What I'd Buy in 2026 for €15,000: A Pro-Studio Sovereign AI Build

Here is what I would buy at €15,000 today, knowing what I know in 2026-05. This is the tier where a serious one-person AI consulting practice or a small studio with two or three full-time engineers actually sits. Below this the workloads start to feel pinched; above this you are buying margin or vanity rather than capability. There are three honest paths and they answer different binding constraints. I run a single DGX Spark at the €8k tier and the configuration below is what I would buy if my consulting practice grew to the point where the Spark stopped being enough. I have not yet operated this scale of build personally, so I will mark the per-component reasoning as architecture-correct rather than measured.

Path A is dual RTX 5090 on a Threadripper Pro workstation, optimized for parallel inference jobs and serious dense-model fine-tuning. Path B is DGX Spark plus a dedicated inference second box (a 5090 desktop), optimized for serving MoE in production while doing image and dense work next door. Path C is a refurbished pro-workstation route (Lenovo ThinkStation P-series, Dell Precision) with dual RTX A6000 Ada or used cards, optimized for buyers who want the enterprise warranty path and predictable resale.

Path A: dual RTX 5090 on Threadripper Pro

To land at €15,000 exactly: drop one of the two 4 TB NVMes to a 2 TB drive and rely on external backup (saves roughly €340), or step the RAM to 192 GB (saves roughly €700). I would step the RAM. 192 GB ECC is enough for a dual-5090 setup because the binding constraint is per-card VRAM, not system memory.

Path B: Spark plus inference second box

Path B is my recommended path for buyers whose business model is “MoE inference in production, plus a strong desktop for everything else.” The Spark handles the customer-facing inference workload; the 5090 desktop handles the image generation, the video work, the model conversion experiments, the second-team-member’s local development environment, and the workloads where the Spark is genuinely weak.

Path C: refurbished pro-workstation

I have not priced this path in detail because the secondhand pro-workstation market (Lenovo ThinkStation P620 / P7, Dell Precision 7960) is highly variable and the listings change weekly. The shape of the configuration is: refurbished P620 or P7 chassis with Threadripper Pro and 256 GB ECC, then add one or two NVIDIA RTX A6000 Ada 48 GB cards (€6,800 list retail per card per Network Outlet’s 2026 listing, with used/refurbished Ada cards trending €4,000 to €5,500 on eBay Europe completed listings in 2026-05). A single-A6000-Ada P620 refurb lands roughly €10,000 to €12,000; a dual-A6000-Ada lands closer to €18,000 to €22,000. Path C is the right answer for buyers who want the warranty story, the data-center thermal design, and the enterprise resale market. It is over budget at the dual-card configuration and slightly under budget at the single-card configuration.

Prices captured 2026-05-22 from Geizhals.de, Geizhals.eu, eBay completed listings, and Network Outlet. They will drift. Re-verify before you buy.

Why each path

Path A: dual 5090. Two 32 GB cards plus tensor parallelism puts 64 GB of fast VRAM on a single workload. A Llama 3.1 70B at FP16 fits cleanly across both cards. Mistral Large dense fits at moderate quant with headroom. Real LoRA fine-tuning becomes feasible on the 7B to 13B class without renting cloud GPUs. The two-card NVLink-less topology of consumer Blackwell is the operational catch; tensor parallelism via PCIe 5.0 is fine for inference and only middling for training. The 7965WX gives you 128 PCIe lanes which is the right shape for two GPUs plus two NVMe drives plus 10 GbE plus expansion headroom. For the reasoning on why I would not go to four cards at €15k (heat, power, blower-card noise, depreciation risk), see the trade-offs in Should You Buy a DGX Spark in 2026?.

