Rubin is coming.
Your GPU strategy should be ready.
10x cheaper tokens. 4x fewer GPUs for training. The Vera Rubin platform ships H2 2026 and will reshape cloud GPU pricing. Here's what you need to know — and what to do right now.
Per-GPU Comparison
Rubin vs Blackwell vs Hopper
| Spec | Rubin | Blackwell (B200) | Hopper (H100) |
|---|---|---|---|
| FP4 Inference | ~50 PFLOPS | ~9 PFLOPS | N/A |
| FP8 Compute | ~25 PFLOPS | ~9 PFLOPS | ~1.98 PFLOPS |
| HBM Capacity | 288 GB HBM4 | 192 GB HBM3e | 80 GB HBM3 |
| Memory Bandwidth | 22 TB/s | 8 TB/s | 3.35 TB/s |
| NVLink Bandwidth | 3.6 TB/s | 1.8 TB/s | 900 GB/s |
| Process Node | TSMC 3 nm | TSMC 4 nm | TSMC 4 nm |
| Transistors | ~336B | ~208B | ~80B |
| Token Cost (est.) | ~$0.005/1M | ~$0.05/1M | ~$0.20/1M |
Rack-Scale Comparison
NVL72 Rack: Rubin vs Blackwell
| Metric | Vera Rubin NVL72 | Blackwell NVL72 |
|---|---|---|
| Total GPUs | 72 Rubin | 72 B200 |
| FP4 Inference | 3.6 EFLOPS | ~720 PFLOPS |
| HBM Capacity | 20.7 TB | 13.8 TB |
| HBM Bandwidth | ~1.58 PB/s | ~576 TB/s |
| NVLink Rack BW | 260 TB/s | 130 TB/s |
| Power Draw | ~120–130 kW | ~120 kW |
Cloud Availability
Who's getting Rubin and when
Rubin entered full production January 2026. Volume is constrained to 200K–300K GPUs by TSMC 3nm and HBM4 supply.
1M+ NVIDIA GPUs (Blackwell + Rubin)
Rubin NVL72 Superclusters, 17+ ZetaFLOPS
Rubin NVL72 in Fairwater AI superfactories
Rubin deployments confirmed
10x token throughput vs Blackwell
Rubin NVL72 in Superintelligence Cloud
Rubin deployments announced
Rubin deployments announced
Decision Tool
Should you wait for Rubin?
Answer three questions. We'll give you a concrete recommendation.
1. What's your primary workload?
2. When do you need compute?
3. Budget sensitivity?
Available Now
Blackwell & Hopper: what you can rent today
While you wait for Rubin, these are the cheapest Blackwell and Hopper instances across our tracked providers.
| GPU | Provider | VRAM | Gen | Price/hr | |
|---|---|---|---|---|---|
5060 RTX5060Ti | Vast.ai | 16GB | Blackwell | $0.0360/hr | Deploy |
5060 RTX5060Ti | Vast.ai | 16GB | Blackwell | $0.0497/hr | Deploy |
5070 RTX5070Ti | Vast.ai | 16GB | Blackwell | $0.0640/hr | Deploy |
5060 RTX5060Ti | Vast.ai | 16GB | Blackwell | $0.0667/hr | Deploy |
5060 RTX5060Ti | Vast.ai | 16GB | Blackwell | $0.0667/hr | Deploy |
5060 RTX5060Ti | Vast.ai | 16GB | Blackwell | $0.0693/hr | Deploy |
5070 RTX5070 | Vast.ai | 12GB | Blackwell | $0.0773/hr | Deploy |
5090 RTX5090 | Vast.ai | 32GB | Blackwell | $0.100/hr | Deploy |
Track the Rubin rollout with us
We'll add Rubin instances to our tracker as soon as cloud providers launch them. Check back or read our deep-dive article for the full GTC 2026 breakdown.