Oracle Cloud Infrastructure GPU Compute Review for AI and Machine Learning Workloads

The research in brief

Oracle Cloud Infrastructure GPU Compute is a good, but not unequivocally strong, fit for GPU cloud compute for AI training. Three of the six included platforms named it during the ranking stage - DeepSeek, Gemini, and Perplexity - giving it a 50% share of included platform responses, an average listed rank of 8.67, and a best listed rank of 7. The strongest reason to consider it is bare-metal 8-GPU A100 and H100 nodes with RDMA/RoCE cluster networking and large local NVMe storage ^[4]. The main limitation is that preemptible pricing is documented at 50% below on-demand but excludes bare-metal instances, so the recommended modern shapes may not qualify ^[1].

Research Snapshot

Why Oracle Cloud Infrastructure GPU Compute Qualified for This Study

Questions explored

• Is Oracle Cloud Infrastructure GPU Compute a good choice for GPU cloud compute for AI training?
• Why did Oracle Cloud Infrastructure GPU Compute qualify for this GPU cloud hosting study?

Oracle Cloud Infrastructure GPU Compute qualified because it was named by three of the six included platforms during the ranking stage and because its documented product line maps directly to the study's criteria: GPU model availability, GPU pricing per hour, multi-GPU cluster support, framework and driver support, and spot or preemptible pricing options.

The three platforms that named it were DeepSeek (rank 9), Gemini (rank 10), and Perplexity (rank 7). Its average listed rank was 8.67 and its best listed rank was 7. The remaining included platforms - ChatGPT, Grok, and Kimi - supplied fit research on the entity but did not name it during ranking discovery, so they are not counted in the mention total.

Qualification is not the same as endorsement. The ranking-stage mention count reflects discovery, not quality, and the fit ratings below come from platform-reported analysis rather than independent testing. Five platforms rated the fit "good" and one rated it "strong" (Grok), but those ratings rest largely on company-owned documentation.

The hosting offering in focus

Questions explored

• Which Oracle Cloud Infrastructure GPU Compute shape should a buyer choose for multi-GPU AI training?
• Is BM.GPU.H100.8 or BM.GPU4.8 the better Oracle Cloud Infrastructure GPU Compute option for distributed training?

The most relevant offering is OCI bare-metal GPU Compute, specifically BM.GPU4.8 (8x A100 40 GB) and BM.GPU.H100.8 (8x H100 80 GB), plus supported virtual-machine GPU shapes. Platform-supplied product names varied slightly: ChatGPT and Perplexity named BM.GPU4.8 and BM.GPU.H100.8; Gemini named BM.GPU.H100.8, BM.GPU.H200.8, and BM.GPU4.8; Grok named BM.GPU.H100.8, BM.GPU.H200.8, and BM.GPU.B200.8; Kimi named BM.GPU.H100.8; DeepSeek named BM.GPU4.8 and BM.GPU.H100.8.

OCI documents A100, H100, H200, Blackwell, L40S, A10, and AMD MI300X/MI355X GPU offerings ^[4]. BM.GPU4.8 provides 8 GPUs and 8x200-Gbps RDMA networking; BM.GPU.H100.8 provides 8 H100 GPUs and 8x2x200-Gbps RoCE-capable RDMA networking ^[4]. Oracle's own GPU page states the company offers bare-metal instances with NVIDIA and AMD GPUs free of virtualization overhead and cites 61.4 TB of local storage per node with H100 GPUs (official:C1).

Note the naming conflict: DeepSeek described BM.GPU4.8 as carrying A100 80GB, while ChatGPT and Perplexity described it as 8x A100 40 GB. Google listed BM.GPU4.8 at 8x A100 SXM 40GB and separately listed BM.GPU.A100-v2.8 at 8x A100 SXM 80GB ^[26]. Buyers should confirm the exact memory configuration of the SKU they order rather than relying on the shape name alone.

What the AI Platforms Agreed About

Questions explored

• What do AI platforms agree Oracle Cloud Infrastructure GPU Compute does well for AI training?
• Does Oracle Cloud Infrastructure GPU Compute support multi-node distributed training with RDMA networking?

The platforms agreed most strongly on three points: bare-metal 8-GPU configurations, high-bandwidth cluster networking, and documented CUDA/NVIDIA software baselines.

