Phi-3 on RTX 5090: Monthly Cost & Token Output
Dedicated RTX 5090 hosting for Phi-3 (3.8B) inference — fixed monthly pricing with unlimited tokens.
Monthly Cost Summary
294 tokens per second. That is not a typo. The RTX 5090 turns Phi-3 into a token-generation machine, producing over 762 million tokens monthly. With 28 GB of free VRAM, you could co-host two or three additional models alongside Phi-3 and still have headroom. At £179/month, the effective cost drops to just £0.23 per million tokens.
| Metric | Value |
|---|---|
| GPU | RTX 5090 (32 GB VRAM) |
| Model | Phi-3 (3.8B parameters) |
| Monthly Server Cost | £179/mo |
| Tokens/Second | ~294.0 tok/s |
| Tokens/Day (24h) | ~25,401,600 |
| Tokens/Month | ~762,048,000 |
| Effective Cost per 1M Tokens | £0.2349 |
The Fastest Phi-3 Deployment Available
At nearly 300 tok/s, the RTX 5090 makes Phi-3 faster than many cloud-hosted larger models. Compare the economics:
| Provider | Cost per 1M Tokens | GigaGPU Savings |
|---|---|---|
| GigaGPU (RTX 5090) | £0.2349 | — |
| Together.ai | $0.10 | Comparable |
| Fireworks | $0.20 | Comparable |
| Azure OpenAI | $0.26 | 10% cheaper |
Break-Even Analysis
Against Together.ai at $0.10/1M tokens, break-even is approximately 1,790M tokens/month. The 5090’s 28 GB of free VRAM enables batching at a scale that can push effective monthly throughput well beyond single-stream capacity, especially under heavy concurrent load.
Hardware & Configuration Notes
28 GB of spare VRAM for a 3.8B model is extraordinary. This setup is ideal for multi-model deployments: run Phi-3 for fast, lightweight queries alongside a 13B or even 70B quantised model for complex tasks, all on one card.
- VRAM usage: Phi-3 requires approximately 4 GB VRAM. The RTX 5090 provides 32 GB, leaving 28 GB headroom for KV cache and batching.
- Quantisation: Running in FP16 by default. INT8 or INT4 quantisation can reduce VRAM usage and increase throughput by 20–40% with minimal quality loss for most use cases.
- Batching: With continuous batching enabled (e.g., vLLM or TGI), you can serve multiple concurrent users from a single GPU, increasing effective throughput significantly.
- Scaling: Need more throughput? Add additional RTX 5090 nodes behind a load balancer. GigaGPU supports multi-server deployments with simple configuration.
Best Use Cases for Phi-3 on RTX 5090
- Multi-model GPU deployments combining Phi-3 with larger models
- Enterprise-scale lightweight LLM serving for hundreds of users
- Ultra-high-throughput batch processing of short-form text
- Real-time AI features embedded across multiple products
- Research environments requiring rapid model output generation
762M Tokens/Month from a Single GPU
Maximise Phi-3 throughput on a dedicated RTX 5090. £399/month, all-inclusive, zero metering.
