Anthropic, an artificial intelligence company, is advancing its largest-ever self-built computing power expansion plan. The company plans to directly purchase nearly 1 million of Google's latest TPU v7 series AI chips from Broadcom, a chip giant, with a total transaction value estimated at up to $21 billion. This move marks a crucial step for Anthropic in the construction of its computing power infrastructure, aiming to deploy powerful dedicated computing power for its own data center system.

Procurement Model and Advantages

The procurement adopts a "purchase + lease" combination model, with 400,000 chips to be directly supplied by Broadcom as rack-level AI systems, and the remaining 600,000 chips to be leased through the Google Cloud Platform (GCP). The overall computing power cluster is expected to have a power consumption of over 1GW, equivalent to the power supply scale of a medium-sized nuclear power plant.

From a technical perspective, the TPU v7 chips demonstrate significant cost advantages and competitive performance. Compared to similar products from Nvidia, the TPU v7 has already surpassed Nvidia's H200 in performance, with some indicators also approaching Nvidia's B200 products. In addition, the TPU v7 provides Anthropic with a lower total cost of ownership (TCO) per effective PFLOP, making it more economical for large-scale model training and inference tasks.

Supply Chain Adjustment and Cooperation

The core of this procurement lies in a significant adjustment of the supply chain path. Broadcom will deliver a complete set of AI computing equipment based on the TPU v7p "Ironwood" to Anthropic in the form of rack-level systems. This means that Anthropic's computing power supply chain will no longer go through Google's cloud service system, but will directly connect with the chip manufacturer. However, as a co-developer of the TPU intellectual property (IP), Google is still expected to earn a considerable licensing income from this transaction.

In Anthropic's own TPU computing power system, three companies including TeraWulf are responsible for supplying the infrastructure, while Fluidstack will be in charge of on-site deployment services. This cooperation model not only improves deployment efficiency but also ensures Anthropic's direct control over computing power resources.

Future Outlook

Anthropic's large-scale procurement of TPU v7 chips is an important strategic layout in the field of AI computing power. By directly connecting with chip manufacturers, Anthropic can not only obtain stronger computing power support but also achieve significant advantages in cost control and performance optimization. With the widespread application of TPU v7 chips, Anthropic is expected to make greater breakthroughs in large model training and inference tasks, further consolidating its leading position in the AI field. In addition, this transaction also provides new ideas and references for other AI companies. In the future, as AI technology continues to develop, the demand for computing power will continue to grow, and companies may pay more attention to the autonomous control and flexible configuration of computing power resources.

Conevo Integrated Circuit Distributor

Conevo, the distributor of chips, focuses on providing efficient and reliable semiconductor solutions to global customers. Through its professional team and advanced supply chain management system, Conevo ensures that customers can quickly obtain the required IC components and receive continuous technical support. The popular IC components of Conevo include, but are not limited to:

● ST62T25CM6 is a high-performance microcontroller, suitable for various industrial and consumer electronic applications, with powerful processing capabilities and low power consumption features.

● LM5009SDC/NOPB is an efficient power management chip, widely used in power conversion and voltage stabilizing circuits, known for its high efficiency and stability.

● W77E516A40PL is a powerful embedded microcontroller, suitable for smart cards and IoT devices, with high integration and low power consumption characteristics.

Website: www.conevoelec.com

Email: info@conevoelec.com