Google targets NVIDIA customers with aggressive TPU sales push; signals major shift in AI chip wars.

Alphabet's decision to sell its Tensor Processing Units directly to external customers represents a fundamental shift in the company's strategy. On April 29, CEO Sundar Pichai announced that Google will begin selling TPUs directly to select customers' data centers, transforming Google from a cloud landlord into a full-blown chip merchant. The move targets some of Nvidia's most valuable relationships.

Google's ambitions are substantial. The company plans to ship 4.3 million TPUs in 2026, ramping to more than 35 million units by 2028—roughly an eightfold increase in two years. Morgan Stanley estimates that selling 500,000 TPU chips alone could generate approximately $13 billion in revenue for Google by 2027, with most revenue expected to land in 2027 and initial shipments to customer data centers beginning within 2026.

Major AI companies are already committing to the chips. Anthropic, the AI safety lab behind Claude, has reportedly committed to deploying up to one million TPUs and investing in approximately 3.5 gigawatts of power capacity starting in 2027. Meta is also said to be in discussions about substantial TPU commitments.

Google's internal groundwork for this commercial move stretches back years. The company has hoarded TPUs for internal use since the mid-2010s, and in 2022 reorganized to give Google Cloud greater control over TPU allocation. These chips have powered internal workloads including Search, YouTube recommendations, and Gemini model training—giving Google proven expertise competitors lack.

While Google isn't alone in developing alternatives to Nvidia, it occupies a unique position. Amazon has its Trainium and Inferentia chips, and Microsoft has its Maia accelerator, but Google's move differs fundamentally: the company is willing to let outsiders buy the silicon rather than just rent access through Google Cloud, putting it in direct competition with Nvidia on hardware sales rather than just cloud services. The scale of Google's deployment plans—jumping from 4.3 million to 35 million TPUs in two years—suggests the company believes the AI accelerator market is far larger than any single vendor can serve.