Google Invests $40 Billion in Anthropic

Google will invest up to 400 billion US dollars in Anthropic, further deepening the cooperation between the two companies. Both are partners and competitors in the artificial intelligence race.

Anthropic said on Friday that Google has committed to investing 10 billion US dollars in cash immediately at the latest valuation of 380 billion US dollars. Google will add an additional investment of 30 billion US dollars after Anthropic achieves performance milestones, while greatly expanding Anthropic's computing power scale.

Anthropic is an important customer for Google's chips and cloud services, and Google is vigorously expanding these businesses to make up for the maturity of its core revenue source - search advertising business growth. Google Cloud will provide 5 gigawatts of computing power for Anthropic in the next five years, starting from 2027, with the possibility of adding several gigawatts later. This agreement is a further expansion of the cooperation agreement signed by Anthropic, Google, and Broadcom earlier this month.

Google's TPU is one of the most competitive alternatives to Nvidia chips. In this industry with extremely large demand for computing power, TPU is a scarce and precious resource for Anthropic and other AI developers.

The End of the Trinity and the Formation of a Dual Confrontation

This investment marks a fundamental restructuring of the AI industry competition landscape.

In the past two years, the first echelon of AI has been defined as the "Three Majors" - OpenAI, Google, and Anthropic, a tripartite stand. Now, this narrative has ended. If we list Anthropic's financing statements in the past six months, we will find an intriguing reality:

• Amazon: 5 billion US dollars in cash, up to 25 billion US dollars, plus 5 gigawatts of Trainium computing power and a 100 billion US dollars AWS procurement contract;

• Google: 10 billion US dollars in cash, up to 40 billion US dollars, plus 5 gigawatts of TPU computing power;

• Nvidia: up to 10 billion US dollars, 1 gigawatt of GPU supply;

• Microsoft: up to 5 billion US dollars, Anthropic procures 30 billion US dollars of computing power from Azure.

The four top Silicon Valley players all appear on Anthropic's shareholder roster. The total computing power commitment exceeds 11 gigawatts - equivalent to the power generation of 10 nuclear power plants.

The pattern has evolved from "three shares of the world" to a dual confrontation between the Anthropic camp and OpenAI. The term "Three Majors" is now outdated.

The Core Division: Big Models + ASIC or Big Models + GPU

The reorganization reveals a clearer technological route division behind the scenes.

Anthropic follows the "big models + ASIC" route. Google's TPU and Amazon's Trainium are custom-designed chips (ASIC) specifically for AI workloads. Anthropic's future computing power base will be mainly supported by these two ASIC systems. One of the reasons Google is willing to lock in Anthropic with 40 billion is because Google's capital expenditure plan for this year is as high as 185 billion US dollars, with a large amount of money invested in data centers and TPU production capacity. If there are no big customers to absorb TPU, it will be the most expensive inventory. Anthropic is not only an entry point for enterprise customers but also the best ballast stone for Google's TPU production capacity.

OpenAI, on the other hand, follows the "big models + GPU" route. OpenAI's core computing power comes from the Stargate project deeply bound with Nvidia - a massive infrastructure plan with a target scale of 500 billion US dollars, with Nvidia GPU as the computing power base. The advantage of this route is that Nvidia's GPU ecosystem is mature and has a complete software stack; however, the problem lies in the long implementation cycle. It is estimated that Stargate will not be fully operational until around 2029, and the physical progress of the first Texas data center is still slow so far.

The underlying logic of the two routes is completely different. The advantage of the ASIC route is high energy efficiency ratio and lower unit computing power cost, but it