In 2025, Nebius exploded onto the European AI scene, positioning itself as the "Anti-Hyperscaler." Their pitch is simple: "We are built for AI. AWS is built for web apps." But marketing aside, how do the numbers actually stack up?

We modeled the Total Cost of Ownership (TCO) for a standard training run: Training a 70B Llama-3 variant on 64 H100 GPUs for one month (720 hours).

1. Compute Costs (The H100 Factor)

Compute is the single largest line item. Let's compare the flagship instances.

• AWS (p5.48xlarge): This is AWS's 8xH100 node.

  On-Demand Price: ~$98.32 / hour per node.

  Per GPU: ~$12.29 / hour.

• Nebius (H100 SXM5):

  On-Demand Price: ~$2.25 - $2.95 / hour per GPU.

  Node Price (8x): ~$18.00 - $23.60 / hour.

The Gap: AWS is roughly 4x-5x more expensive on raw on-demand compute. Even if you commit to a 3-year Compute Savings Plan on AWS (saving ~50%), the price comes down to ~$6.00/GPU/hr, which is still double Nebius's on-demand rate. For pure FLOPs per dollar, Nebius wins by a landslide.

2. Networking & Egress (The Hidden Killer)

This is where hyperscalers make their margin. "Data Gravity" is expensive.

• AWS Egress: Standard pricing is ~$0.09 per GB for data transfer out to the internet. If you move 1 Petabyte of training data out, that is a $90,000 bill.

• Nebius Egress: Typically charges ~$0.015 per GB or includes generous allowances. Egress within their storage network is free.

For research teams that frequently move checkpoints (model weights) between clouds or to Hugging Face, the "Egress Tax" on AWS can effectively lock you in. Nebius's low egress fees encourage a multi-cloud approach.

3. The Network Performance (InfiniBand vs EFA)

Pricing is irrelevant if the training takes twice as long. In distributed training (64+ GPUs), "All-Reduce" operations—where GPUs synchronize gradients—are the bottleneck.

• AWS EFA (Elastic Fabric Adapter): Runs on Ethernet. Bandwidth is 400 Gbps (on P4) or higher on P5. It is fast, but it has higher latency and jitter than InfiniBand.

• Nebius InfiniBand: Uses Nvidia Quantum-2 InfiniBand networking with 3.2 Tbps bandwidth per host.

The Impact: On large clusters, InfiniBand can reduce training time by 20-30% compared to Ethernet-based solutions due to lower latency. If your job finishes 20% faster, you pay 20% less for compute. This is a "Speed Dividend."

The "Ecosystem tax"

So why does anyone use AWS? Because of the Ecosystem.

• AWS: You get IAM, S3, CloudWatch, SageMaker, VectorDBs, Load Balancers, and a thousand other services integrated instantly. It is "batteries included."

• Nebius: You get Compute, Kubernetes, and basic Object Storage. You have to build the rest yourself.

Final Verdict: Build vs. Serve

The smartest CTOs use Nebius as their "Gym" (where models are trained) and AWS as their "Stage" (where models are served to customers).