OpenAI plans to purchase Google TPUs to reduce reliance on Nvidia AI chips

OpenAI is reportedly diversifying its AI infrastructure by planning to purchase Google's Tensor Processing Units (TPUs). This move signals a strategic shift for the ChatGPT developer as it seeks to reduce its heavy reliance on Nvidia's dominant GPU ecosystem to secure the massive computing power required for next-generation models.

While neither company has officially confirmed the arrangement, the reported plan involves a transition from renting to owning hardware. OpenAI began renting Google's TPUs in June 2025, though the company initially described those efforts as early testing rather than a large-scale deployment. A shift toward direct purchasing would allow OpenAI to own the silicon rather than paying for hourly cloud access. This transition is supported by a recent OpenAI job listing for an engineer to lead the optimization of frontier models on TPUs, a critical step since switching chip architectures requires complex software compilers to translate AI models into instructions the specific hardware can execute. This effort is further bolstered by the presence of Richard Ho, OpenAI's hardware program lead, who previously spent nearly nine years at Google leading its Cloud TPU project.

The Diversification Strategy

This move toward Google silicon does not suggest an immediate exit from the Nvidia ecosystem, but rather the implementation of a multi-vendor strategy. OpenAI remains one of Nvidia's largest customers, with current and planned commitments totaling approximately 12 gigawatts of computing power, potentially rising to 16 gigawatts. The financial scale of this relationship is immense, with some estimates suggesting a total spend of $600 billion through 2030. Furthermore, Nvidia has provided up to $105 billion in credit support for an OpenAI data center campus in Ohio, a facility that is largely restricted to Nvidia hardware.

To further insulate itself from supply chain bottlenecks and pricing pressure, OpenAI is pursuing a three-pronged approach to hardware:

• Continued large-scale procurement of Nvidia GPUs.
• The reported purchase of Google TPUs.
• The development of in-house custom silicon.

The third pillar of this strategy involves "Jalapeño," a custom inference chip designed in collaboration with Broadcom. Announced in June 2026, Jalapeño is intended for deployment by the end of 2026. OpenAI has also explored options with AMD, granting the company the right to buy up to 160 million shares in exchange for the deployment of 6 gigawatts of AMD chips.

Market Shifts and Infrastructure Demand

For Google, the reported deal validates the commercial viability of its TPU business. Alphabet has recently expanded the sale of its processors to external customers, moving beyond using them exclusively for internal workloads. Google's current offerings include the eighth-generation TPU 8t for training and the TPU 8i for inference. Alphabet's finance chief, Anat Ashkenazi, has indicated that while some revenue from TPU sales will be recognized by the end of 2026, the bulk of that financial contribution is expected in 2027.

This shift occurs against a backdrop of staggering infrastructure spending. Alphabet, Amazon, and Microsoft are collectively expected to spend as much as $600 billion on AI infrastructure in 2026 alone. While Nvidia continues to dominate the market, its growth has been hampered by supply constraints. Nvidia management has noted that demand continues to far exceed its production capacity, a bottleneck expected to persist through the end of fiscal 2028.

The broader implication of OpenAI's move is a gradual shift in leverage. By establishing credible alternatives to Nvidia, OpenAI gains significant protection against supply shortages and increases its bargaining power during price negotiations. As other major players, such as Anthropic, also secure multi-gigawatt TPU capacity from Google and Broadcom, the industry is moving toward a fragmented hardware landscape. The success of this transition will likely depend on whether TPU software can close the gap with Nvidia's established ecosystem, potentially ending the era of single-vendor dominance in AI compute.