How 10Gbps dedicated infrastructure eliminates the hidden bandwidth tax killing AI app profit margins.
What is the cloud egress tax? Cloud egress is the fee public cloud providers (AWS, Azure, Google Cloud) charge for every gigabyte of data your server sends outbound to users. Inbound data is typically free; outbound is billed per GB — creating a model where the more successful your application becomes, the higher your infrastructure bill climbs. For data-intensive AI applications, this outbound billing can account for 30–60% of a total monthly cloud invoice.
What is the cloud egress tax — and why does it hit AI apps hardest?
Public cloud providers have a straightforward pricing asymmetry: data flowing into your server (ingress) is free. Data flowing out to your users (egress) is metered and billed per gigabyte.
For a static website, egress costs are negligible. But modern AI applications are different:
- A real-time voice synthesis response streams 200–400 KB of audio per request
- A live AI image generation result delivers 1–4 MB per generation
- A video AI processing pipeline can output 50–200 MB per job
When an AI application reaches even modest scale, the egress volume becomes the dominant infrastructure cost, not compute.
Real numbers: what AWS, Azure, and GCP charge for egress in 2025
Figures drawn from publicly available pricing pages as of May 2025 (GBP).
| Provider | First 10 TB/month | Next 40 TB/month | Beyond 150 TB/month |
|---|---|---|---|
| AWS (London) | ~£0.074/GB | ~£0.070/GB | ~£0.052/GB |
| Azure (UK South) | ~£0.070/GB | ~£0.060/GB | ~£0.050/GB |
| GCP (europe-west2) | ~£0.095/GB | ~£0.080/GB | ~£0.060/GB |
| Bare Metal (Unmetered) | £0 (Included) | £0 (Included) | £0 (Included) |
What this means in practice: An AI application transferring 100 TB of outbound data pays approximately £6,500–£8,500 in egress fees alone on AWS or Azure. At 500 TB/month, that figure exceeds £30,000 per month.
The "success penalty"
The egress billing model creates a structural problem: the more successful your product becomes, the faster your costs accelerate.
- Month 1 (1,000 users/day): ~5 TB egress → ~£350/month. Manageable.
- Month 12 (100,000 users/day): ~500 TB egress → ~£35,000/month.
This is the success penalty — your cloud bill grows quadratically while your revenue grows linearly.
What unmetered 10Gbps bare metal actually means
"Unmetered bandwidth" means a dedicated 10Gbps physical port with no monthly data transfer cap. You pay a flat monthly server price inclusive of bandwidth.
The physical NIC advantage: On a cloud VPS, your network interface is virtualised and shared. Network contention from a neighbour directly degrades your throughput. On dedicated bare metal, you have exclusive access to the physical NIC. No hypervisor layer.
UK latency for real-time AI
For real-time voice and conversational AI, latency above 150ms ruins natural conversation flow. Server round-trip to a UK user from a US AWS region adds 80–120ms. Localized bare metal infrastructure in the UK measures 2–8ms to major UK ISPs, keeping interactions perfectly real-time.
Who should move to bare metal?
Good fit:
- AI apps with sustained, predictable traffic
- Workloads where egress volume exceeds ~20 TB/month
- Latency-sensitive applications serving a concentrated local user base
Better served by cloud:
- Early-stage prototypes with unpredictable traffic
- Workloads requiring rapid horizontal auto-scaling (spiky demand)
(Originally published on the eServers Engineering Blog)
Top comments (0)