https://eurokommerz.at/blog/ Deep dives on AI infrastructure: GPUs, memory sizing, virtualisation and the parts vendor slides skip. en Mon, 14 Sep 2026 09:00:00 +0200 https://eurokommerz.at/blog/vmware-renewal-standard-vvf-vcf-which-tier/ https://eurokommerz.at/blog/vmware-renewal-standard-vvf-vcf-which-tier/ Mon, 14 Sep 2026 09:00:00 +0200 Standard and Enterprise Plus stop at vSphere 8 U3, 9.x lives only in VVF and VCF: features per tier, per-core rules, vSAN entitlements, and the renewal traps. https://eurokommerz.at/blog/rtx-pro-6000-vs-5000-vs-4500-blackwell/ https://eurokommerz.at/blog/rtx-pro-6000-vs-5000-vs-4500-blackwell/ Mon, 14 Sep 2026 09:00:00 +0200 96, 72, 48, 32 or 24 GB: memory, bandwidth, MIG and vGPU for each RTX PRO Blackwell card, which LLM fits where, and why CAD does not need the big one. https://eurokommerz.at/blog/nvidia-ai-enterprise-licensing-explained/ https://eurokommerz.at/blog/nvidia-ai-enterprise-licensing-explained/ Mon, 14 Sep 2026 09:00:00 +0200 Per GPU, not per VM: how NVIDIA AI Enterprise is counted, what ships with H100 NVL and H200 NVL, when vGPU compute needs it, and what runs without it. https://eurokommerz.at/blog/immutable-backup-what-it-protects-against/ https://eurokommerz.at/blog/immutable-backup-what-it-protects-against/ Mon, 14 Sep 2026 09:00:00 +0200 Hardened repositories and S3 Object Lock explained, the attacks immutability stops and the ones it does not, and how to size the lock window against dwell time. https://eurokommerz.at/blog/how-many-users-per-rtx-pro-6000/ https://eurokommerz.at/blog/how-many-users-per-rtx-pro-6000/ Mon, 14 Sep 2026 09:00:00 +0200 NVIDIA publishes no tokens/s for this card. The KV cache arithmetic, the 1,597 GB/s server figure and the vLLM log line that prints your own concurrency. https://eurokommerz.at/blog/h200-nvl-in-an-existing-server-checklist/ https://eurokommerz.at/blog/h200-nvl-in-an-existing-server-checklist/ Mon, 14 Sep 2026 09:00:00 +0200 Slot, sense pin, cooling, support list, NVLink, BIOS and licence: the seven gates that decide whether 141 GB of HBM3e can go into the server you already have. https://eurokommerz.at/blog/which-gpu-for-a-first-ai-project/ https://eurokommerz.at/blog/which-gpu-for-a-first-ai-project/ Mon, 14 Sep 2026 09:00:00 +0200 Model size, users at once, where the box lives, what follows the pilot: 273 GB/s against 1,597 GB/s and 4.8 TB/s, and which of the three a first project needs. https://eurokommerz.at/blog/fine-tuning-vs-rag-which-one/ https://eurokommerz.at/blog/fine-tuning-vs-rag-which-one/ Mon, 14 Sep 2026 09:00:00 +0200 RAG changes the input, fine-tuning changes the weights. What each one fixes, what full tuning, LoRA and QLoRA cost in VRAM, and what the papers really found. https://eurokommerz.at/blog/what-fits-on-a-dgx-spark-128gb/ https://eurokommerz.at/blog/what-fits-on-a-dgx-spark-128gb/ Mon, 14 Sep 2026 09:00:00 +0200 No VRAM figure, one shared pool: what the 128 GB really holds at BF16, FP8 and FP4, why 200 billion means 4-bit, and what one unit can actually fine-tune. https://eurokommerz.at/blog/dgx-b300-vs-8x-rtx-pro-6000-server/ https://eurokommerz.at/blog/dgx-b300-vs-8x-rtx-pro-6000-server/ Mon, 14 Sep 2026 