What I'd Buy in 2026 for €4,000: A Mid-Tier Sovereign AI Build

Here is what I would buy at €4,000 today, knowing what I know in 2026-05. There are two honest paths at this budget, and the choice is binding on workload rather than aesthetics. Path A is a new RTX 4090 24 GB on an upper-mid AM5 platform with 128 GB of DDR5, optimized for throughput on dense and MoE language models that fit in 24 GB. Path B is a used RTX A6000 48 GB on a Threadripper-class workstation board, optimized for the model classes that need more than 24 GB in one card.

I have not personally tested either build end to end. I run a DGX Spark at roughly the same euro outlay, which is the third honest option at this price point and gets its own comparison below. The Path A and Path B picks are conservative, sourced from current Geizhals listings, and the prices below are captured 2026-05-22.

Path A: new 4090 build

That total is over budget at the 4090’s current Geizhals floor of €2,690. To come in under €4,000, drop the RAM to a single 64 GB kit (saves €630) and drop the NVMe to a 2 TB Samsung 990 PRO at roughly €280 (saves €220). Adjusted total: €4,389. Still over budget by €389. To land at €4,000 exactly, you either accept a slightly lower-end 4090 SKU (the floor moves week to week), buy the GPU on a Mindfactory sale (historically a 5 to 10 percent discount window appears monthly), or accept that the build is €4.4k rather than €4.0k. I prefer the third option; the budget envelope is not the constraint that matters, the workload-fit is.

Path B: used A6000 build

The used A6000 is the price-mover. eBay completed listings range from roughly €2,200 to €2,800 in 2026-05, with the low end being cards from data-center decommissioning and the high end being lightly-used workstation pulls with the original box. The 2 TB NVMe is estimated because the current Geizhals listing for the 4 TB at €499.99 implies the 2 TB at approximately €280; I have not pulled a specific 2 TB SKU’s current price and want to mark that line honestly as estimate, verify before buying.

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

Why each pick, the short version

4090 over 5090 at this tier. The 5090 at €3,469 to €3,889 is faster but pushes the build well past €5k for the same VRAM envelope. The 4090 is the price-correct dense-inference card at €2,690 because the per-token throughput delta to the 5090 does not justify the €800 to €1,200 cost delta unless you are specifically planning to use NVFP4 quantization. For the NVFP4 trade-offs see NVFP4 Quantization Explained; short version, the format is real and the speedup is real, but it is a Blackwell-only path and 24 GB caps you well below the model classes where NVFP4 actually changes the workflow.

A6000 48 GB over 4090 24 GB. This is the workload-fit pivot. Models that need more than 24 GB in one card become first-class citizens. Llama 3.1 70B at Q6 or Q8 quantization fits in 48 GB with headroom for context. Mistral Large dense fits at moderate quant. Fine-tuning small LoRAs has scratch space. The A6000 is the cheapest path to 48 GB of NVIDIA VRAM in a single card; the alternative is two 3090s with NVLink, which is mechanically feasible but operationally noisier and harder to cool.

7800X3D over 7700. At this tier the €100 cost delta is rounding error and the 3D V-cache helps the rare workloads that mix gaming with inference on the same box. If this is strictly an inference workstation, drop to the €2k tier’s Ryzen 7 7700 pick and pocket €110. I included the X3D here because the readers writing in at the €4k tier more often run mixed workloads (one box, used for both day-job development and inference experimentation).

128 GB DDR5. Two times the €2k build’s RAM. The reason is the second card and the room for CPU-offloaded layers when a model just barely overflows VRAM. 128 GB is the threshold below which model loading at the 70B+ class starts to feel slow because of page-cache churn.

4 TB NVMe (Path A) or 2 TB NVMe (Path B) plus a 4 TB SATA backup. Models accumulate fast. At this tier I assume you are running three to five models concurrently, each 30 to 80 GB on disk. The 4 TB primary plus an external backup is the smallest config that does not constantly trip over itself. The backup-without-bankruptcy approach lives in Backing Up 119B Parameters Without Bankruptcy; the same logic applies one tier down.

