Anthropic's self-developed chip strategy: hiring the founder of Google's TPU business, a "bet against Nvidia" on the eve of a 2 trillion IPO.

In the critical window for a $2 trillion valuation IPO, Anthropic is shifting its strategic focus from "acquiring computing power" to "creating computing power," attempting to take control of the foundational infrastructure that supports its explosive growth. On Friday, the developer of Claude announced the official hiring of Amir Salek, the founder of Google's custom chip business, to establish a self-sufficient AI custom chip design team. At the same time, the company is actively advancing external computing power procurement: it has reached an initial procurement agreement worth approximately $250 million with British chip startup Fractile, expanded collaborations with Google and Broadcom to secure around 3.5GW of next-generation TPU capacity by 2027, and plans to co-design custom hardware with Samsung. This series of actions sends a clear signal: leading AI large model companies are accelerating their efforts to break free from absolute reliance on single chip suppliers like Nvidia. Through a dual-track strategy of "self-developed + diversified external procurement," Anthropic aims not only to alleviate the computing power supply bottleneck but also to achieve structural optimization in inference costswhere self-controlled computing power is becoming the core variable determining the valuation ceiling of large model companies. The acquisition of Amir Salek is a key step for Anthropic in entering the realm of autonomous semiconductor manufacturing. Salek worked at Nvidia for eight years, founding and leading the system-on-chip design team, and participating in the research and development of GPUs and Tegra chips; in 2013, he was recruited by Google to start the custom chip business from scratch, leading the development of the first seven generations of TPU chips, which remain a backbone of Googles AI infrastructure today. After leaving Google in 2022, Salek served as Senior Managing Director at Cerberus Capital Management, focusing on investments in the semiconductor and AI sectors. In his new position at Anthropic, Salek will report to James Bradbury and be responsible for establishing and leading an AI custom chip design team. In addition to self-development, Anthropic also plans to collaborate with Samsung to design hardware and models to further enhance operational speed and efficiency. This layout highly aligns with industry trends. Competitor OpenAI has developed the Jalapeno chip in collaboration with Broadcom specifically designed for inference loads, Google DeepMind has long relied on its own TPU, while Meta is advancing its MTIA accelerator. By developing its own chips, Anthropic can precisely adjust hardware design based on its own model architecture, establishing barriers at the foundational computing power level and reducing reliance on external generic chips. With billions of dollars "hoarding cards," Anthropic is building a computing power network that does not rely on a single supplier. Before its self-developed chip is realized, Anthropic has shown strong aggressiveness in expanding external computing power. On the chip procurement side, in addition to the initial $250 million order from Fractile, Anthropic is also significantly deepening cooperation with established giantsexpanding collaborations with Google and Broadcom to secure around 3.5GW of next-generation TPU capacity by 2027. This figure is approximately equivalent to the electricity consumption of a medium-sized city and does not include the 1GW capacity committed for delivery in 2026 in Google Cloud's October agreement last year. Broadcom plays a key role in this: it is responsible for developing and supplying custom TPUs and has committed to supplying network components for Googles next-generation AI racks until 2031. On the data center side, Anthropics contracted scale has reached the level of several tens of billions of dollars, with partners including traditional cloud vendors, chip giants, and emerging computing power suppliers, including capacity agreements signed with Riot Platforms and Volta Infra Holdings. It is noteworthy that the market estimates that over half of its computing power is long-term supported by TPUs on Google Cloud. Given the backdrop of slowing overall computing power growth at Google itself, Anthropic has to accelerate "finding cards" externally. The accounting behind autonomous computing power: solidifying the physical foundation for a $2 trillion IPO. Anthropic's aggressive investments in chips and computing power directly correspond to its rapidly improving financial model. As of the end of July, the company's annualized revenue has surpassed $65 billion, with initial revenue exceeding $11.5 billion in the second quarter, growth of over 14 times year-on-year; the gross margin of inference infrastructure surged from 38% to over 70%, and adjusted operating profit turned positive for the first time. Chief Financial Officer Krishna Rao referred to this as the companys "most important investment in computing to date," aimed at meeting the exponential growth of its customer base. For Anthropic, which is about to enter the public market with a valuation of $2 trillion or higher, controlling computing power pricing and supply chain autonomy is not just about costs; it is also about whether it can deliver on that steep growth curve to investors in the second half of AI.