On GPU availability, ChatGPT, DeepSeek, Perplexity, Gemini, Grok, and Kimi all described OCI as offering A100 and/or H100 bare-metal shapes ^[4]. Google additionally listed H200, B200, GB200, and AMD MI300X ^[25], and Grok listed H200 and B200 ^[17].

On multi-GPU cluster support, DeepSeek cited RDMA over Converged Ethernet (RoCE) for multi-node GPU training ^[6], and Gemini described non-blocking Clos topologies with NVIDIA ConnectX-7 NICs supplying up to 1.6 Tbps per node on H100 bare-metal shapes ^[10]. Perplexity cited Oracle documentation on bare-metal GPU cluster scaling for AI/HPC workloads ^[8].

On framework and driver support, OCI's H100 quick start documents Oracle Linux 8/9 and Ubuntu 22.04/24.04, NVIDIA drivers 580/590 Open, CUDA 13.0/13.1, NVIDIA tooling, and NCCL topology considerations ^[2]. Google reported SLURM orchestration on BM.GPU.H100.8 multi-node clusters and Kubernetes with the NVIDIA GPU Operator ^[9].

Agreement among platforms does not establish product quality. Much of this agreement traces back to the same company-owned Oracle documentation.

Where the AI Platforms Disagreed or Were Uncertain

Questions explored

• Is preemptible pricing available on Oracle Cloud Infrastructure GPU Compute bare-metal H100 shapes?
• How much does Oracle Cloud Infrastructure GPU Compute charge per hour for BM.GPU.H100.8?

The sharpest conflict concerns preemptible pricing on the recommended shapes. Oracle's preemptible documentation states preemptible instances are priced at 50% below on-demand and that bare-metal instances are not available as preemptible instances; it lists VM.GPU2 and VM.GPU3 among supported GPU shapes ^[1]. Yet several platform responses bundled "Preemptible Instances" with BM.GPU4.8 and BM.GPU.H100.8 as the relevant plan (DeepSeek, Perplexity, Kimi). Grok went further, reporting BM.GPU.H100.8 at $40/hr preemptible ^[34]. That claim conflicts with the company documentation excluding bare metal from preemptible capacity. Buyers should treat preemptible availability on modern bare-metal GPU shapes as unverified.

Pricing transparency is the second area of uncertainty. Oracle's published service table lists BM.GPU4.8/A100 40 GB at $2.6212 per GPU-hour, roughly $20.97 per 8-GPU node-hour ^[18]. Google reported BM.GPU4.8 at $3.05 per GPU-hour ($24.40/node-hour) and BM.GPU.H100.8 at $10.00 per GPU-hour ($80.00/node-hour) ^[26]. Grok reported the same $80/hr H100 figure via an independent tracker ^[34], and Kimi reported approximately $10.00 per GPU-hour from a third-party aggregator ^[24]. Perplexity could not reliably extract the exact value from Oracle's price list and rated pricing confidence "low" ^[15]. OpenAI noted that the retrieved official pricing material did not clearly establish the current BM.GPU.H100.8 compute rate and directed buyers to the OCI Cost Estimator ^[20].

Other unresolved items: whether OCI offers NVLink-domain-level networking versus RDMA-only (Kimi); availability of B200, GB200, or GB300 (Kimi, DeepSeek); exact preemptible discount percentages for H100 (Kimi); and region-specific shape availability for a required US region and node count (ChatGPT, DeepSeek, Perplexity, Gemini).

Features for this workload

Questions explored

• Does Oracle Cloud Infrastructure GPU Compute support PyTorch, TensorFlow, CUDA, and NCCL for AI training?
• How much local NVMe storage does Oracle Cloud Infrastructure GPU Compute provide for training checkpoints?

For distributed AI training, OCI's documented strengths are node-level GPU density, RDMA cluster fabric, local NVMe, and standard Linux/NVIDIA software stacks.

Storage and checkpointing: BM.GPU4.8 includes 27.2 TB of local NVMe, while BM.GPU.H100.8 documentation describes approximately 54 TB usable local NVMe ^[4]. Oracle's own page claims 61.4 TB of local storage per node with H100 GPUs (official:C1). These figures differ, and local NVMe is instance-local - it should not be treated as durable shared storage without a separate persistence design (ChatGPT).