Path B: Spark plus second box. This is the path I would buy if I were scaling up from my current single-Spark setup. The Spark stays as the production inference target (MoE 100B+ at ~45 tok/s per Spark Arena Rank 4 Made Me Add Qwen3.6), the second box becomes the workstation where you write code, generate images, convert quantizations, and run the experiments you do not want the production endpoint to host. The architecture-correct split is “serving on the Spark, development on the 5090.” The cost of the split is roughly €5,000 over a single-Spark setup, which is the right cost to pay if your business model is “customer pays me for the fact that inference never leaves my premises.”

Path C: refurbished workstation. This is the path for buyers who are uncomfortable with consumer-card depreciation curves and want the enterprise resale story. A ThinkStation P620 or Dell Precision 7960 holds its value better than a homebuilt Threadripper Pro rig and the warranty paths through Lenovo or Dell are real. The cost is paying enterprise prices for last-generation Ampere cards. For most one-person consultancies this is overkill; for small studios that bill enterprise customers, the warranty story alone can justify the markup.

Why the components, the short version

Threadripper Pro 7965WX over 7955WX. 24 cores versus 16 cores. The €900 cost delta is the right tier for a consultancy box that also runs parallel inference jobs, document-processing pipelines, and the occasional fine-tune. For pure inference, the 7955WX is enough; for the workloads that bottleneck on CPU (chunking, embedding generation, retrieval index builds), the extra cores matter. The 7965WX gives 128 PCIe lanes either way.

256 GB ECC DDR5 RDIMM. ECC matters at this tier because the box is going to run for months between reboots and silent memory errors corrupt model weights in ways that are hard to debug. The 256 GB capacity is the threshold below which you cannot comfortably hold a large model in system RAM as a CPU-offload fallback when a workload spills. The ECC RDIMM is non-negotiable at this tier; non-ECC DDR5 saves €600 and costs you one weekend of debugging in year two.

Mirrored 4 TB NVMe. Models are the asset. Losing the model directory to a single-drive failure is a multi-day recovery from external backup. A simple ZFS mirror or mdadm RAID 1 across two NVMe drives makes the single-drive failure a non-event. The cost is one extra €619 drive. For the backup-without-bankruptcy story at the next layer down (what to do when the mirror itself fails), see Backing Up 119B Parameters Without Bankruptcy.

1600 W ATX 3.1 PSU (Path A). Two 5090s under full load can pull 1200 W combined. Add the Threadripper Pro at 350 W, two NVMe drives at 20 W, fans and RAM at 50 W. 1600 W is the safe headroom; 1200 W will work for steady-state but trips on the transient spikes that happen when both cards ramp simultaneously. ATX 3.1 means the 12V-2x6 connector is native; no dongles.

UPS sized for graceful shutdown. The Power Failure Recovery: DGX Spark in 30 Minutes procedure assumes a UPS that can hold the box up for at least eight minutes under load while the model unloads cleanly and the journals flush. At the dual-5090 power envelope (1,500 W under sustained load), the math wants a 2000 to 2200 VA online double-conversion UPS rather than the cheaper line-interactive boxes. The online UPS is also better behaved during brownouts, which are more common in Germany than the marketing for backup power supplies admits.

What this runs, what it does not

Path A runs well: parallel inference jobs (two model endpoints simultaneously, each with 32 GB headroom), dense Llama or Mistral 70B at FP16 with tensor parallelism, real LoRA fine-tunes on the 7B to 13B class, large-context inference (128k+ context fits), image generation at production resolution on either card independently. Does not run well: 119B+ MoE language models in production (the Spark is still the architecture-correct answer for that class), full-parameter fine-tuning on 70B+ models without quantization-aware tricks (the consumer-card NVLink-less topology is the constraint).

Path B runs well: the union of “Spark workloads” and “5090 workloads” with operational separation. The Spark serves the customer-facing inference; the 5090 box handles image, video, model conversion, quantization experiments, development. The split lets you upgrade one without disrupting the other. Does not run well: workloads that need 64 GB of VRAM in one model with tensor parallelism (Path A is the architecture-correct answer for that class).

