For AI developers, cloud solutions offer flexibility and speed in the early stages. But as workloads grow, they often create unpredictable bills, inefficiencies, and vendor lock-in. Increasingly, AI companies are approaching Punch Technology to explore custom-built AI servers as a more cost-effective alternative.
This is the exact situation that led to us recently completing a project for a UK-based AI developer. After reading the 37signals article, ‘Why we’re leaving the cloud’, this fast-growing AI startup realised that they too were quickly approaching the tipping point, where the benefits of readily available, quick to start, cloud solutions were being outweighed by rapidly rising costs, and decided to explore a Punch custom server build.
Our sales and customer service team worked to consult with the client on their requirements and the possible solutions. We knew that we needed to offer a system which offered value as a one-off investment vs. ongoing cloud costs, that it needed to be scalable and appropriate for on-premises hosting. We also knew that we would be able to offer a solution that was fully tuned for specific AI workloads, a distinct advantage over the more generalist cloud hardware.
The client’s requirements were primarily GPU power and VRAM, which meant the solution could go one of two ways, either; multiple GPUs in one system, requiring custom water block cooling solutions and scaling the cost massively, or using fewer GPUs, with more VRAM per card to offer the equivalent processing power.
After researching GPU options, and comparing benchmark results, we opted for Nvidia RTX Pro 6000 Blackwell Workstation editions. Given the customer’s specific use case, their AI application actually works better with more VRAM per card, rather than a lot of VRAM split across a lot of cards, which is why the 96GB Pro 6000’s made for a great option.
The Nvidia RTX Pro 6000 Blackwell Workstation edition outperforms all its nearest competition in Blender benchmarks – not directly related to the AI application, but a good general indication of performance
The solution we delivered can handle the requirements of our customers application and is scalable with additional machines. Not only that but we were able to more than halve the customers recurring cloud solution costs, saving them the best part of £250,000 over the next 5 years.
“Our cloud costs were spiralling. Punch helped us design a server that not only cut our expenses but also improved performance for our models.”
UK-based AI developer
Cloud solutions are perfect for prototyping and rapid setup, but as AI workloads stabilise and scale, owning your infrastructure often delivers better economics and performance. Punch Technology helps developers make that transition with confidence by designing and building custom AI servers that save money, reduce complexity, and unlock full control. Talk to us today to see how much you could save by moving beyond the cloud.
| RESULTS AT A GLANCE |
|---|
| £250,000 saved over 5 years compared to cloud costs |
| Performance tuned specifically for AI workload (VRAM-heavy tasks) |
| Scalable design with future machines |
TECHNICAL BUILD SUMMARY
Corsair Obsidian 1000D Case
AMD Ryzen Threadripper Pro 9975WX CPU
ASUS WRX90E-SAGE SE Motherboard
NVIDIA RTX PRO 6000 96GB CPU
NVIDIA RTX PRO 6000 96GB CPU
128GB DDR5 RDIMM 4800MHz Memory (4x32GB)
1TB WD Black SN7100 M.2 PCIe 4.0 NVMe SSD
4TB Crucial T705 M.2 SSD
4TB Crucial T705 M.2 SSD
be quiet! Silent Loop 3 420mm all-in-one liquid cooler
3300W Dual PSU
Windows 11 Professional
10 x Silent Wings Pro 4 PWM120mm