09:00:00 +0200 2.1 TB of HBM3e with NVLink at 14 kW, or 768 GB of GDDR7 on PCIe in 4U: memory arithmetic, MLPerf v6.0 results, facility demands and the workloads each fits. https://eurokommerz.at/blog/backup-is-not-disaster-recovery/ https://eurokommerz.at/blog/backup-is-not-disaster-recovery/ Mon, 14 Sep 2026 09:00:00 +0200 A successful backup job is evidence about RPO, not RTO. What a restore leaves behind, how long 82 TB really takes, and what proven recovery has to show. https://eurokommerz.at/blog/gpu-server-for-70b-llm-worked-example/ https://eurokommerz.at/blog/gpu-server-for-70b-llm-worked-example/ Sun, 13 Sep 2026 09:00:00 +0200 Llama 3.3 70B, 30 users, 8k context: the memory arithmetic, why one 96 GB card only just works in FP4 and not in FP8, the H200 NVL option, the server around it. https://eurokommerz.at/blog/liquid-cooling-for-gpu-racks/ https://eurokommerz.at/blog/liquid-cooling-for-gpu-racks/ Sat, 12 Sep 2026 09:00:00 +0200 Where air stops (20 to 50 kW per rack), what rear doors, cold plates and immersion each carry, ASHRAE water classes, flow rates, CDU sizes, retrofit realities. https://eurokommerz.at/blog/how-many-gpus-fit-in-one-server/ https://eurokommerz.at/blog/how-many-gpus-fit-in-one-server/ Fri, 11 Sep 2026 09:00:00 +0200 Why a 2U takes four 350 W cards but two 600 W ones, where PCIe switches come in, what an eight-GPU server draws, and the number that stops most orders. https://eurokommerz.at/blog/dgx-spark-two-nodes-and-beyond-llms/ https://eurokommerz.at/blog/dgx-spark-two-nodes-and-beyond-llms/ Thu, 10 Sep 2026 09:00:00 +0200 How two units link, what NVIDIA supports up to four, measured two-node token rates, image and video generation, fine-tuning, and the honest answer on gaming. https://eurokommerz.at/blog/rtx-pro-6000-vs-h200-nvl-llm-inference/ https://eurokommerz.at/blog/rtx-pro-6000-vs-h200-nvl-llm-inference/ Wed, 09 Sep 2026 09:00:00 +0200 96 GB GDDR7 at 1.6 TB/s with FP4, or 141 GB HBM3e at 4.8 TB/s with NVLink. Model footprints, the bandwidth arithmetic behind tokens per second, MLPerf results. https://eurokommerz.at/blog/l40s-vs-rtx-pro-6000-server-edition/ https://eurokommerz.at/blog/l40s-vs-rtx-pro-6000-server-edition/ Tue, 08 Sep 2026 09:00:00 +0200 48 GB at 350 W or 96 GB at up to 600 W: memory, bandwidth, MIG, vGPU seats, server fit, and the cases where the older L40S is still the right order. https://eurokommerz.at/blog/rag-on-company-data/ https://eurokommerz.at/blog/rag-on-company-data/ Fri, 04 Sep 2026 09:00:00 +0200 Almost every RAG failure is a retrieval failure, not a model failure. Permissions at retrieval time, chunking and metadata, the evaluation set nobody builds, and the index that quietly goes stale. https://eurokommerz.at/blog/private-llm-vs-cloud-api/ https://eurokommerz.at/blog/private-llm-vs-cloud-api/ Thu, 03 Sep 2026 09:00:00 +0200 The break-even is a token-volume question. The four cost lines nobody counts, the arithmetic on your own numbers, and when residency decides first. https://eurokommerz.at/blog/vsphere-where-the-waste-is/ https://eurokommerz.at/blog/vsphere-where-the-waste-is/ Wed, 02 Sep 2026 09:00:00 +0200 Oversized VMs, zombie workloads, forgotten snapshots