Path C runs well: the workloads that fit the A6000 Ada’s 48 GB VRAM per card, with enterprise warranty and predictable resale. Does not run well: any of the Blackwell-only features (no NVFP4 path on Ada Lovelace, no MXFP4, no current-gen FLOPS for diffusion).

For the model-class trade-offs that decide between the three paths, see Mistral Small 4 vs Qwen 3.6 vs GLM 5 on DGX Spark and NVFP4 Quantization Explained. For the workstation-versus-server framing at this tier, see DGX Spark vs Mac Studio for Local LLMs; the same architecture-pivot applies one tier up.

Monthly power cost, three jurisdictions

Path A averages roughly 350 W under realistic mixed-use (eight hours active, sixteen hours idle), which is 256 kWh per month. Path B averages roughly 250 W combined (Spark idle plus 5090 active during work hours), which is 183 kWh per month. Path C depends heavily on which Ada card configuration; single-card lands roughly 200 W average (146 kWh), dual-card lands roughly 320 W (234 kWh).

Hardware amortization over three years is €417 (Path A €15k) per month. Power adds €18 to €87. Total cost of operation runs €435 to €505 per month, which is the right scale for a consultancy that bills €4,000 to €15,000 per month in revenue. Below that revenue floor, you are buying margin you do not yet have; the self-hosted-vs-cloud cost model makes the break-even visible.

Compare to the other tiers

Below this tier, the €8k premium build is the right answer for a one-person practice that is at or near the 1,000-calls-per-day threshold but not yet running parallel workloads. The €4k mid-tier build is the right answer for buyers whose primary workload still fits in 48 GB on a single card. The €2k beginner build is the entry point for buyers measuring their workload before committing to the architecture.

The €15k tier specifically makes sense when one of three statements is true: you have parallel inference jobs that need true concurrency rather than queuing; you are doing real fine-tuning rather than LoRA-only experiments; or you are running a consultancy where one workstation is the production inference target and a separate box is the development environment. If none of those three is true, the €8k tier is probably the architecture-correct answer for less money.

For the broader strategic framing on why a one-person sovereign-AI practice scales to this tier at all, see the comparison anchor in Self-Hosted AI vs Cloud APIs: Real Total Cost. The year-one revenue retrospective is in draft and will be published once the year has actually closed; the consulting-pricing piece is in draft until a stable cohort of paid engagements has shipped.

If I had it to do again

I have not yet operated this tier of build personally, so the “if I had it to do again” paragraph is forward-looking rather than retrospective. The discipline I would impose on myself: do not buy the second card until the first card is genuinely saturated. Most operators at this tier discover that a single 5090 or a single Spark is at 30 to 40 percent utilization most days and the second card sits idle. If you can measure six months of sustained 70 percent utilization on the first card before buying the second, the parallel-card argument is honest. Without that measurement, the second card is a vanity purchase.

The other discipline is the observability layer. At this tier the failure modes (a card fan dies, the PSU drops a rail, the UPS battery degrades, the network 10 GbE link flaps) all benefit from monitoring that you set up before the failure rather than after. The pattern lives in Self-Hosted Observability: The One-Person AI Stack; the punchline is that a Prometheus plus Grafana plus Loki stack on the second box is the cheapest insurance you will ever buy at this tier.

For the operational disciplines that decide whether a €15k build pays for itself or sits idle, see Five DGX Spark Disasters I Survived and The Quiet Pattern Among Sovereign Engineers. Both articles apply to the dual-5090 path even though they were written from the Spark side; the failure modes are architecturally similar.

Book a Stack Audit

If you want a second pair of eyes on whether Path A, Path B, or Path C matches your actual workload and revenue, the Stack Audit is two hours, fixed-fee, ends with a configuration recommendation and a power-cost projection for your jurisdiction. About a quarter of audits at this tier end with “buy the €8k build instead, you do not yet have the workload to justify the second card.” The honesty is the product.

To discuss your workload, use the contact strip in the footer. Or read the €8k version, the €4k version, or the €2k version if you are sizing down rather than up.