Networking: BM.GPU4.8 provides 8x200-Gbps RDMA; BM.GPU.H100.8 provides 8x2x200-Gbps RoCE-capable RDMA ^[4]. Google described 8 dedicated 200 Gbps ConnectX-7 NICs providing 1.6 Tbps per node with GPUDirect RDMA ^[27].

Software: OCI GPU quick starts provide provisioning, approved images and drivers, OKE guidance, and diagnostics ^[3]. OCI GPU instances support NVIDIA CUDA with documentation for PyTorch and TensorFlow, and preinstalled drivers are available via Oracle Cloud Marketplace ^[5]. OCI Data Science supports specific GPU shapes with reservation and service-limit mappings ^[7].

Orchestration: OCI documents operation with Oracle Kubernetes Engine and provides GPU quick starts ^[3]. Google reported SLURM orchestration on multi-node H100 clusters ^[9]. However, the reviewed material does not establish that OCI provides a turnkey managed distributed-training control plane comparable to specialized GPU-cloud services (ChatGPT), and Kimi characterized OCI as lacking the broad managed ML platform ecosystem of AWS SageMaker, Google Vertex AI, or specialized neocloud orchestration ^[33].

Pricing, renewal terms, and ongoing costs

Questions explored

• How much does Oracle Cloud Infrastructure GPU Compute cost per hour for A100 and H100 nodes?
• Are there setup, egress, or storage fees on top of Oracle Cloud Infrastructure GPU Compute hourly rates?

Published rates conflict across sources, and the H100 rate is not cleanly verified from Oracle's own retrieved material.

Preemptible instances are documented at 50% of regular instance pricing, with the discount applied in addition to other discounts such as UCM negotiated rates (Evidence ID perplexity:c1, Evidence ID google:1.3.7, official:C2). Unused capacity reservations are priced at 85% of regular instance pricing (Evidence ID google:1.1.5, official:C2).

Additional fees reported across platforms include block volume storage and performance provisioning, object storage, network egress, optional software licensing including NVIDIA AI Enterprise, and taxes (ChatGPT, DeepSeek, Gemini). Google and Grok both reported 10 TB of free monthly egress, with Grok citing $0.0085/GB thereafter ^[31].

Contract terms: PAYG is available without an annual Universal Credits commitment; Universal Credits involve an annual spend commitment with excess usage invoiced per the contracted rate card ^[22]. Google reported that high-end GPU configurations typically require OCI Universal Credits and that reserved capacity contracts run 1 to 3 years with typical discounts of 30% to 60% (Gemini). Deepseek noted committed contracts may carry termination fees or minimum spend requirements (DeepSeek). Oracle's price list states that cancelling service prior to the end of a commitment interval does not stop metering (official:C2).

Pricing confidence varied: Grok and Gemini rated it high, ChatGPT, DeepSeek, and Kimi rated it moderate, and Perplexity rated it low.

Best Suited For

Questions explored

• Who gets the most value from Oracle Cloud Infrastructure GPU Compute for AI training?
• Is Oracle Cloud Infrastructure GPU Compute best for large-scale multi-node training rather than small experiments?

OCI GPU Compute is best suited to teams running substantial distributed training on whole 8-GPU bare-metal nodes.

Platforms converged on these buyer profiles: large training jobs using 8-GPU bare-metal nodes (ChatGPT); distributed training that benefits from RDMA-capable cluster networking (ChatGPT, DeepSeek, Gemini); teams needing substantial local NVMe for checkpoints or datasets (ChatGPT, Perplexity); buyers already using OCI networking, storage, Kubernetes, or Oracle Universal Credits (ChatGPT, DeepSeek, Kimi); and fault-tolerant experimentation on supported preemptible VM GPU shapes (ChatGPT, DeepSeek).

Grok rated the fit "strong" specifically for large-scale multi-node AI/ML training clusters and workloads needing high-bandwidth RDMA interconnect and local NVMe storage (Grok). Google described the best fit as large-scale multi-node distributed training requiring non-blocking networks, bare-metal access to eliminate virtualization overhead, and workloads with large data transfers benefiting from free egress tiers (Gemini).

