Oracle to Offer Google TPUs on Its Cloud Infrastructure Starting in 2027

Oracle is bringing Google’s Tensor Processing Units to its own cloud infrastructure in 2027, breaking a long-standing hardware exclusivity barrier. The agreement marks the first time Google has allowed a rival hyperscaler to host its custom artificial intelligence silicon inside competing data centers.[1][2]

The deployment will feature Google’s eighth-generation chips, specifically the TPU 8t designed for model training and the TPU 8i optimized for inference. Oracle plans to scale the offering over a multi-year period to meet surging enterprise demand for specialized compute.[1][2]

Oracle executives are framing the integration as a massive, permanent shift in their infrastructure strategy. Karan Batta, senior vice president of Oracle Cloud Infrastructure, described the deal as a “multi-year, multi-gigawatt partnership” that will mirror the native Google Cloud Platform experience.[2]

“What I can tell you is that this is not an experiment,” Batta told DatacenterDynamics. “This is going to be a cloud gigawatt-scale investment from Oracle over time, and the best way to think of it is that when Google launches a new TPU, we will have it on day one.”[2]

Google has spent more than a decade developing its TPUs as a highly efficient alternative to general-purpose graphics processing units. Until now, developers wanting to leverage that silicon had to move their data and workloads entirely into Google’s walled garden.[2]

Breaking the walled garden

The hardware arriving in Oracle facilities represents Google’s most advanced architecture. The TPU 8t delivers nearly three times the compute performance per pod compared to the previous generation, backed by 19.2 terabits per second of bidirectional scale-up bandwidth.[2]

To support that throughput, Google is introducing a new networking architecture called the Virgo Network. The design features 400 gigabits per second of scale-out networking bandwidth, utilizing high-radix switches that reduce network layers and support a four-fold increase in overall data center capacity.[2]

The inference-focused TPU 8i, meanwhile, is engineered with higher memory bandwidth to serve latency-sensitive applications. A single pod can scale to 1,152 chips, delivering 11.6 exaflops of FP8 compute performance and housing 331.8 terabytes of high-bandwidth memory.[2]

For Oracle, the addition of Google’s silicon is about giving enterprise customers leverage and choice. “Different AI workloads have different compute needs,” said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure.[2]

“Bringing Google TPUs to OCI gives customers access to multiple leading AI compute architectures in one cloud,” Thiagarajan said. “That choice gives them the flexibility to choose the right infrastructure for the performance, throughput, and economics of each AI workload.”[2]

The reality behind the announcement

Despite the aggressive marketing, this is an announcement of future intent rather than a capability shipping today. The hardware will not actually be available on Oracle Cloud Infrastructure until calendar year 2027, leaving a significant gap between the hype and the deployment.[1][4]

That delay highlights the profound logistical friction of multicloud hardware integration. Dropping Google’s proprietary Virgo Network and liquid-cooled TPU pods into Oracle’s existing data center footprint requires extensive physical retrofitting, not just a simple software update.[2]

Google’s motivation for opening its hardware ecosystem stems from the sheer cost of silicon research and development. To justify the billions spent designing the eighth-generation TPUs, Google needs to monetize the chips across a broader market share than its own cloud platform currently commands.[1][2]

“The next generation of AI will require a diverse range of compute,” said Amin Vahdat, Google AI and Infrastructure senior vice president and chief technologist.[1]

Oracle is not the only external partner gaining access to Google’s hardware. Google is simultaneously working with the investment firm Blackstone to create a United States-based cloud platform called Crux AI that will also offer TPU access.[2]

Depending on which project launches first, Oracle could be the first to bring the TPUs online outside of Google’s platform. Both initiatives signal a strategic pivot for Google, transforming its proprietary silicon from an exclusive cloud advantage into a distributed hardware product.[2]

The Switzerland of AI compute

The Google deal is part of a broader strategy to position Oracle Cloud Infrastructure as a neutral host for diverse hardware. Alongside the TPU announcement, Oracle confirmed it is adding Cerebras wafer-scale AI chips to a lineup that already relies heavily on Nvidia GPUs.[3]

Cerebras specializes in massive, single-wafer chips designed to train massive models faster than clustered GPUs. By offering Cerebras, Google, and Nvidia hardware under one roof, Oracle is attempting to commoditize the accelerator layer and compete directly with Amazon Web Services and Microsoft Azure on pure flexibility.[3]

Oracle is pairing this hardware neutrality with software flexibility. The company recently announced new integrations with OpenAI’s ChatGPT, allowing enterprise customers to route workloads through various foundational models while keeping their proprietary data secured within Oracle’s cloud boundaries.

Data gravity remains the underlying driver for these partnerships. Moving petabytes of corporate data from Oracle databases into Google Cloud simply to run a TPU training job incurs massive egress fees and latency penalties; bringing the TPU to the data solves the economic equation.[1][2]

The arrangement also provides Oracle customers with a vital negotiating lever against Nvidia. By offering a credible, high-performance alternative for both training and inference, developers can bypass Nvidia’s supply chain constraints and premium pricing models.[3]

As the artificial intelligence hardware market matures, the rigid boundaries between competing cloud providers are beginning to fracture. The next major bottleneck for developers will no longer be finding an available TPU, but securing the gigawatts of power required to turn it on.[2]