850 W PSU. A 4090 alone pulls up to 450 W under load. The A6000 is gentler at roughly 300 W. The 7800X3D plus board plus drives add 130 W headroom. 850 W is the safe floor for the 4090 path; the A6000 path could drop to 750 W but the saved cost is €30 and the headroom is worth it.

Path A versus Path B versus Spark at €4k

The Spark at €4,769 from the NVIDIA Founders listing on Geizhals is the third honest option and gets its own deep-dive in the Spark decision tree. The decision among the three pivots on whether your model roadmap is dense (Path A or Path B) or MoE (Spark), and whether you want a single box that you administer as a Linux server (Spark) or a desktop with a discrete card (Path A and Path B). See also the DGX Spark vs Mac Studio comparison for the workstation-versus-server framing.

What this runs, what it does not

Path A runs well: Llama 3.1 70B at Q4 (very fast), Mistral Small 3.x at FP16, Qwen 30B-class at Q8, Stable Diffusion XL and Flux at production resolution, dense models up to the 24 GB ceiling. Does not run well: Qwen 3.6 119B MoE (the active-parameter footprint plus the routing table do not fit in 24 GB cleanly), Mistral Large dense at usable quant, anything labelled 100B+ dense.

Path B runs well: Llama 3.1 70B at Q6 or Q8 (clean), Mistral Large dense at Q4, dual-model serving (a 7B plus a 70B on the same card), small LoRA fine-tunes (real, not symbolic). Does not run well: 119B MoE class with full context (still spills the routing table to system RAM), latest Blackwell-only quantization formats (no NVFP4 path on Ampere).

For the model-class trade-offs that decide which path wins for your workload, see Mistral Small 4 vs Qwen 3.6 vs GLM 5 on DGX Spark. The relative model rankings translate; the absolute throughput numbers do not because those were measured on Blackwell.

Monthly power cost, three jurisdictions

The 4090 inference-idle is roughly 20 W. Under load it pulls 400 W to 450 W. The A6000 idles at 15 W and loads to 280 W to 300 W. A realistic mixed-use profile (eight hours active, sixteen hours idle) averages around 180 W for Path A and 140 W for Path B. I will use 180 W as the conservative centerline; that is 131 kWh per month.

Hardware amortization over three years is €112 (Path A €4,000 envelope) to €126 (Path B €4,500 envelope) per month. Power adds €9 to €45. Total cost of operation: €120 to €170 per month, still well below cloud-API for sustained workloads. The break-even math is in Self-Hosted AI vs Cloud APIs: The Real Total Cost.

Compare to the other tiers

Below this tier, the €2k beginner build is the right answer for workloads that fit in 24 GB and do not need new-card warranty coverage. Above this tier, the €8k premium build is the right answer for sustained MoE workloads and for the operators who want the Spark’s unified-memory architecture. The €15k pro-studio build is the floor for two-card parallel jobs and serious fine-tuning.

If I had it to do again

The mistake I see most often at this tier is buying Path A when the workload was Path B (or vice versa). The trap is the GPU’s VRAM number on the spec sheet, which the buyer treats as a binary check (does the model fit yes or no) when it is actually a continuous variable (how much context, what quant, what batch size, what serving framework). Spend two evenings before you buy this build doing a paper exercise on three specific models you intend to run, at three specific quantization levels, with three specific context lengths. If all nine cells fit in 24 GB, Path A is correct. If three or more cells need 48 GB, Path B is correct. If any cell needs 80 GB or more, you are in the €8k tier and the €4k tier is going to disappoint.

The other discipline is to read Five DGX Spark Disasters I Survived before buying any of these paths. The disasters are operational, not architectural; they happen to every local-inference box, not just the Spark. Knowing what they look like in advance saves at least one weekend.

Book a Stack Audit

If you want a second pair of eyes on which of Path A, Path B, or Spark matches your actual workload, the Stack Audit is two hours, fixed-fee, ends with a configuration recommendation. About a third of audits end with “rent cloud for six months, here is what to measure.” The honesty is the product.

Contact via the footer (Nostr or email). Or read the €8k version next if your workload is past the 48 GB ceiling.