Probably Not Best Suited For

Questions explored

• Who should not choose Oracle Cloud Infrastructure GPU Compute for AI training?
• Is Oracle Cloud Infrastructure GPU Compute a poor fit for small single-GPU training experiments?

OCI GPU Compute is probably not the best fit for small workloads, interruption-intolerant jobs, or buyers who need preemptible capacity on current-generation bare-metal shapes.

Platforms identified these exclusions: small workloads needing one modern GPU at a time (ChatGPT, Gemini); training jobs that cannot tolerate interruption (ChatGPT, Perplexity); buyers specifically requiring preemptible BM.GPU.H100.8 or BM.GPU4.8 capacity (ChatGPT); teams prioritizing the broadest managed ML ecosystem or the most granular GPU-instance selection (ChatGPT, DeepSeek, Kimi); hobbyists or small teams wanting cheap single-GPU VMs on a fractional basis (Gemini); and teams without Oracle expertise seeking the fastest path to distributed training (Kimi).

Bare-metal shapes require whole-node allocation, which can be inefficient for small experiments (ChatGPT). Google noted that most flagship AI-training GPUs must be rented as whole 8-GPU bare-metal nodes rather than single-GPU allocations (Gemini).

When Another Option May Be Better

Questions explored

• What is a better alternative to Oracle Cloud Infrastructure GPU Compute for cheap single-GPU AI training?
• When should a buyer choose AWS, Azure, or Google Cloud over Oracle Cloud Infrastructure GPU Compute?

Several platforms named specific circumstances where alternatives may serve better.

For inexpensive single-GPU or fractional-GPU experimentation, ChatGPT recommended a specialized GPU cloud (ChatGPT). Google recommended specialized clouds such as Lambda Labs or RunPod for micro-sizing and lower pay-as-you-go rates, and named Vast.ai for volatile ultra-cheap spot markets (Gemini). Kimi named AWS EC2 Capacity Blocks ($5.19–6.88), Nebius ($3.85), and Vast.ai marketplace ($1.73) as lower GPU-hour rates ^[24].

For mature managed ML ecosystems, ChatGPT recommended AWS, Azure, or Google Cloud when the team depends heavily on their managed ML, data, identity, or orchestration ecosystems (ChatGPT). Deepseek named AWS SageMaker, Azure ML, or Google Vertex AI for managed AI services and community support (DeepSeek). Kimi named CoreWeave, Lambda, or AWS SageMaker for more integrated AI tooling (Kimi).

For immediate access to newer silicon, Kimi named Crusoe, Nebius, or CoreWeave for GB200/GB300 NVL72 racks (Kimi). For per-second billing granularity, Kimi named RunPod or Vast.ai (Kimi). For EU data residency, Kimi named Nebius ^[33].

For guaranteed access to a specific GPU generation, region, node count, or contractual capacity SLA, ChatGPT recommended another provider when those guarantees matter more than OCI's bare-metal pricing and networking (ChatGPT).

Questions to ask before choosing a provider

Questions explored

• What should a buyer confirm with Oracle Cloud Infrastructure GPU Compute before signing a contract?
• How can a buyer verify Oracle Cloud Infrastructure GPU Compute H100 capacity and pricing before committing?

Platforms supplied overlapping verification lists. The recurring items:

• Current on-demand hourly rate for BM.GPU.H100.8 in the exact US region (ChatGPT, DeepSeek, Perplexity, Kimi).
• Whether Oracle can provision the required number of H100 or A100 nodes concurrently and what service-limit increase is needed (ChatGPT, Perplexity, Gemini).
• Whether the required shapes are available through capacity reservations and what reservation commitment or cancellation terms apply (ChatGPT, DeepSeek, Gemini).
• Which current GPU VM shapes, if any, support preemptible capacity in the target region (ChatGPT, Perplexity, Kimi).
• Total storage, object-storage, inter-region transfer, and internet-egress costs for the planned pipeline (ChatGPT, DeepSeek, Perplexity).
• Which NVIDIA driver, CUDA, NCCL, PyTorch, and TensorFlow combinations are supported for the intended container images (ChatGPT, Perplexity, Kimi).
• How RDMA networking, placement, topology, and NCCL configuration are exposed across multiple bare-metal nodes (ChatGPT, Kimi).
• What checkpoint, restart, and replacement automation is available if using preemptible capacity (ChatGPT, DeepSeek).
• What support response times and escalation paths apply to GPU capacity, networking, and driver failures (ChatGPT, DeepSeek, Kimi).
• Whether the planned purchase uses PAYG or Universal Credits, and whether annual minimums, expiration rules, or unused-credit consequences apply (ChatGPT, Gemini).
• Whether a minimum Universal Credits contract is required to access bare-metal GPU shapes or whether PAYG self-service is fully active in the target region (Gemini).
• Whether the intended training framework fully supports OCI's RoCE v2 cluster networking topology out of the box or requires custom driver and network configuration (Gemini).

