Google TPUs emerge as a major alternative to Nvidia chips

The TPU moves into the open

Google's Tensor Processing Units (TPUs) are transitioning from a private internal tool to a widely available merchant product. For years, these custom chips were primarily used within Google's own cloud infrastructure to power its services. Now, the hardware is being sold and rented to a broad range of external customers.

The Information reported that Oracle will offer Google's eighth-generation TPUs on its own cloud infrastructure starting in 2027. This makes Oracle the first major cloud provider outside of Google to host the hardware. This move turns the TPU into a competitive offering for large enterprises and AI labs that share customers with both companies.

This shift is significant because it opens the door for third-party providers to offer Google's silicon. While Google remains a primary source, the availability of TPUs on rival clouds expands the reach of the hardware. It allows companies to diversify their hardware sources while still accessing Google's specialized AI architecture.

Major labs join the TPU ecosystem

Several of the biggest names in artificial intelligence are now making large-scale commitments to Google's hardware. OpenAI, a long-time Nvidia customer, reportedly plans to buy Google's TPUs to power its models. The company even posted a job listing for an engineer specifically to lead the optimization of its frontier models on these chips.

Anthropic's recent filings provide a clear look at the scale of this shift, showing roughly $400 billion in compute commitments tied to TPUs. This represents about half of Anthropic's total compute commitments of nearly $800 billion. The company credits using a mix of different chips, including those from Nvidia, Google, and Amazon, as a key reason for its efficiency.

These moves suggest that frontier AI labs are no longer relying solely on one hardware provider. By diversifying into TPUs, these companies can potentially find better performance-per-watt or lower costs for specific tasks. It proves that a major AI lab can be built and scaled using Google's custom silicon as a primary foundation.

The shift toward AI inference

While training large models often gets the most attention, the next phase of the AI boom is focused on inference. Inference is the process of running AI models for everyday use by businesses and developers. Because this happens constantly, the cost per token and latency are the most important metrics for success.

Google responded to this demand by splitting its eighth-generation TPUs into two distinct versions. The 8t is designed specifically for training models, while the 8i is optimized for inference. This specialization helps Google target the market where volume is highest and where cost-efficiency is the primary driver for customers.

Every major chipmaker is now competing in this inference space to win over third-party developers. This includes Nvidia with its Blackwell architecture, AMD with its Helios racks, and Amazon with Trainium. Google's TPUs are now positioned as a serious contender in this high-volume market, competing for the same developer base.

Google's growing hardware business

Alphabet's recent third-quarter results show that the demand for TPUs is translating into significant growth for Google Cloud. Cloud revenue more than doubled to $30.7 billion, with operating income rising to $11.1 billion. Alphabet noted that TPU system sales were a meaningful contributor to this growth, though they did not provide a specific dollar amount.

Analysts are already projecting the financial impact of this hardware expansion. One estimate suggests that TPUs could bring in about $9.4 billion in the second half of this year. By 2028, some analysts believe the TPU business could reach $138 billion, representing about 34 percent of Google Cloud's revenue.

While these numbers are large, they are still a fraction of the total market dominated by Nvidia. However, the ambition for the TPU business has doubled in recent estimates. This suggests that Google is successfully positioning its chips as a primary alternative for the next generation of AI infrastructure.

Nvidia's response and the future

Nvidia remains the current leader in the AI chip market, but it is facing increased pressure from Google's expanding reach. Nvidia has historically maintained its lead through its CUDA software platform, which many developers use to run models. However, some reports suggest that OpenAI might share its own software tools to make it easier to run models on different types of chips.

If software becomes more portable across different hardware, it would weaken Nvidia's primary moat. This would make it easier for customers to adopt TPUs, AMD chips, and other alternatives without significant friction. A more aggressive Google could make the market more competitive, potentially benefiting the broader ecosystem by lowering costs.

The AI chip market is not yet a zero-sum game because demand is currently so high. Both Nvidia and Google can find room to grow as more companies build AI agents and reasoning models. For now, the TPU has become Google's strongest challenger in the race to power the next wave of AI applications.