Anthropic Gets Ready to Build its Own Chips – Moby THE GIST
Our analysts just identified a stock with the potential to be the next Nvidia. Tell us how you invest and we’ll show you why it’s our #1 pick. Tap here.
In the dog-eat-dog world of AI, every company needs an edge. That edge comes down to one thing: Network, which bleeds into every LLM must-have, from energy, compute, to employees, and perhaps most importantly, chips.
Anthropic, whose founders left OpenAI in late 2020 and early 2021 to start the company, knows that better than anybody. So it’s no surprise CEO Dario Amodei is deploying major talent and capital for its new in-house chips team, forcing everyone from Elon Musk and Nvidia to OpenAI and Broadcom to choose sides.
WHAT HAPPENED
Anthropic officially confirmed that it’s building its own in-house “custom silicon team,” according to an Anthropic spokesperson who spoke to Business Insider and a publicly posted job listing.
We can assume that because Anthropic is assembling its own in-house silicon team to design custom chips, they’re talking about Application-Specific Integrated Circuits, or ASICs. ASICs are purpose-built for better performance-per-watt and lower cost per FLOP — the raw volume of mathematical work, i.e., computing power. These ASIC chips are going to bring costs down for Anthropic’s everyday driver models to help them compete with OpenAI’s GPT-5.6 family and Chinese open-weight models like Qwen, DeepSeek, and GLM.
The AI company, which boasts an official post-money valuation of $965 billion following its Series H funding round, joins Google, Amazon, Microsoft, Meta, and OpenAI, the latter of which recently announced the Broadcom-built “Jalapeño” inference chip in June. The chip has yet to be deployed to any data centers.
Anthropic joining the ranks of these hyperscalers doesn’t mean they are ditching their chips altogether, though. They will continue to utilize Amazon’s AWS and Google Cloud while leveraging Nvidia and AMD chips for training and “frontier model discovery.”
Samsung is rumored to be in early talks with Anthropic for manufacturing, per The Information.
At the same time, Clive Chan, poached from OpenAI’s custom-chip team, has extensive experience working with, but not for, Broadcom, which continues to expand its market share in ASIC dominance.
One stock. Nvidia-level potential. 30M+ investors trust Moby to find it first. Get the pick. Tap here.
WHY IT MATTERS
The clue that Anthropic may end up working solely with Broadcom for their ASIC chips is Chan himself. He was the “second hardware hire” on OpenAI’s custom-chip Jalapeño program and also worked with Tesla Dojo, Google, and SpaceX. Note how we didn’t say Chan used to work at Broadcom, but instead worked on OpenAI’s chip program, which Broadcom co-designed and built. That puts Chan in the room while building out OpenAI’s first real attempt at custom silicon alongside Broadcom’s design team on “Jalapeño.” Chan isn’t bringing OpenAI’s designs, any of Broadcom’s IP, or even a relationship with Broadcom itself — he’s bringing experience and everything he learned on the job.
All Amodei has to do is connect his company with the fabless semiconductor company to take the next step to get their ASIC chips in production, further expanding Broadcom’s already dominant market share of 70%, per Hashrate Index. This is up from the 60–80% range, as Bloomberg Intelligence noted earlier this year.
It’s a win-win for Broadcom, but not necessarily great for Musk and his partnership with Nvidia, despite Jensen Huang and co continuing to provide GPUs to Amodei for training.
When Amodei, Chan, and their new in-house chip design team complete their ASIC chip, it means Claude is going to have a cheap, powerful, more efficient daily driver, given how well Fable 5 as well as Mythos 5 have been doing across numerous benchmarks the past year.
Where does that leave Grok? Today, Elon Musk’s prized model isn’t even in the top 20 on the AI model rankings on OpenRouter.ai. Nvidia will happily keep selling to Musk until the money runs out, but the real problem is simpler: Musk is nuking capital on a model that actual AI researchers won’t touch, while his competitors are building custom ASICs to ship better, user-specific models that actually work.
WHAT’S NEXT
We should expect OpenAI and Anthropic to drop their custom ASIC chip over the next year. Businesses already locked into Anthropic’s ecosystem via Claude are eagerly awaiting even more performance, further boosting Anthropic’s recurring revenue.
We will likely see an explosion of cheap inference, perfect for Agentic AI workflows and lead AI users in the U.S., suddenly given a domestic option instead of Chinese open-weight models. If that happens, the federal government may be able to let out a sigh of relief.