The consensus assessment

Oracle Cloud Infrastructure GPU Compute is a good, but not unequivocally strong, fit for GPU cloud compute for AI training. Five of six included platforms rated the fit "good" and one rated it "strong," but those ratings rest heavily on company-owned documentation rather than independent validation.

Its strongest case is an AI startup or data-science team running substantial distributed jobs on 8-GPU A100 or H100 bare-metal nodes with RDMA networking and local NVMe. Its main weaknesses are limited applicability of preemptible pricing to the recommended modern bare-metal shapes, whole-node allocation, uncertain current H100 pricing, and capacity/availability verification requirements.

The consensus is conditional rather than enthusiastic. Buyers who need fractional GPU access, guaranteed preemptible capacity on current-generation hardware, or fully transparent public H100 pricing should verify those points directly before committing. Buyers already inside the Oracle ecosystem, or those whose workloads justify whole 8-GPU nodes, have a defensible case for OCI.

How This Review Was Produced

This review was produced from platform-supplied fit research collected for the study "Best GPU Cloud Hosting for AI and Machine Learning Workloads," with a run research date of 2026-10-03. Six platforms supplied fit-research responses: ChatGPT, DeepSeek, Gemini, Grok, Kimi, and Perplexity. Three of those six named Oracle Cloud Infrastructure GPU Compute during the ranking stage.

Each platform response included a fit rating, use-case findings, pricing and terms, strengths, limitations, and questions to verify before buying. Those responses were synthesized into the sections above. Citation IDs in parentheses map to the source list below.

Platform-reported research dates differed from the run date: DeepSeek reported 2026-06-01, while Gemini, Grok, Kimi, ChatGPT, and Perplexity reported 2026-10-03. Platform-reported dates are provenance metadata and do not independently prove freshness.

Research limitations

Several limitations apply to this review.

Platform mentions count only platforms that named the entity during ranking discovery; all six included platforms evaluated fit, but three did not name it during ranking. Company-owned citations materially outnumber independent citations in the supplied evidence, so company claims are not described here as independently verified. The supplied URLs were collected from platform responses and were not independently validated by the writer stage. Citations are platform-reported evidence, not independently verified facts.

Deepseek's research was conducted without search enabled, so its claims require explicit verification before being treated as current facts. Platform-reported research dates differ from the authoritative run date.

Conflicting product names, pricing, and capabilities were described rather than resolved. Specifically: BM.GPU4.8 memory configuration (40 GB vs 80 GB), H100 hourly pricing, preemptible availability on bare-metal shapes, and local NVMe capacity figures all conflict across sources. No personal testing, customer experience, or independent verification was performed for this review.

Explore more vps cloud and dedicated server hosting guidance in the category directory.

