OpenAI’s reported plan to buy Tensor Processing Units (TPUs) from Google could give the AI company another source of computing capacity and greater leverage with chip suppliers, without replacing its reliance on Nvidia GPUs.
For data center operators, a multichip strategy could improve compute availability and bargaining power, but software porting, performance tuning and infrastructure costs will determine whether it pays off.
“Compute is the constraint for OpenAI and labs like it,” said Matt Kimball, vice president and principal analyst for data center technologies at Moor Insights & Strategy. “It’s all about how fast it trains, what it can afford to serve. And then how much leverage it has when negotiating.”
The Information reported the potential purchase on Tuesday, citing anonymous sources familiar with the matter. Neither Google nor OpenAI has confirmed the arrangement.
Kimball said the potential deal would extend OpenAI’s existing strategy of buying Nvidia GPUs, developing custom silicon with Broadcom and adding another accelerator architecture. Renting TPUs through Google Cloud provides a way to test the technology, while purchasing hardware could give OpenAI more control over where capacity is deployed and how it is used.
Software could be the biggest hurdle
Google uses TPUs to train and serve its Gemini models, and other AI developers also use the accelerators for large-scale workloads. Their ability to support frontier AI workloads is established, Kimball said. Adapting OpenAI’s software stack to achieve competitive performance is the harder task.
OpenAI has built much of its infrastructure around PyTorch, Nvidia’s CUDA platform and Triton, its GPU programming language. Google’s TPU ecosystem relies more heavily on JAX and the XLA compiler. Running a model on another architecture can require specialized work on kernels, memory use and distributed execution.
“Getting a frontier model to run on a TPU is one thing,” Kimball said. “Getting it to run with tuned kernels for attention, mixture-of-experts routing and long-context inference, is months of specialized work.”
The cost of that engineering, along with performance differences, would factor into whether TPUs offer a financial advantage over Nvidia GPUs.
OpenAI’s familiarity with Google’s technology could ease the work. Kimball noted that OpenAI’s chip team is led by an executive who helped build TPUs at Google. That familiarity could help, though it would not eliminate the need to optimize workloads for the new architecture.
Inference might offer the first savings
Kimball expects the earliest economic benefits to emerge in inference, the computing required to generate AI responses. A purpose-built accelerator can improve the cost of serving workloads, but the result depends on throughput, latency, utilization, memory requirements and operating costs.
A TPU deployment could also provide another route to capacity as demand for AI computing grows. But it would not free OpenAI from the semiconductor industry’s supply constraints. GPUs and TPUs both depend on advanced manufacturing, packaging, high-bandwidth memory and data center power.
Adding a supplier gives OpenAI another potential source of hardware, rather than an escape from those bottlenecks. Kimball said the ability to negotiate with a credible alternative supplier could be valuable even if Nvidia remains OpenAI’s primary accelerator platform.
Google expands its TPU market
Oracle announced Tuesday that it plans to offer Google TPUs through Oracle Cloud Infrastructure, with deployments beginning in 2027. The agreement will make OCI the first hyperscale cloud outside Google to offer the accelerators.
“The next generation of AI will require a diverse range of compute,” said Amin Vahdat, Google’s senior vice president and chief technologist for AI and Infrastructure, in the announcement. “TPUs give customers a powerful, purpose-built option for demanding training and inference workloads.”
The deal reflects Google’s effort to expand TPU access beyond its own cloud. For customers, offering the chips through OCI could provide another accelerator option without requiring them to operate their own hardware.
A potential OpenAI purchase would extend that strategy to a company competing with Google in AI models and services, although the reported transaction remains unconfirmed.
Kimball does not see the development as a direct threat to Nvidia. He pointed to Nvidia’s software ecosystem, broad availability through cloud providers and hardware vendors, and roadmap serving customers from enterprises to sovereign AI programs.
“Demand continues to soar,” he said. “And it’s growing faster than any single supplier can support.”
Multiarchitecture computing needs better tools
Running several accelerator architectures can complicate infrastructure operations. Different programming environments and performance characteristics make workloads harder to move and tune across hardware.
Kimball said broader adoption will require orchestration and control-plane tools that automate deployment and optimization across architectures. Without them, the engineering overhead could erode the benefits of adding another chip supplier.
For OpenAI, the test will be whether TPUs deliver useful capacity and competitive workload economics after software optimization and operating costs are included. For infrastructure teams considering a similar strategy, hardware choice is only one part of the calculation.
Shane Snider is a senior news writer at TechTarget, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. You can reach Shane at [email protected] or on LinkedIn.
