NEOXIS AD - GPU Compute Project

About the Project

NEOXIS AD is a GPU compute venture for AI and machine-learning workloads, in which Aura Digital participates as a 5% co-founder. The joint-stock company is established and the public website neoxis.ai is live (React + Tailwind, built by Aura Digital). The first GPU server was delivered in July 2026 and is installed on colocation in a third-party data center - NEOXIS does not operate its own data center yet.

Current status

• NEOXIS AD established; Aura Digital is a 5% co-founder
• First server delivered and racked on colocation, July 2026; Vast.ai-verified at 98.79% reliability, listed at $1.05/GPU/hour
• Delivery, unboxing and rack installation handled by Aura Digital
• neoxis.ai website live (React + Tailwind)

The first server - delivered and installed (July 2026)

• 8× NVIDIA RTX PRO 6000 Blackwell Server Edition (Passive) - 96GB GDDR7 each, 768GB VRAM total
• 2× AMD EPYC 9754 - 128 cores, 2.25 GHz, 256MB L3, Socket SP5
• 24× Micron 64GB DDR5-4800 RDIMM ECC - 1.5TB RAM
• 2× Samsung 990 PRO 4TB M.2 NVMe (PCIe 4.0)
• 10GbE network card - 2× 10GBase-T ports (PCIe 3.0 x16)

The machine is a 4U dual-socket platform (the same one behind our AURA 4U-G8E2 configurator), assembled to order, delivered on a pallet and racked by Aura Digital in the colocation data center.

The vision

• GPU compute for AI training and inference for clients in Bulgaria and the region
• Capacity offered through GPU marketplaces (Vast.ai / RunPod) and directly
• Grow from colocated servers toward dedicated capacity over time

Passing Vast.ai verification

A GPU machine only earns on a marketplace once it is verified. Vast.ai does not take the spec sheet at face value - it measures the host under load and publishes the result to every prospective renter. The NEOXIS node cleared all four gates:

• Reliability 98.79% - the platform's rolling uptime score. Below roughly 95% a machine stops appearing near the top of search results, whatever its hardware.
• 935.61 TFLOPS aggregate compute across the eight RTX PRO 6000 Blackwell cards.
• 1,396.4 GB/s measured GPU memory bandwidth - the number that decides whether large-context inference is fast or merely possible.
• 6,687 MB/s NVMe throughput and CUDA 13, so dataset staging is not the bottleneck.

Listed at $1.05 per GPU-hour. What a machine actually earns depends on occupancy, which nobody can guarantee in advance - the arithmetic is set out honestly in our Vast.ai verification case study.

What the build taught us

Three things carried over into how we now spec every rental-oriented server:

1. Host memory is not a spec-sheet luxury. 1.5 TB of DDR5 across 24 RDIMMs is what lets tenants cache large datasets in RAM instead of hammering storage - and idle-time storage thrash is a common cause of poor reliability scores.
2. Passive server-edition cards need the chassis designed around them. RTX PRO 6000 Blackwell Server Edition has no fans of its own; it lives or dies on the chassis pressure path. This is exactly why we validate inlet temperature at the middle slots.
3. Colocation beats a hurry to build. Renting rack space in a proper facility got the machine earning months earlier than any own-facility timeline would have, with someone else carrying the power and cooling risk.

What Makes It Special

The venture targets a real market gap - Bulgaria has little specialized GPU capacity for AI. Aura Digital's role as co-founder is a natural extension of our hands-on experience building GPU servers (see the 8x Vast.ai servers project) and positions the company in the growing AI-infrastructure market.

Status: First server Vast.ai-verified (98.79% reliability) and live on colocation since July 2026; capacity ramp-up ahead.