Sources

Company-Owned Sources

1. Preemptible Instances: https://docs.oracle.com/en-us/iaas/Content/Compute/Concepts/preemptible.htm
2. OCI GPU Quick Start: NVIDIA H100: https://docs.oracle.com/en-us/iaas/Content/Compute/gpu-quick-start/nvidia/H100/README-H100.htm
3. OCI GPU Quick Starts: https://docs.oracle.com/en-us/iaas/Content/Compute/gpu-quick-start/overview.htm
4. Compute Shapes: https://docs.oracle.com/en-us/iaas/Content/Compute/References/computeshapes.htm
5. GPU Instances - Oracle Cloud Infrastructure: https://docs.oracle.com/en-us/iaas/Content/Compute/References/gpu.htm
6. GPU Cluster Networking - Oracle Cloud Infrastructure: https://docs.oracle.com/en-us/iaas/Content/Compute/Tasks/gpu-cluster-networking.htm
7. Supported Compute Shapes: https://docs.oracle.com/en-us/iaas/Content/data-science/using/supported-shapes.htm
8. Deploy a Bare Metal GPU Cluster for AI Workloads in a Dedicated: https://docs.oracle.com/en/solutions/deploy-bare-metal-gpu-cluster-for-ai/index.html
9. OCI File Storage with HPMT for AI Model Training and Fine-Tuning | cloud-infrastructure: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE5CQzplsGJRp6PdfGoWIHAC49YxfGalwCSbSwmDHX9XJ7rRfkXi1p2Zy80HfTF-ciyXUoKPyWa0dnpQPKBoxHwYJDcUrDkCtsAjRewg5L-77LZ2yY2jypRjTRk99pvjaczR7GQR60g4RXAZcGXH-8e6YbMqoUwEBZfNR7-UHr8ee6X29HChfQfqIyEDUD6fbc=
10. Now Generally Available: The Largest, Fastest AI Supercomputer in the Cloud: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEUhDdDRz96jXoktMPP84lb79X5ox6MR_x3FEZgjhkuYrm8kBpeBmXRqoy8oONMHmpUwP7hFL42W2PkD1sJGto4OkKxZfQUSLzXVt_shGzzrEB9hZd4DmVSRDv8lsb6ORFk4v38IdrqpzDChcoxdji__tM3bkshC6Ne0YDJ97s-nq68aRyrCw7JgKQ6fSZRB2Dw4A==
11. Precios de procesamiento de Oracle Cloud Compute: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3jte6NLzHq-7_kV2EswfmEHd1bdxe16h4SCrHQzjxthfF5MS9l66BOMHZuisk2BnrKp8HE6M18BC1ozcthUeC8xrL9rd4MlLsDQxuVgCnjibx4R-xxBuGb5iUvyFQ1KkCoV2SFh4wQqm06F4SSz4=
12. Lower Compute Costs by up to 50% with Preemptible Instances - YouTube: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGMaL6Vt2zhvVnbLQUr-n3_5nO1rL6SuqAlxLReL6dQf_dDQMS5UsRjFSGDmZxLEiWZB9Qy2e64OgWVcwi20OYKJ-MaesH8c4lN7bLerlP4fSSe7-WJAqgMvCCjjj3CrDQ=
13. Oracle Cloud Infrastructure Data Science: https://www.oracle.com/artificial-intelligence/data-science/
14. OCI Price List - Oracle: https://www.oracle.com/asean/cloud/price-list/
15. OCI Price List: https://www.oracle.com/bz/cloud/price-list/
16. Oracle Cloud Infrastructure Billing: https://www.oracle.com/cloud/billing/
17. GPU, Virtual Machines and Bare Metal: https://www.oracle.com/cloud/compute/gpu/
18. OCI IaaS and PaaS Services Highlights: https://www.oracle.com/cloud/iaas-paas/
19. GPU Compute Pricing - Oracle Cloud Infrastructure: https://www.oracle.com/cloud/price-list/#compute-gpu
20. Cloud Pricing: https://www.oracle.com/cloud/pricing/
21. Oracle Cloud Regions - GPU Availability: https://www.oracle.com/cloud/regions/
22. Cloud Universal Credits: https://www.oracle.com/cloud/universal-credits/
23. Official pricing and terms source: https://www.oracle.com/cloud/price-list/#pricing-compute

