The magic of dedicated GPU hosting for a growing product is amortisation. £500/month feels significant at 10 customers. At 1,000 customers, it’s 50p per customer. Understanding this curve helps with pricing and growth planning.
Contents
The Curve
For a single £500/month server serving up to 10,000 customers at realistic utilisation:
| Customers | GPU cost/customer/month |
|---|---|
| 10 | £50.00 |
| 100 | £5.00 |
| 1,000 | £0.50 |
| 5,000 | £0.10 |
| 10,000 | £0.05 |
At scale the GPU cost per customer is negligible as a fraction of subscription revenue. This is why SaaS margin expands with scale.
Capacity
One server handles a finite customer count. Plot expected requests per customer against server throughput to estimate the ceiling. When you approach 60-70% utilisation, provision the next server before peak traffic hits the ceiling.
Doubling capacity (two servers) keeps per-customer cost flat as you scale – amortisation continues linearly.
Planning
Rule of thumb: one dedicated 5090-class server comfortably supports 2,000-5,000 concurrent active users for chat workloads. Adjust based on your average requests per user.
Plan for the server count you need at 1.5x your forecast customer count. Provisioning a new server is fast (days, not weeks), but overcommitted capacity affects SLAs.
