Anthropic Is Building Its Own AI Chips
The company behind the Claude chatbot is hiring chip designers to make its AI faster and cheaper to run, following a path already taken by OpenAI, Google and Meta.
Key points
- Anthropic confirmed in August 2026 that it is forming a custom silicon team to design its own AI chips.
- The company is already in talks with Samsung as a potential manufacturing partner, according to a report first published by The Information last month.
- Anthropic currently buys computing power from AWS, Google, Nvidia and AMD, but rising demand for Claude has made that arrangement harder to sustain at scale.
- OpenAI unveiled its own chip, called Jalapeño, in June 2026, built with semiconductor company Broadcom.
- Alphabet (Google's parent) and Meta have both been running on proprietary chips for years.
Anthrophe wants to stop renting and start owning, at least when it comes to the silicon, the physical chips, that power its AI. The company is advertising roles on a "custom silicon team," a group that will design specialised processors built specifically to run Claude, TechCrunch AI first reported.
Why would an AI company bother making its own chips?
Right now, running a large AI model costs enormous amounts of money because companies must rent access to powerful chips from suppliers like Nvidia. Building your own chip means you can design it around exactly what your software needs, cutting waste and, eventually, cost.
Anthropec already has deals in place with AWS, Google, Nvidia and AMD to access computing hardware. But as demand for Claude grows, buying time on someone else's chips puts a ceiling on how fast the company can expand. Designing its own silicon removes that ceiling.
Who else is doing this?
Several of the biggest names in AI have taken this road. OpenAI revealed its first in-house chip, named Jalapeño, in June 2026. Jalapeño was manufactured by Broadcom and built for inference workloads, meaning it handles the moment a user types a question and the model generates an answer, the most common and expensive task an AI system does every day.
Google's DeepMind has long relied on chips called TPUs, short for Tensor Processing Units, which Alphabet builds internally. Meta has been developing its own accelerator chips, called MTIA, for similar reasons.
A rough comparison of where the major players stand:
| Company | In-house chip | Status (mid-2026) |
|---|---|---|
| Google / DeepMind | TPU (various generations) | Long in production |
| Meta | MTIA accelerator | In development |
| OpenAI | Jalapeño (via Broadcom) | Unveiled June 2026 |
| Anthropic | Custom silicon (unnamed) | Hiring now |
What does this mean for people who use Claude?
In the short run, nothing changes. Chip design takes years before it reaches production. The practical benefit for everyday users would come later: faster responses, lower outage risk during busy periods, and potentially cheaper access to Claude-powered tools as Anthropic's running costs fall.
The honest note worth keeping in mind: most AI companies that announce chip ambitions take three to five years to ship real silicon at scale. Hiring a team is step one of a very long process.
Anthropec did not respond to a request for comment.
Takeaway: If you use Claude for work today, nothing changes. But watch this story over the next two to three years. The companies that control their own chips will have a real cost advantage over those that keep renting, and that will eventually show up in pricing.