Independent Sources

24. How to Choose a GPU Cloud Provider for AI Model Training | GPUSmith: https://gpusmith.com/articles/en/choose-gpu-cloud-provider-ai-training
25. Oracle Cloud vs Ori Cloud: GPU Compute Price Comparison: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEo_rpoPIbFVSu--1A4qCauD462ObSyCeykh3yNlFA_9iCPwuSoaeeWsZH8vkdqLupjC7DfJ3pYeudZDtStYd4lrGV9CYN2cQ19Rrm5I-qqYEHDVxFU-rf1HuDkQK5eRQF5yTEbeetgkP9HnTMm9yeRg7u8KVBQ6CurceSH3oj6Ti5ZM4haYNsB
26. Oracle Cloud (OCI) GPU Pricing 2026: H100, H200, B200 and A100 Per-Hour Cost: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEtBKFofeCQVmHyJXtjc5GRTbUToFBOlw2RTwKf17HozS_lLiXveCqzNbC9EDnTUZEnsMiK1IMUvJ8bIQktrC3wsL2RvZ6CXRmAeXOTbRid6oej8QvenfTPYAySChxbJLfAKDm_Se3-tXEJ52A8RURV0f88a7s=
27. Clos Networks - Patterns that Connect: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEwIykgJCcwiffWcJI8jYSrtLjrG3SO_V3x71Y7V4qjH6iVYt5t3vjGbwOmhQTl6m7mMvL5SmKQDbKn9w2iQVMyADORZCg81ZWZGr7Sa8vJ3RbPwz_9ct86andUQmXkeSk3Gc1uvxo19cY5
28. (PDF) OCUDU-on-OCI with NVIDIA GPUs - ResearchGate: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEwZnAcN1Ko4mcNzowk-KJqHXDvD21wVl8QxU5ThN9flQPE6AzIRnwmKt4hFdWHIfzVS5kUV8GWdM1B76UMgv7NYZWUdmvy7MaX6a61Mkeq0hGZAQsEOFQMYTHHfqo_MkXSyDLse8pLk1MsjjBLH91OGE5BBWEPyg1sfh929KFDbxiNlLbt
29. H100 GPU Cloud Pricing Comparison: AWS vs GCP vs Azure | GPUSmith: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEyAb2DnKoWWwWmKZ1aUt49QLnTOadO3BSOxpLUGXbDxrsReB0pzAIDlD7cQDq5s2wdwLjvmsqrZpCImMvYtCX-sJN3nHia4noh9KD6Rj2ENLSMS-MiC9sd8oUeiTx3N57EiE0k9dJ_ihi_u-AOeP752aLXS5WCbw==
30. What an H100 hour lists for at Azure and Oracle, next to the rental marketplaces - Hugging Face Forums: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF8EpbO_o_oGBi8OBFCwIQrGxo669LqJr-A7ncyFKQMihejWEvmN-bWoJa4XFyDbDjzen-JXyyAFvh_KJ4DoA_4vaiwaCHnxFYCumqT4rrleu69xPZPX5bvyrBgABB4YjXAvVvnpsLhSbAbN_Rwh9J238b8THL6fcoFysv4GOMVO5Is1nAX8jBQHsiGkyZjXo7LFwTbCyaAiGIYWq1avV9j0nykp_6p-zPBR-Cq_uTKeSy4cOGb9Fmn0c5pbYNEEw==
31. OCI Costs Overview & How OCI Compares to AWS/Azure/GCP - Finout: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGvzjcj38XpYIhuXFcF8y9P2tqg5E_sO5irPV1tQ867f7Uw9iMlPuz_ujfpUytqmI47BP1YxXEtGcEVr4cXIT9cKyBnbIeUX0fG6sJ-bF1qIjt79CZUbRTyspK8FjubiUMLOQ==
32. Oracle Cloud GPU Pricing Explained: How OCI Bills GPUs and When It Fits: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHa5FZtOCoV-LX6JwxO9PRMVuAxOtR6-y8dIHrQTRHUrxJpuJ1k68mkjGuk72VDHfy_JnsFvSttYS-ysht1ZRvpeNjuY9gry8XCktAVwHhlv0f0J6LiN3m27vopgKaHIjivpqaWipVY2pdAZmLAfz-tmSVVo_Ywmw==
33. Nebius AI Cloud - Overview, Features & Use Cases | THE DAILY BRIEF: https://www.beri.net/tools/nebius-ai-cloud
34. OCI BM_GPU_H100_8 Pricing | Economize: https://www.economize.cloud/resources/oci/pricing/compute/BM_GPU_H100_8/
35. Oracle Cloud (OCI) GPU Pricing 2026: H100, H200, B200 and A100 Per-Hour Cost: https://www.thundercompute.com/blog/oracle-cloud-oci-gpu-pricing