Recommended Free Tools
Oracle said it signed a multibillion-dollar contract with AMD to build a cluster of 30,000 Instinct MI355X accelerators. That was a commitment to build, not proof that all 30,000 were installed and operating. Since then, Oracle has announced OCI capacity scalable to as many as 131,072 MI355X GPUs and made an eight-GPU instance generally available. Those are distinct milestones: the larger figure is a maximum announced scale, while the public cloud offering is an eight-GPU configuration.
What Oracle committed to
During Oracle’s fiscal 2025 third-quarter earnings discussion in March 2025, Chairman and CTO Larry Ellison said Oracle had signed a multibillion-dollar contract with AMD to build a cluster of 30,000 MI355X GPUs. Oracle’s quarterly results announcement provides the reporting context; the commitment was described in the earnings-call transcript.
The public statement establishes a contract to build a cluster. It does not disclose the contract’s dollar value, customer or end user, facility location, delivery milestones, final network and cooling design, or whether “30,000” referred to individual accelerator modules or a phased cluster target. The multibillion-dollar description is Oracle’s characterization; it does not support a reliable per-accelerator price estimate.
What has happened since the announcement
| Date | Publicly announced milestone | What it establishes |
|---|---|---|
| March 2025 | Oracle disclosed a multibillion-dollar AMD contract to build a 30,000-MI355X cluster. | A stated contract commitment, not confirmation of a completed deployment. |
| June 12, 2025 | Oracle and AMD announced OCI infrastructure scalable to as many as 131,072 MI355X GPUs. | A later maximum scale for the announced platform, not evidence that this many GPUs had been installed. Oracle and AMD announcement. |
| October 14, 2025 | Oracle announced general availability of the BM.GPU.MI355X.8 bare-metal shape. | Customers could access an eight-accelerator OCI configuration, subject to region, quota and capacity. Oracle availability announcement. |
These milestones do not establish that Oracle completed and accepted all 30,000 accelerators, or that it had 131,072 in service. The availability of an eight-GPU cloud shape is a customer-access milestone, not a deployment count for the larger cluster.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →#1 Best Overall
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
What the MI355X brings
The MI355X is part of AMD’s Instinct MI350 series and uses its fourth-generation CDNA architecture. AMD lists 288 GB of HBM3E memory and up to 8 TB/s of memory bandwidth per accelerator, along with 16,384 stream processors, 1,024 matrix cores, and support for low-precision formats including MXFP4 and MXFP6. AMD gives June 12, 2025 as the product launch date. These are manufacturer specifications, not independent benchmark results. See AMD’s MI355X specifications.
Large memory capacity can help keep larger models or more of their working state on accelerators, potentially reducing the need to split a workload across devices. The real benefit depends on the model, precision, batch size, software and distributed setup; specifications alone do not determine training or inference speed.
What OCI customers can access
Oracle’s generally available BM.GPU.MI355X.8 is a bare-metal instance with eight MI355X accelerators. Oracle’s October 2025 announcement lists this configuration:
Rank #2
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
- 2.3 TB aggregate HBM3E across the eight accelerators
- 128 fifth-generation AMD EPYC CPU cores and 3 TB of DDR5 system memory
- 61.44 TB of local storage
- 400 Gbps front-end networking and 3,200 Gbps cluster networking
Oracle’s launch material gave $8.60 per hour as a price signal for the shape at the time of its October 14, 2025 announcement. It is not a guaranteed current or universal rate: price can vary with region, billing arrangement, contract, availability and additional services. Check Oracle’s OCI price list for current regional pricing and confirm capacity and quota for the intended region before planning a workload.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →An hourly instance rate is only one part of operating cost. Storage, networking, data transfer, support, setup and idle time can affect total spend. An eight-GPU bare-metal instance may suit development, inference and distributed work, but a small or occasional job may not use enough of the capacity to justify it.
Why the announcement matters to Oracle, AMD and cloud buyers
Thirty thousand accelerators would be a very large installation compared with a conventional enterprise AI system. Capacity at that scale could serve model pretraining, fine-tuning, high-volume inference, multimodal workloads, AI agents or scientific computing. Oracle did not identify a customer or specify which workload the original cluster would serve, so none of those uses should be treated as a confirmed assignment.
