--- title: "AI infrastructure buying guide - five stages, planned right" description: "A staged investment guide for sovereign AI infrastructure. Five stages from on-paper sizing through dedicated AI rooms, with realistic spend ranges and the principles that protect early hardware investment." url: "https://www.lmtek.com/sl/ai-solutions/ai-infrastructure-buying-guide" locale: "sl" --- Prevod v pripravi Ta stran še ni v celoti prevedena v slovenščino. Vsebina je trenutno prikazana v angleščini. [Odpri angleško različico →](/ai-solutions/ai-infrastructure-buying-guide) Buying guide · five stages # Start small. Leave room to scale. The cheapest mistake at Stage 1 is buying hardware that cannot host the next GPU generation. The most expensive mistake at Stage 3 is realising your facility was the constraint all along. Below is a five-stage buying guide for sovereign AI infrastructure - what each stage looks like, what it costs, and what you learn before the next one. The stages ## From on-paper to dedicated AI infrastructure. Stage 00 ### Scoping the first workload Pick ONE use case · estimate token volume, latency, concurrency · identify which data classes are involved → Answers: _"What's the smallest real workload we can run?"_ Footprint None yet - sizing exercise on paper Spend Zero hardware spend What you learn Workload profile (params, throughput, SLA) · the foundation for every later spend decision Stage 01 ### Starter node Single workstation or 1U/2U server · 1–2 GPUs · run quantised open-weight model (8B–30B class) · inference only · single use case · limited users → Answers: _"Does this work end-to-end inside our perimeter?"_ Footprint 1 box · single-room cooling Spend Low five figures EUR What you learn Real latency · real model quality on your data · real ops burden Stage 02 ### Production single-node Properly racked server · 2–8 GPUs · redundant power · supported OS · monitoring · backups · larger model or higher concurrency · 1–3 use cases sharing the box → Answers: _"Can we run this as a real internal service?"_ Footprint 1 rack unit in existing server room Spend Low to mid six figures EUR What you learn Where the bottleneck is - GPU memory, interconnect, storage IO, or cooling Stage 03 ### Multi-node cluster 2–8 nodes · high-speed interconnect · shared storage · workload scheduler · larger models · fine-tuning · multi-team access → Answers: _"Can we serve the whole company from this?"_ Footprint Dedicated rack(s) · liquid cooling starts paying off Spend High six to low seven figures EUR What you learn Whether your facility (power, cooling, network) is the next constraint Stage 04 ### Dedicated AI infrastructure Purpose-built environment · dense GPU racks · liquid cooling as default · capacity planning tied to roadmap · DR / failover · governance and audit at platform level → Answers: _"Is AI now a core piece of company infra?"_ Footprint Dedicated room or co-lo cage Spend Seven figures EUR and up · ongoing What you learn How to forecast and amortise - this is now a capex programme, not a project Design principles ## Four principles that protect your early spend. The decisions you take at Stage 1 either compound favourably or compound against you. These four principles - applied early - are the difference between a Stage 4 estate that grew gracefully from a Stage 1 box, and a Stage 4 estate that required throwing away a generation of hardware along the way. Principle 01 ### Buy chassis and cooling that can host the next GPU generation Today's 700 W GPUs are tomorrow's 1 000+ W GPUs. Air cooling at high density runs out of headroom faster than the GPU you bought. Specifying liquid cooling at Stage 1 is cheaper than rebuilding at Stage 3. Principle 02 ### Standardise on an inference stack portable across stages Whatever runtime, container layout, and observability stack you pick at Stage 1 should still work at Stage 4. The cost of porting a poorly-chosen runtime across three rack moves is high. Principle 03 ### Treat Stage 1 hardware as a lab asset after Stage 2 - not as throwaway The starter node becomes the development and evaluation environment once the production single-node is live. Plan for that from the start; the box keeps earning its keep. Principle 04 ### Decide the Stage 4 power and cooling envelope at Stage 2 The hardest constraint in scaling AI infrastructure is usually the facility, not the silicon. If you wait until Stage 3 to think about Stage 4 power and cooling, you have already missed the design window. The platform ## RM-4U8G is engineered for Stages 2 and 3. A production single-node deployment (Stage 2) and a multi-node cluster (Stage 3) are the same chassis, the same cooling loop, the same power topology. You upgrade by replacing GPUs and motherboard - the platform stays. That continuity is the buying-guide thesis in physical form. The Stage 2 starter - two RTX 6000 Ada GPUs in the same chassis - is the same Stage 3 maximum - eight H200 NVL GPUs. Same chassis, same cooling, no rebuild. ![RM-4U8G exploded view]() Reference figures ## What you would actually buy at each stage. Reference figures · examples drawn from real customer deployments. Final sizing is specified per workload - these are not bundles LM TEK sells, they are sketch BOMs to anchor the buying conversation. Stage Configuration sketch Indicative spend Stage 01 2× RTX 6000 Ada, 1× CPU, 256 GB RAM, 2× 3.84 TB NVMe ~€80k–€120k Stage 02 4× RTX A6000, 2× CPU, 512 GB RAM, 4× 3.84 TB NVMe ~€180k–€280k Stage 03 8× H200 NVL, 2× CPU, 1 TB RAM, 8× 7.68 TB NVMe (per node × 2–4 nodes) ~€800k–€2M