Google's TPU Father, Jumping Ship to Anthropic?
Early this morning, the AI world was flooded with this news.
Foreign media revealed a bombshell: Anthropic has successfully poached former Google executive and the famous "TPU Father," Amir Salek!
It seems Anthropic is going to make its own chips.
Previously, Anthropic had already poached key self-developed chip talent from OpenAI.
Clearly, the real battle for AGI has now shifted from the software layer to the chips themselves!
TPU Father + OpenAI Defector, This Lineup Is Frightening
To self-develop chips, the core element is always the "genius minds."
Anthropic's poaching this time is a precise strike.
In the semiconductor and AI computing circles, the name Amir Salek is well-known.
He is the renowned father of Google's TPU, the core soul of the project for a full 9 years.
Before joining Google, Salek worked under Jensen Huang, leading Nvidia's SoC design team, deeply versed in the underlying logic of GPU and complex chip architecture.
He joined Google in 2013. Over these 9 years, he not only built Google's custom silicon team from scratch but also led the delivery of the first seven generations of TPUs in one go!
It's important to know that the TPU was Google's "computing power artifact" specifically built for AI machine learning to break free from Nvidia dependency.
It was precisely because of the TPU that Google gained the confidence to challenge Nvidia.
After leaving Google in 2022, Salek wasn't idle, joining the top global private equity firm Cerberus Capital Management as a Senior Managing Director.
This gave him expertise not just in technology but also in capital operations and the commercial expansion of enterprise chip projects.
Now, this "TPU Father," a blend of technology, management, and business acumen, officially reports to James Bradbury, Anthropic's Head of Compute.
Moreover, Anthropic didn't just poach from Google's garden; they also raided OpenAI's deck.
Just this June, Clive Chan quietly jumped ship to Anthropic.
He was the second hardware employee on OpenAI's self-developed chip team and an early core member of OpenAI's highly confidential in-house chip project (codenamed Jalapeno).
Now, Anthropic's chip team perfectly combines the "founding lead of Google TPU" + "core brain of OpenAI's in-house chip."
Coincidentally, this month Anthropic just announced the formal establishment of its internal Custom Silicon department. It seems they are ready to flip the table.
Why Build Your Own Chips?
So, having raised so much money and backed by Amazon AWS and Google Cloud, why go through the trouble of building your own chips? Isn't it easier to just buy Nvidia's H100?
The answer is: You can afford to buy them, but you can't afford the cost; it's better to become self-reliant than be dependent.
Jensen Huang's GPUs are great, but the machines are expensive and in short supply.
It is reported that Anthropic's compute spending for 2026 is estimated at $19 billion, and its inference costs in 2025 exceeded budget by 23%. The company's gross margin in 2025 was 40%, showing a significant gap from the 77% gross margin target it informed investors it aimed to reach ahead of its anticipated Nasdaq listing.
For a giant like Anthropic aiming for AGI, the "Nvidia tax" should be minimized as much as possible.
Building your own chips is the true way.
Moreover, general-purpose GPUs are not tailor-made for Claude.
Apple's M chips can outperform a host of traditional PC processors precisely because of software-hardware co-design.
Anthropic wants to do the same for Claude.
If those who understand Claude's large model algorithms sit together with those designing the underlying chips, it enables deep integration of "model + chip."
Future versions of Claude could achieve significantly higher inference speeds and intelligence levels at much lower costs.
Furthermore, with an internal chip division, Anthropic wouldn't have to fight for scarce data center resources and wait in line.
This allows them to not only customize hardware according to their needs but also effectively alleviate supply chain bottlenecks.
Not Just Chasing Silicon
Anthropic's "Infrastructure Frenzy" Blueprint
Moreover, while recruiting Amir Salek, Anthropic is conducting an astonishing global "compute infrastructure shopping spree."
First, their primary preparation is to stabilize existing giants.
Anthropic has not severed ties with current suppliers. Reports indicate they are still purchasing large quantities of chips from Nvidia, Amazon, and Google.
The second preparation is heavy investment in external startups, buying future production capacity.
Anthropic has also made an exceptionally heavy investment in external chip companies.
Recently, they issued a substantial initial commitment order worth $250 million to a UK AI chip startup called Fractile.
It is reported that Fractile is developing a novel hardware architecture, claiming to enable LLM inference with extremely low power consumption.
Anthropic's move is akin to both building their own chips and securing the production capacity of promising third-party chip startups.
The third preparation is aggressively acquiring land, scrambling for data centers and power.
Recently, Anthropic quietly reached a series of infrastructure agreements, with partners even including Riot Platforms Inc. (one of the largest mining companies in the US) and Volta Infra Holdings Ltd.
The cyclical nature of a certain market has left many mining farms seeking transformation.
Anthropic is entering the fray, snapping up assets aggressively.
Clearly, they are building their own exclusive compute fortress with an aggressive posture.
Now, the competition among major model giants has completely changed.
It used to be about algorithms and data, later about fundraising.
Now, the flames of war have spread to chip manufacturing.
Whoever can first complete the loop of "model architecture + underlying silicon" will gain a cliff-like lead in inference costs.
Now, OpenAI has its Jalapeno chip, and Anthropic has a Claude-dedicated TPU led by Amir Salek.
Whoever controls the compute power will define AGI.
References:
https://www.bloomberg.com/news/articles/2026-08-21/anthropic-taps-google-chip-veteran-as-part-of-push-into-hardware
https://www.techtimes.com/articles/325089/20260820/anthropic-bet-250m-unshipped-silicon-why-fractile-earned-pre-revenue-contract.htm
This article is from the WeChat public account "New Zhiyuan", author: ASI Revelation, editor: Aeneas