Rank #3
- Built for Running LLMs Locally: RDNA 4, 128 AI Accelerators, up to 1,531 TOPS (INT4) for fast inference and fine-tuning
- 32GB GDDR6 VRAM for Large AI Models: 256-bit, up to 640GB/s bandwidth, run large language and multi-modal AI models without offloading
- Multi-GPU Scaling for Local AI Clusters: PCIe 5.0 and 2-slot design support dense multi-GPU builds for local AI training and inference clusters
- Diecast Shroud and Backplate: Wave-pattern design cuts memory temperature by up to 16%, keeping clocks steady during long AI training runs
- Phase-Change GPU Thermal Pad: Delivers superior thermal conductivity for consistent performance and longevity under heavy AI loads
For Oracle, a large accelerator fleet can support OCI’s position as a place to rent AI compute, including for customers that want to run models near Oracle databases and other cloud services. For AMD, a hyperscaler deployment offers a route to make its accelerator platform available beyond customers building their own data centers. A reasonable interpretation is that Oracle is broadening its supply options rather than relying on one accelerator supplier; that is analysis, not a stated explanation of the contract.
Oracle and AMD said the platform could deliver more than 2× better price-performance than the previous AMD GPU generation in large-scale AI training and inference, and up to 2.8× higher throughput for certain AI deployments. These are vendor claims. Without a named workload, precision, software configuration, comparison baseline and test conditions, they do not establish a universal advantage over Nvidia or any other platform.
What a buyer should check before choosing MI355X
Workload fit and software
MI355X runs in AMD’s ROCm software ecosystem rather than Nvidia’s CUDA ecosystem. A team should verify support for its framework, model, inference server, kernels, quantization method and distributed-training setup before committing. Oracle has described ROCm 7.0 work and performance validation for model serving; its technical details are vendor material, not proof that every application will run equally well.
Rank #4
- 24GB GDDR7 ECC Memory: handles large AI, 3D and rendering files smoothly
- Powerful CUDA Compute - 8,960 CUDA cores for fast graphics and computing power
- AI & Ray Tracing Boost - Tensor of the 5th generation and RT cores of the 4th generation
- PCIe 5.0 x16 interface - fast data connection with modern systems
- 4 × DisplayPort 2.1 - Multi-monitor support for professional workflows
CUDA-dependent applications or custom CUDA kernels may need porting and validation. AMD provides ROCm documentation and an Instinct customer acceptance guide for teams evaluating software and system requirements.
Capacity, networking and total cost
GPU counts alone do not predict distributed performance. A large cluster also depends on networking, storage throughput, cooling, power delivery, scheduling, fault handling and how well software uses the interconnect. Oracle’s eight-GPU shape lists 3,200 Gbps of cluster networking; that specification does not by itself establish performance at 30,000-GPU scale.
- Check that the target OCI region has capacity and that the account has the needed quota.
- Benchmark the actual model, precision, sequence length and batch size on the intended software stack.
- Include storage, networking, data movement, support and idle capacity in cost estimates.
- Compare a full workload on OCI with alternatives in the cloud where the data and operations already reside.
For comparison, AWS lists its accelerated-computing options at Amazon EC2 accelerated computing; Microsoft documents Azure GPU-optimized virtual machines; Google lists Google Cloud GPU platforms; CoreWeave describes its GPU cloud; and Vultr lists its GPU cloud. The exact GPU families, prices and regional availability vary, so these links are starting points for checking current fit—not like-for-like price comparisons.
What remains unconfirmed
The public announcements cited here do not establish the 30,000-unit cluster’s location, a named end customer, its delivery and installation schedule, acceptance of all units, utilization, or production performance at that scale. Early reporting anticipated arrivals by mid-2025, but that expectation is not evidence of completion. The later announcement of capacity scalable to 131,072 GPUs likewise describes a maximum platform scale rather than a verified operating fleet.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