and stale cluster headroom: the four places capacity goes, how to measure them honestly, and why more vCPU often makes a VM slower. https://eurokommerz.at/blog/rpo-rto-how-to-choose/ https://eurokommerz.at/blog/rpo-rto-how-to-choose/ Wed, 02 Sep 2026 09:00:00 +0200 RPO is how much data you can lose, RTO how long you can be down. Why cost climbs steeply near zero, how to tier systems, the dependency order that breaks plans. https://eurokommerz.at/blog/vmware-licensing-after-broadcom/ https://eurokommerz.at/blog/vmware-licensing-after-broadcom/ Tue, 01 Sep 2026 09:00:00 +0200 How VCF and vSphere Foundation cores are counted, what the 16-core minimum does to a small cluster, what the 72-core story really was, and where the number gets larger than it needs to be. https://eurokommerz.at/blog/gpu-rack-power-and-cooling/ https://eurokommerz.at/blog/gpu-rack-power-and-cooling/ Mon, 24 Aug 2026 09:00:00 +0200 Rack density in kW, the airflow arithmetic, the point where air cooling stops and liquid starts, and the four checks that decide whether a delivery switches on. https://eurokommerz.at/blog/nvidia-dgx-spark-real-benchmarks/ https://eurokommerz.at/blog/nvidia-dgx-spark-real-benchmarks/ Sun, 09 Aug 2026 09:00:00 +0200 NVIDIA’s own figure for gpt-oss 120B: 55 tok/s generation, 1,725 tok/s prefill. The 273 GB/s ceiling, 128 GB unified memory, 140 W TDP, two-node pairing. https://eurokommerz.at/blog/nutanix-gpu-ai-cluster/ https://eurokommerz.at/blog/nutanix-gpu-ai-cluster/ Wed, 05 Aug 2026 09:00:00 +0200 Which NX nodes (G9 and G10) take L40S, H100 NVL, A16 and L4, how many cards each holds, what AHV supports, and why power and airflow, not case size, block the next card. https://eurokommerz.at/blog/rtx-pro-6000-blackwell-vs-rtx-6000-ada/ https://eurokommerz.at/blog/rtx-pro-6000-blackwell-vs-rtx-6000-ada/ Tue, 28 Jul 2026 09:00:00 +0200 96 GB GDDR7 vs 48 GB GDDR6, 1,792 vs 960 GB/s, FP4 and MIG that Ada never had, and the cases where RTX 6000 Ada still does the job. https://eurokommerz.at/blog/rtx-pro-6000-workstation-max-q-server-comparison/ https://eurokommerz.at/blog/rtx-pro-6000-workstation-max-q-server-comparison/ Mon, 20 Jul 2026 09:00:00 +0200 Same 96 GB silicon, three power envelopes: 600 W, 300 W and 400–600 W. Which one takes vGPU, how many VMs per card, and which belongs in a server. https://eurokommerz.at/blog/mig-vgpu-how-many-vms-per-gpu/ https://eurokommerz.at/blog/mig-vgpu-how-many-vms-per-gpu/ Fri, 10 Jul 2026 09:00:00 +0200 MIG splits the silicon, vGPU splits time and needs a per-user licence. Maximum instances and smallest slice for each card, and where the setup breaks. https://eurokommerz.at/blog/h200-nvl-vs-h100-nvl/ https://eurokommerz.at/blog/h200-nvl-vs-h100-nvl/ Sun, 28 Jun 2026 09:00:00 +0200 Identical compute silicon, all the difference in memory: 141 vs 94 GB and 4.8 vs 3.9 TB/s. Where the gain reaches 3.4×, and where it is exactly zero. https://eurokommerz.at/blog/how-much-vram-llm-needs/ https://eurokommerz.at/blog/how-much-vram-llm-needs/ Mon, 15 Jun 2026 09:00:00 +0200 Weights, KV cache, CUDA-graph overhead and concurrent users: the working formulas and sizing tables for choosing a GPU for local LLM inference.