, or Gemini are exposed to whatever supply-chain constraints affect those providers. Anthropic exploring chip independence is partly about cost, but it is also about control. For enterprise customers who care about uptime guarantees and throughput commitments, a vertically integrated AI provider is a more stable long-term partner.

That said: Anthropic designing its own chips in 2026 or 2027 does not solve any supply problems in 2026. If your business is deploying AI today, this news changes nothing about how you should procure and deploy AI tools now. What it does suggest is that the foundational infrastructure investment across the major AI labs is real, serious, and likely to produce better, more reliable, and cheaper AI services over the next three to five years.

The window for building operational AI competency is open. These infrastructure investments are the reason it will stay open.

---

If you’re building with Claude or Codex right now, grab the free Working With Claude field guide. Thirty-two pages on the full ecosystem, Claude Code in depth, and how to roll agents out properly. Get the free guide.