Not NVIDIA this time: why we ordered AMD GPUs for our next AI factories

Until now, there was no way around NVIDIA for us either. The RTX PRO 6000 (Blackwell) are top GPUs – and still a good choice if you want to stay in the NVIDIA ecosystem. Our first AI factories run on them. For the next ones we have now deliberately ordered AMD – Supermicro servers with 8× AMD Instinct MI350P. Why not NVIDIA as the base this time? Three reasons:

• Memory (VRAM). 144 GB per GPU, 1,152 GB in the server. VRAM decides which models run – with AMD, significantly larger models fit on fewer GPUs.
• Runs in a normal data center. Air-cooled, no hyperscaler site required. We build AI factories for mid-market companies that want sovereign AI – near peak quality, without a hyperscaler data center.
• Future-proof hardware. Our customers’ models change constantly. The large memory also carries tomorrow’s models: the hardware stays, the models get better.

In short: what GEKI does

GEKI builds and operates AI factories for inference – for companies that want to use AI regionally and cheaply for their own applications: without hyperscaler supplier risk, without their data in US data centers, with full data sovereignty – but that don’t want to deal with hardware, the inference stack or the data center.

You get OpenAI-compatible API endpoints from us – the convenience of a hyperscaler, only in Austrian and German data centers. You choose the location. We run the AI factory 24/7 and manage models, routing and scaling. You use AI for your applications.

What we ordered

A Supermicro server with 8× AMD Instinct MI350P (144 GB HBM3e each), AMD EPYC CPUs, 1 TB of RAM and fast NVMe storage. That’s 1,152 GB of GPU memory in a single machine.

Main reason: VRAM

Modern AI models need one thing above all: a lot of VRAM. Size decides which model can run at all – that was the main reason for AMD. The MI350P also uses HBM3e, not GDDR like the RTX PRO 6000: significantly more bandwidth. Example: DeepSeek V4 Flash (284B parameters) needs about 160 GB just for the weights. It does not run on a single RTX PRO 6000 (96 GB). On two of them (192 GB) it gets too tight for multiple requests at once. With 2× MI350P (288 GB) there is enough headroom.

(More in our VRAM comparison: how much memory Qwen, DeepSeek and Kimi really need.)

Runs in a normal data center

We deliberately chose the air-cooled PCIe variant (MI350P). Each GPU draws about 600 W – so the whole server runs in a normal data center rack, with no special liquid cooling and no hyperscaler site required. That’s exactly why we can offer more regional locations.

That’s our point: we build AI factories for mid-market companies that want sovereign AI – near peak quality with open models, but without a hyperscaler data center. The hardware sits where your data should be: in an Austrian or German data center – you choose the location.

Future-proof hardware

The models our customers run change constantly – every few months a better open model arrives. The large memory of the MI350P is therefore an investment in the future: the same factory carries DeepSeek V4 Flash or GLM 5.3 today – and tomorrow’s next, larger generation. The hardware stays, the models get better.

From one GPU to a full server

Through GEKI, the right configuration is always available – from a single GPU to a full server. Entry already from 1 GPU. Our recommendation (all open models):

GPUs	Example models	Why it fits well
1	Qwen3.8 27B	compact, a strong model on little hardware – a good entry
2	DeepSeek V4 Flash	strong agent model, open – a lot of performance for the money
4	GLM 5.3 Flash	more speed and quality for demanding workflows and more parallel users
8	GLM 5.3 (full version)	near-frontier quality for the most demanding applications

You tell us what you want to run – we deliver the right configuration.

The point

Hardware alone is interchangeable and will keep changing. The difference is the software layer on top – inference routing, utilisation, multiple models. It turns cost-efficient hardware into an economical AI factory. That’s exactly why I could choose AMD this time: because our operations deliver the efficiency, not the most expensive GPU.

GEKI operates fully managed AI factories with open models in Austrian and German data centers. The convenience of a hyperscaler (OpenAI-compatible API endpoints), but local, sovereign and affordable. You take care of your application – we take care of everything underneath.