The AI computing landscape is undergoing a major transformation. According to a recent report by the semiconductor research firm SemiAnalysis, AI company Anthropic has reached a significant partnership with Broadcom, directly purchasing nearly 1 million TPU v7p "Ironwood" AI chips to deploy large-scale AI training and inference infrastructure in its own data centers. This move means that Anthropic has chosen to bypass Google, the original developer of TPU, and directly obtain complete system solutions from chip manufacturer Broadcom, allowing Google only to participate as an IP licensing provider in the transaction.

 Behind the $21 Billion Order: Building AI Computing Independent of Cloud Giants

Broadcom CEO Poonam Chaudhary confirmed in December 2025 that Anthropic had placed a total order of $21 billion (approximately 147.2 billion RMB), covering rack-level AI systems based on TPU v7p. These systems will be supported by infrastructure provided by companies such as TeraWulf, with Fluidstack responsible for on-site deployment and operations, forming a fully self-controlled computing loop by Anthropic.

Although Google remains the intellectual property holder of the TPU architecture and will receive IP licensing fees from the transaction, it no longer acts as a hardware supplier or cloud service intermediary. This arrangement significantly reduces Anthropic's reliance on Google Cloud, enhancing its independence in model training, data security, and cost control.

 Why Bypass Google? Security, Cost, and Strategic Autonomy Are Key

Analysts believe that Anthropic's decision stems from three considerations:

- Data Sovereignty: The Claude series models involve a large amount of enterprise-level sensitive data, and on-premises deployment can avoid data being handled by third-party cloud platforms;

- Cost Optimization: Directly purchasing complete system solutions can avoid cloud service premiums, significantly reducing long-term operational costs compared to renting TPU Pods;

- Technical Autonomy: Gaining control over the underlying computing stack allows for deep customization of chip drivers, communication protocols, and scheduling systems, improving training efficiency.

This move also reflects that leading AI companies are accelerating their "de-clouding" efforts, shifting from "renting computing power" to "building computing power" to gain control over infrastructure in the AGI competition.

 TPU Ecosystem Fragmentation: Is Broadcom the Biggest Winner?

For Broadcom, this collaboration marks its successful upgrade from a "chip supplier" to an "AI system integrator." Previously, the TPU ecosystem was entirely controlled by Google. Now, Broadcom, leveraging advanced packaging and system integration capabilities, transforms Google's IP into standardized products available for external sales, opening up a new revenue stream.

For Google, although it has lost opportunities for hardware sales and cloud integration, the IP licensing model still brings stable income and expands the industry influence of the TPU architecture—though its control is greatly reduced.

 AIbase Observations: The War for Computing Sovereignty Has Begun

Anthropic's multi-billion-dollar purchase is not just a business decision but also a concentrated expression of the "computing sovereignty" awareness in the AI era. When large models become national strategic assets, whoever controls computing power holds the initiative in AI development.