Anthropic confirmed on August 5, 2026 that it is building its own team to design custom chips. The company wants to co-design hardware and models so Claude runs faster and cheaper as demand for its services climbs. It marks a clear shift for a lab that until now leaned entirely on accelerators from outside providers.
Key takeaways
- Anthropic confirmed an internal custom silicon team on August 5, 2026
- The goal is to co-design hardware and models to cut Claude inference cost and latency
- Job listings cover silicon engineers paid 320,000 to 485,000 USD a year
- The company keeps its current partners, namely Trainium, TPU, NVIDIA cards and AMD chips
- The Information reports Anthropic held early talks with Samsung on manufacturing
What Anthropic actually announced
Business Insider first reported the move, and Anthropic confirmed it to TechCrunch. The new team will focus on custom silicon, meaning chips designed for the specific needs of the Claude models. It is the first time the company wants to help build hardware rather than only buy it.
We work from the chip level up with our silicon partners, and we are deepening that investment by building a custom silicon team.
— Anthropic spokesperson
Job listings describe silicon engineer roles paying 320,000 to 485,000 USD a year. They span front-end design, verification, physical design, analog and mixed-signal work, plus foundry relations and packaging. The company wants people who have already shipped silicon to production.
Anthropic stresses that it is not dropping its current suppliers. It still relies on Trainium accelerators (Amazon), TPU chips (Google), NVIDIA GPUs and AMD processors.
Why build silicon when partners exist
The reason is both technical and financial. Co-designing the chip and the model lets hardware fit Claude's exact workloads, which cuts response time and the cost per query. At the scale a growing customer base demands, even a few percent of savings means hundreds of millions of dollars a year.
Anthropic enters silicon with huge commitments to outside partners. In April 2026 it announced multi-gigawatt TPU capacity, and according to The Information it plans to spend about 200 billion USD on Google Cloud over five years. Broadcom disclosed 128 billion USD in purchase commitments through 2028, almost all tied to Google chips. An in-house chip is meant to ease that dependence over time.
How rivals compare
Anthropic starts later than its rivals. OpenAI unveiled its Jalapeño chip in June 2026, built with Broadcom and aimed at inference, meaning the cheap running of finished models. Google DeepMind has trained and served models on its own TPUs for years, and Meta is developing MTIA accelerators. For Anthropic, custom silicon is a way to avoid falling behind in a race where a hardware edge maps directly to model cost and speed.
| Company | Chip | Status |
|---|---|---|
| OpenAI | Jalapeño (with Broadcom) | Inference, since June 2026 |
| Google DeepMind | TPU | Training and inference, for years |
| Meta | MTIA | Accelerators in development |
| Anthropic | Custom silicon | Team being built, starting later |
Why it matters
The decision shows that for the biggest AI labs, cloud access alone is no longer enough. Control over silicon is becoming part of strategy, not just a line in the budget. If Anthropic tailors a chip to Claude's architecture, it can depend less on the prices and schedules of suppliers that are also its competitors. Yet this is a multi-year bet with real risk. Chip design is expensive, slow and easy to get wrong. Success will hinge on whether the company keeps the team together and takes a first project all the way to production.
What's next
- Silicon engineer hiring is already underway, spanning front-end design through foundry relations
- According to The Information, the company is in early talks with Samsung on manufacturing, with no formal deal yet
- A first in-house chip is years away, as design and validation typically take several years
