OpenAI plans to buy Google TPUs directly from Alphabet to expand its computing capacity for training and running artificial intelligence models, The Information reported on Tuesday. The proposed hardware deal marks a significant shift for the ChatGPT creator, which has relied almost entirely on Nvidia processors to build and operate its systems.
Key points
- OpenAI is preparing a direct purchase of Google’s custom Tensor Processing Units to expand its training and inference capacity.
- The deal would reduce OpenAI’s reliance on Nvidia, which remains its primary hardware supplier and an investor.
- OpenAI began renting Google TPUs in 2025 and is recruiting engineers to adapt its frontier models to Google’s architecture.
- Google has started selling its custom chips directly to selected corporate buyers alongside cloud rentals.
Details of the planned purchase
The Information reported the planned transaction citing a person familiar with the matter. Financial terms, delivery schedules, and the specific volumes of chips involved have not been disclosed. Google did not immediately respond to requests for comment, while OpenAI declined to comment on the report.
The move represents an escalation of commercial relations between two companies that compete aggressively across consumer and enterprise markets. Both organisations offer rival flagship chatbots, developer platforms, and enterprise search tools. Despite that rivalry, surging demand for computational power has driven OpenAI to look beyond its traditional suppliers to secure hardware wherever possible.
OpenAI has made earlier moves to tap Google infrastructure. The company began renting Google TPUs through Google Cloud in 2025, according to reporting by Reuters and The Information. OpenAI also announced earlier plans to distribute its frontier models through Google Cloud, signalling an appetite to work with Alphabet when compute access or commercial distribution justifies it.
Why OpenAI wants Google TPUs
Securing sufficient silicon remains the central operational bottleneck for leading AI labs. OpenAI requires immense computing clusters to train frontier foundation models and run round-the-clock inference workloads for hundreds of millions of ChatGPT users. Relying on a single supplier has exposed developers to hardware backlogs and steep pricing.
Nvidia has served as OpenAI’s core hardware foundation since the research lab’s inception. Thousands of Nvidia graphics processing units, or GPUs, power OpenAI’s supercomputing clusters. Nvidia has also invested billions of dollars into OpenAI. Even so, the startup has sought alternatives to avoid single-vendor dependence and lower unit costs for running its models at scale.
Google designed its Tensor Processing Units specifically to accelerate neural network operations. While Nvidia GPUs are general-purpose processors adapted for machine learning, Google TPUs feature application-specific integrated circuit designs tuned to tensor mathematics. That specialised architecture can deliver high power efficiency and reduced inference latency for particular model types.
Adapting frontier models to new silicon
Transitioning models to a different chip family requires substantial engineering work. AI frameworks, kernels, and model pipelines built around Nvidia’s proprietary CUDA software stack cannot execute on Google TPUs without modification. Code must be adapted to Google’s compiler and software ecosystem, such as XLA and OpenXLA.
OpenAI has already begun assembling internal expertise to manage that transition. The company has posted vacancies to recruit an engineer dedicated to leading the optimisation of its frontier models on TPUs. The hire indicates that OpenAI is investing in long-term engineering pipelines to ensure its software runs natively across Google hardware rather than treating the chips as a temporary stopgap.
Engineering teams that support multiple silicon platforms gain bargaining power against hardware suppliers. By building portable software stacks, OpenAI can allocate workloads between Nvidia GPUs, Google TPUs, or future custom chips based purely on availability and price performance.
Google’s shifting hardware strategy
For most of the past decade, Google kept its TPUs exclusive to its own data centres and cloud customers. Companies that wanted access to Google hardware had to build inside Google Cloud Platform. Google used that exclusivity as a selling point to lure machine learning teams away from rival cloud providers.
That stance has changed this year. Google has begun selling its TPUs directly to selected corporate customers, allowing them to install the processors inside their own facilities or third-party infrastructure. The shift lets Google monetise its chip research outside the confines of Google Cloud revenue.
Google has also expanded distribution partnerships with competing cloud platforms. Google and Oracle recently agreed to a multi-year deal that will place Google TPUs inside Oracle Cloud Infrastructure starting in 2027. That arrangement allows Oracle customers to run workloads on Google’s custom processors without migrating data or services to Google Cloud.
What happens next
Alphabet is scheduled to report its quarterly earnings on Wednesday. Investors are watching for data on capital spending and signs that Google’s heavy internal investments in custom silicon and data centre capacity are yielding returns in both cloud and hardware sales.
For OpenAI, any direct delivery of Google TPUs will involve long deployment cycles. Data centres must prepare specialized cooling and networking architectures to support large clusters of the processors. Work on adapting OpenAI’s model codebases will run in parallel as the company prepares its infrastructure for multi-architecture training and inference.
