Google is expanding its custom AI chip business by promoting Tensor Processing Units (TPUs) to external customers with financial guarantees, manufacturing arrangements, and software tools, positioning TPUs to compete with NVIDIA in the data‑center market.

The strategy blends chip design, cloud services, software support, and project financing to place TPUs in more data centers and attract external customers into Google’s AI compute ecosystem.

TPU shipments are forecast to reach about 4.3 million units in 2026 and could climb toward 35 million by 2028, though these are projections rather than confirmed sales.

A broader $35 billion infrastructure deal involving Broadcom, Anthropic, Apollo Global Management, and Blackstone is linked to five U.S. data centers using Google chips, though commercial terms remain largely undisclosed.

Google has expanded its production network, reportedly ordering more than three million TPUs from Intel for 2028 and exploring new designs with Marvell Technology to boost capacity.

Google provided a $3.2 billion financial guarantee for the Lake Mariner data center campus in western New York, a project that will host TPU-based computing systems and supply capacity to Fluidstack for Anthropic’s Claude AI models.

Google introduced TorchTPU to help PyTorch developers run workloads on TPUs, reducing effort to migrate tasks from NVIDIA systems.

Other major players like Amazon, Microsoft, and Meta are building their own AI chips, highlighting competitive dynamics as many still rely on NVIDIA GPUs for AI workloads.

Beyond chip design, Google is financing TPU deployments, taking on a financing role similar to NVIDIA’s support for large computing clusters.