Oracle has made a move that may redefine its position in the AI field. The company has officially announced that it will deploy up to 50,000 AMD Instinct MI450 GPUs in its Oracle Cloud Infrastructure (OCI) cloud. This gigantic purchase was confirmed by Larry Ellison during the Oracle CloudWorld 20 25 event in Las Vegas in October. The deployment will begin in the third quarter of 2026 and will deliver a new Oracle AI supercluster built on AMD’s Helios architecture.
The project aims to significantly strengthen OCI’s capacity to train and run large-scale AI models. With this move , the Oracle AI supercluster enters into a direct competition with tech giants such as Amazon, Microsoft and Google, which already run huge AI data centers with accelerators from Nvidia.
Who is Oracle and why does it need an AI supercluster ?
Oracle is one of the largest providers of enterprise database solutions, cloud services and IT infrastructure. In recent years, the company has focused heavily on AI-transforming its cloud to offer power for generative artificial intelligence, machine learning and simulations.
In addition to its own projects, Oracle also provides computing capacity to partners – among them, for example, OpenAI, which uses part of their GPU clusters to train models. According to Ellison, the new Oracle AI supercluster is designed to handle models with trillions of parameters, at lower power and better scalability than existing solutions.
Oracle is also among the companies looking for an alternative to Nvidia. The shortage of GPUs on the market and the high prices of Blackwell chips have forced several players to look for new partnerships – and AMD became the logical choice. The AI supercluster itself is a huge cluster of compute nodes that combine thousands of GPUs into a single unit capable of processing gigantic amounts of data in real time. Their MI450 will offer high performance, large HBM4 memory, and open UALink interconnect to enable better integration with existing OCI systems.
GPU purchase and collaboration with AMD
The official announcement came on October 14, 2025 during Oracle CloudWorld. The company confirmed that it will build approximately 700 racks that will collectively hold 50,000 units of MI450 GPUs in the first phase. The estimated value of the contract is between $3.5 billion and $4 billion , according to analysts, making it the largest GPU purchase in AMD’s history outside of Nvidia.
The project is the result of a long-term collaboration Oracle has been building with AMD since 2023 – when Oracle deployed the first clusters with MI300X and later MI355X GPUs. The new MI450s, however, are in a completely different league and deliver massive improvements in performance and efficiency. According to the available data, the new system is expected to consume approximately 200 MW of power in its final configuration. Oracle claims that despite this high consumption, the Oracle AI supercluster will be more efficient than current Nvidia solutions thanks to advanced cooling and optimized architecture.
Oracle AI Supercluster performance and specifications
The new AMD Instinct MI450s are among the most powerful accelerators AMD has created to date for artificial intelligence. The GPUs are manufactured using TSMC’s 2nm process and use HBM4 memory with record-breaking throughput. These accelerators will be deployed as part of AMD’s Helios rack architecture, which combines GPUs, processors and networking elements into a single computing unit. Eight MI450 GPUs are housed in a single rack module – together they form an extremely high-performance compute node that forms the core of the Oracle AI supercluster.
| Specification | Value |
|---|---|
| Manufacturing process | 2 nm (TSMC) |
| Memory | 432 GB HBM4 |
| Memory throughput | 19.6 TB/s |
| System architecture | AMD Helios rack-scale |
| GPU in a single module | 8 × MI450 |
| Performance per module | ~3.2 exaFLOPS (FP4) |
| DPU network throughput | 800 Gbps per chip |
| GPU ↔ GPU interconnect | UALink over Ethernet (UALoE) |
Each accelerator is interconnected with Pensando Vulcano DPU chips that take care of data transfer and communication between each GPU. This combination creates a unified memory domain that allows the accelerators to work together efficiently without performance loss. Communication within the rack is done via UALink over Ethernet (UALoE), an open standard that replaces closed solutions like NVLink. AMD thus brings a more flexible and affordable approach to powerful AI hardware.
Oracle AI Supercluster Summary
This purchase is a strategic step for Oracle towards greater independence from Nvidia. The partnership with AMD gives it access to powerful MI450 GPUs and paves the way to its own world-class AI infrastructure. The new Oracle AI supercluster will bring tremendous computing power for generative AI, scientific simulations and enterprise solutions. The first MI450 racks will be deployed in the second half of 2026, with further expansion in 2027.
If the project succeeds, Oracle Cloud Infrastructure will rank among the world’s highest-performing AI clouds and become the first provider with public access to AMD MI450 GPUs.