In August 2025, Andreessen Horowitz (a16z) published a detailed article about their personal AI workstation: a Threadripper Pro machine equipped with four NVIDIA RTX 6000 Pro Blackwell Max-Q GPUs, each with 96 GB of VRAM. In total, that’s 384 GB of GPU memory under one desk – the kind of setup normally reserved for datacenters.

The price? Easily 35 000 €–40 000 € in Europe for the GPUs alone, with a full system reaching +50 000 €. For startups, independent researchers, or small consulting firms, that’s a prohibitive cost.

With Sentra Brain, I design and deliver private AI systems for clients who need serious local AI power – for fine-tuning, inference, or running private agent frameworks – but at a realistic budget. That’s why I explored: how can we build a “cheap” (or at least cheaper) AI workstation in Spain/Europe that still delivers 4-GPU performance?

The Target: A “Cheap” 4-GPU AI Workstation

The goal is not to match a16z’s monster spec euro for euro, but to replicate the concept: four high-VRAM GPUs, a many-core CPU with enough PCIe lanes, plenty of ECC memory, fast NVMe storage, and a PSU/case that can handle it.

Instead of spending 50 000 €, we aim for something in the 14 000 €–16 000 € range, while keeping performance suitable for real-world LLM training and inference.

This is the kind of Sentra Brain-ready workstation I would recommend to clients who want on-prem AI without cloud costs or data leakage risks.

Step 1: Choose the Right GPUs (the Heart of the Build)

• Option A (Recommended): NVIDIA RTX A6000 48 GB (Ampere)

  • Used/refurbished market in Europe: 2 600 €–3 000 € each

  • Four cards = ~11 000 €

  • 192 GB total VRAM, blower design (perfect for 4-GPU stability)

  • ECC memory, NVLink support, designed for 24/7 workstation duty

• Option B (Ultra Budget): NVIDIA RTX 3090 24 GB

  • Used market: 700 €–800 € each

  • Four cards = ~3 000 €

  • 96 GB total VRAM, but open-air cooling and higher power draw make 4-GPU use tricky

• Option C: NVIDIA RTX 4090 24 GB

  • New: ~1 700 €–1 900 € each in Spain

  • Four cards = ~7 000 €

  • Extreme performance, but power draw and cooling are problematic without dual PSUs or water-cooling

Verdict: For professional, stable use, the RTX A6000 (48 GB) is the sweet spot. Half the VRAM of Blackwell 96 GB, but at one-third the cost.

Step 2: CPU & Motherboard (PCIe Lanes Matter)

• CPU: AMD Threadripper PRO 5975WX (32 cores)

  • Used/OEM market: ~1 300 €–1 500 €

  • 128 PCIe 4.0 lanes, 8-channel memory – critical for four GPUs at full x16 bandwidth.

• Motherboard: WRX80 workstation board (Asus Pro WS WRX80E, Supermicro M12SWA, etc.)

  • Used: ~500 €–700 €

  • Supports 4–7 PCIe x16 slots, E-ATX/SSI-EEB size.

Step 3: Memory (ECC for Stability)

• 128 GB DDR4 ECC Registered (8×16 GB, 3200 MT/s)

  • Market price in Spain: ~350 €–400 €

  • 8 channels fully populated = maximum bandwidth.

• Optional upgrade: 256 GB (8×32 GB) if you handle very large datasets (~800 €).

Step 4: Storage (Fast, but Affordable)

• 2× 2 TB NVMe Gen4 SSDs (Samsung 990 Pro, WD SN850X, etc.)

  • ~150 € each → 300 € total

  • Either stripe (RAID0) for 14 GB/s throughput or split OS vs data.

• Optional bulk: 4 TB SATA SSD or HDD for archives (~200 €–300 €).

Step 5: Power, Cooling, and Case

• PSU: Corsair AX1600i (1600 W, 80+ Titanium) → ~450 €–500 €

• Cooling: 360 mm AIO for CPU (Arctic LF II 420 TR4, etc.) → ~180 €

• Case: Phanteks Enthoo Pro 2, Corsair 7000D Airflow, or Fractal Define 7 XL → ~200 €–250 €

• Extra fans (high airflow, e.g., Arctic P14) → ~80 €–100 €

Total Cost (Spain/Europe, Q3 2025)

Component	Estimated €
4× RTX A6000 48 GB (used)	11 000 €–12 000 €
CPU (TR Pro 5975WX)	1 400 €
Motherboard (WRX80)	600 €
Memory (128 GB ECC)	400 €
Storage (2× 2 TB NVMe)	300 €
PSU (1600 W)	500 €
Cooling + Case + Fans	500 €
Total	14 500 €–15 500 €

Compare that to the 50 000 € a16z setup – we’re at ~30 % of the cost while keeping ~70–80 % of the capability.

Trade-Offs

• VRAM: 192 GB vs 384 GB (but still enough for most LLMs with quantization).

• PCIe 4.0 vs PCIe 5.0 (no real loss for current GPUs).

• Used components (less warranty, but massive cost savings).

• Noise & heat: a 1.5 kW system will sound like a server.

Why This Matters for Sentra Brain

Many SMEs, law firms, or advisory firms in Spain ask me the same thing:

“Can we run large AI models on-premise, without paying cloud bills or sending client data outside?”

The answer is yes – with workstations like this. With Sentra Brain, my team and I help clients design, source, and deploy cost-efficient AI hardware (or hybrid infrastructure), and then customize the software stack so it’s ready for private LLMs, agents, and secure integrations.

Instead of renting A100s or H100s in the cloud for 10–20 €/hour, you own your machine. Within months, the workstation pays for itself. And, critically, your data never leaves your office.

Conclusion

The a16z “personal AI workstation” showed what’s possible at the bleeding edge – but also how expensive it can be. With careful part selection, a bit of the used market, and a focus on value, you can build a 15 000 € AI supercomputer under your desk in Spain or Europe.

That’s exactly the kind of solution I bring to clients with Sentra Brain: powerful, affordable, and entirely under your control.