Oracle said at its AI World 2026 show in Las Vegas that Google TPUs on OCI will become available to cloud customers starting in 2027. The announcement marks the first time Google’s specialised AI chips will be offered in a hyperscale environment other than Google’s own. Oracle expects capacity to expand over several years and will use the TPUs for large language model training and post-training, as well as inference for reasoning and agentic workloads.
The move gives OCI customers access to hardware designed around Google’s own AI infrastructure while expanding Oracle’s effort to reduce its heavy reliance on Nvidia. OCI president Mahesh Thiagarajan said different AI workloads have different compute requirements, so adding Google TPUs gives customers flexibility to choose infrastructure based on performance, throughput and economics. Oracle has not confirmed how many TPUs will be available in 2027 or later.
Why Google TPUs on OCI matters for enterprise customers
Cloud buyers increasingly want alternatives to a single accelerator supplier. Models built for general-purpose GPUs can serve many workloads, but specialised silicon may offer better economics for particular jobs. By adding Google’s Tensor Processing Units, Oracle is bringing a platform closely associated with its own large-scale AI operations into a broader enterprise market.
Oracle’s case is not that one chip type should replace every other option. Instead, the company wants OCI to act as a selection layer for customers running training, fine-tuning, inference and high-performance computing. Nathan Thomas, Oracle senior vice president and product manager for OCI, said enterprises and HPC organisations are more comfortable when they have a wide range of options. He said Oracle works closely with suppliers it considers best in breed and does not want to dictate customer choices or close off alternatives.
Google released its eighth-generation TPUs in April, with the family divided for the first time into specialised components. The TPU 8t is aimed at training, while the TPU 8i targets inference. Oracle has not detailed the exact chip configurations, availability schedule or commercial terms for its OCI offering. Those details will matter to developers and infrastructure teams deciding whether to build new deployments around a non-Nvidia platform.
Oracle broadens its accelerator portfolio
The TPU announcement was part of a wider hardware strategy shown at Oracle AI World. The company also plans to deploy Cerebras hardware in OCI next year. Cerebras introduced its CS-4 rack-scale accelerator in August, built around three Wafer Scale Engine 3 Turbo chips. Cerebras says the system can deliver up to thirty times the tokens per second per user of GPUs, although that is a company claim rather than an independently established result across all workloads.
Oracle will also offer SambaNova’s fifth-generation SN50 Reconfigurable Dataflow Units. Released in February, these accelerators are intended to work alongside GPUs, with GPUs handling prefill and the RDUs handling decode. The mixed design reflects a broader industry trend: instead of assuming that one accelerator should handle every stage of an AI request, cloud platforms are separating tasks according to their computational and economic characteristics.
Oracle continues to offer Nvidia hardware and announced general availability of Nvidia Vera Rubin NVL72 systems and Vera CPUs. It also plans to introduce HGX Rubin NVL8 systems, although it did not specify when those systems would become available. This combination of new and existing products suggests that Oracle is broadening choice rather than making a clean break from the market leader.
The scale of Oracle’s bet is significant. It has signed agreements with Nvidia worth tens of billions of dollars to equip datacenters, including a Texas project that has been reported to be in difficulty. Adding Google and Cerebras hardware creates operational complexity, but it could give customers leverage when comparing long-term capacity, software support and cost. Oracle did not provide a capital figure specifically for the TPU deployment.
Revenue growth supports a larger infrastructure build
Oracle presented the accelerator expansion alongside a strengthening financial position. Sales rose from $57.4 billion in fiscal 2025 to $67.4 billion in fiscal 2026. OCI revenue reached $7.4 billion in the first quarter of fiscal 2027, an increase of 121 percent from a year earlier. Remaining performance obligations stood at $664 billion, up $209 billion over the previous year.
Those figures support Oracle’s argument that enterprise demand for cloud and AI services is becoming substantial. They also show why the company is willing to make unusually large capital commitments. Capital expenditure reached $55.7 billion in fiscal 2026 and could exceed $90 billion in the current fiscal year. That spending is being made amid public resistance to new datacentre construction and continuing debate about the pace of enterprise AI adoption.
Oracle’s infrastructure choices therefore carry strategic weight. Google TPUs can make OCI more relevant to customers that want access to Google-developed hardware without moving workloads into Google Cloud. Cerebras may appeal to customers prioritising high-speed inference, while SambaNova’s system provides another route for optimising model serving. Nvidia remains central to the broader platform, but Oracle is making clear that it does not want customers to see Nvidia as the only route to AI scale.
Security moves closer to enterprise data
Hardware choice is central to OCI’s growth, but Oracle used AI World to argue that infrastructure must also make AI applications safe enough for enterprise use. The company introduced its Deep Trust AI Architecture, which will be embedded in AI Database, Autonomous AI Database and APEX AI Application Generator. Its purpose is to protect sensitive data while AI applications use it and to control agent activity across organisational boundaries.
Oracle executive vice president Juan Loaiza and product management vice president Maria Colgan said trust should be incorporated from the start rather than added after an application is built. Placing enforcement in the underlying infrastructure can apply controls to both existing applications and new ones. The approach is particularly relevant as agents gain access to databases, enterprise resources and third-party systems that may hold different permissions.
The company is also expanding its OpenAI partnership by connecting models, ChatGPT and Codex with Oracle applications, data and infrastructure. Oracle AI Workspace is presented as one interface for accessing OpenAI frontier models alongside enterprise data and cloud services. Its ChatGPT plugins expose applications including Oracle Fusion, NetSuite, AI Database and OCI to agents and ChatGPT users.
Oracle rolled out 25 new agentic applications across its Fusion suites using the Fusion Claw execution runtime. The applications cover human resources, enterprise resource planning, supply chain, customer management and payroll. Oracle also introduced Fusion Companion and published a Git repository containing setup and configuration practices. Co-chief executive Mike Sicilia framed the strategy around making software largely disappear into the completion of work.
What enterprises should watch before 2027
Oracle has established the timing of the Google TPU launch, but several practical questions remain open. Customers will need clear information about regional availability, supported instance types, networking, data transfer charges, monitoring tools and software compatibility. They will also need evidence about throughput and cost for their own models rather than relying on broad performance comparisons from chip vendors.
The broader mix of accelerators raises another issue. Moving workloads between Nvidia, Google, Cerebras and SambaNova can require different software paths and operational skills. Oracle’s advantage will depend partly on making those differences manageable through consistent cloud services, orchestration and support. Capacity that expands gradually may also determine how quickly customers can replace established configurations. Whether Google TPUs on OCI arrive on schedule and at competitive cost, and whether Oracle can hold the accelerator mix open without alienating the Nvidia relationship that still underwrites the buildout, will decide how much this Google TPUs on OCI announcement changes the market.
For now, the central message is clear: Oracle is using its cloud growth and financial scale to widen the range of enterprise AI infrastructure it can offer. Google TPUs on OCI gives customers a notable new choice, while the planned launch of Google TPUs on OCI makes 2027 an important test of whether Oracle can turn accelerator diversity into a practical competitive advantage.