Affordable GPU Supercomputers
for Offline, Private AI.
Norman Supercomputers, Inc. designs and builds on-prem, air-gapped GPU supercomputers for hobbyists, extreme gamers, tech enthusiasts, businesses, and startups who need serious AI performance without surrendering their data to the cloud.
All initial systems are in design and validation.
Models & pricing: TO BE ANNOUNCED.
Join the mailing list to be the first to hear when pre-orders open.
NS-Series GPU Supercomputer
Our reference designs are tuned around next-generation RTX 50-series GPUs, ECC memory, and high-core Threadripper & Ryzen CPUs, all configured for 24/7 offline AI workloads.
As a Florida-based family business, we build and test each system as if we were putting it in our own home lab. No mystery parts, no aggressive upselling — just honest, high-performance machines tuned for real-world AI workloads.
NS-Series Supercomputers
Three tiers of local AI power. All models and pricing are currently TO BE ANNOUNCED, but these are the design targets we’re working toward.
NS-64 Pioneer
Designed as an “AI starter” supercomputer for hobbyists, small teams, and serious gamers who want to run large models locally without breaking the bank.
- Target GPUs: 2× RTX 5090 (≈ 64 GB total VRAM)
- CPU: Ryzen 9 7950X3D (16 cores)
- Memory: 128 GB DDR5 ECC
- Storage: 4–8 TB NVMe SSD (AI-ready)
- Power: ~1,400W, air-cooled
Internal testing target: ~40–50 tokens/sec on a 70B LLM (fp16), and smooth Stable Diffusion XL image generation at 2–3 it/s.
NS-128 Vanguard
Our “best value” concept: a balanced, multi-GPU system for small businesses and AI-heavy startups that need serious throughput without going into datacenter pricing.
- Target GPUs: 4× RTX 5090 (≈ 128 GB total VRAM)
- CPU: Threadripper PRO 7955WX (16 cores)
- Memory: 256 GB DDR5 ECC
- Storage: 8–16 TB NVMe SSD (RAID options)
- Power: ~2,400W, tuned air-cooling
Internal goal: ~75–85 tokens/sec on a 70B LLM (fp16), 4.5–5.5 it/s on SDXL, and practical fine-tuning for 13B models in 4–5 hours.
NS-192 Titan
A concept “AI beast” for teams that want a serious local alternative to rented cloud clusters, with room to push bigger models and heavier fine-tunes.
- Target GPUs: 6× RTX 5090 (≈ 192 GB total VRAM)
- CPU: Threadripper PRO 7960X (24 cores)
- Memory: 512 GB DDR5 ECC
- Storage: 16–24 TB NVMe SSD (RAID / scratch)
- Power: ~3,600W, custom loop cooling
Design target: ~100–110 tokens/sec on 70B LLMs (fp16), 6.5–7.5 it/s SDXL, and 13B fine-tunes in around 3–3.5 hours.
All specifications and names are preliminary and may change. Final model names, prices, and configurations are TO BE ANNOUNCED.
Why on-prem, offline AI?
Cloud GPUs are great — until you factor in privacy, latency, and surprise bills. Norman Supercomputers exist so you can keep your data and your compute under one roof.
Data stays in your building
Fine-tune models on proprietary code, medical records, financial data, or any sensitive information without sending a single token to a third-party cloud.
Predictable cost, no “surge pricing”
Once your system is paid for, the only ongoing cost is power and maintenance. No more watching usage dashboards to avoid surprise multi-thousand-dollar bills.
Instant experimentation
Spin up local LLMs, diffusion models, and fine-tunes as quickly as you can type a command. No queueing, rate limiting, or waiting for capacity.
Family-owned accountability
You’re not a ticket number. When something matters, you can talk to the people who actually design and build these systems — not just a reseller.
Core software stack (planned)
100% Offline by defaultEvery Norman Supercomputer is designed to ship with a curated, local-first AI toolkit:
- Ubuntu 24.04 LTS + NVIDIA CUDA 12.6, tuned for multi-GPU workloads.
- Ollama for local LLMs (chat, coding, embeddings), pre-configured with popular open models.
- ComfyUI for Stable Diffusion XL and image/video workflows.
- Axolotl and supporting tools for fine-tuning 7B–70B class models.
- ECC RAM configurations for long-running training jobs.
- No cloud accounts required; everything can run completely disconnected.
Software selections are also evolving and may be tailored per customer at launch.
Design targets & performance envelope
Below is the internal reference table we’re using to shape the NS-Series. These numbers are based on our current component assumptions and are for planning only; final systems will be benchmarked and documented at launch.
| Metric | NS-64 Pioneer | NS-128 Vanguard | NS-192 Titan | PewDiePie-style Lab* |
|---|---|---|---|---|
| Total VRAM | 64 GB | 128 GB | 192 GB | ~424 GB |
| 70B LLM (fp16) inference | 40–50 tokens/sec | 75–85 tokens/sec | 100–110 tokens/sec | ~120–140 tokens/sec |
| Fine-tune 13B model | 6–8 hours | 4–5 hours | 3–3.5 hours | 2–3 hours |
| Stable Diffusion XL | 2–3 it/s | 4.5–5.5 it/s | 6.5–7.5 it/s | 8–10 it/s |
| Power draw (full load) | ~1,400W | ~2,400W | ~3,600W | ~3,500W |
| Est. monthly electric (24/7) | ≈ $100 | ≈ $200 | ≈ $300 | ≈ $500+ |
| Noise level | 40–45 dB | 45–50 dB | 55–60 dB | 65–70 dB |
*“PewDiePie-style lab” refers to a roughly comparable 10-GPU enthusiast build (8× modded 4090 + 2× 4000 Ada). It’s included here only as a reference point for extreme community builds and is not a product we sell.
- $5K → $10K class: +100% cost → +80–100% performance (sweet spot).
- $10K → $20K class: +100% cost → +25–30% performance (diminishing returns).
- Above $20K: steep cost for marginal gains, mostly edge-case laboratories.
- 128 GB VRAM covers ~95% of real-world local AI use cases.
- Extra VRAM mainly helps for very large, less quantized models and exotic experiments.
Join our mailing list
Get notified when Norman Supercomputers become available, including model reveals, pricing, and early-access opportunities. Opt-out anytime.
Privacy Policy (mailing list)
1. Purpose of collection
We collect your email address (and optional name / interest information) solely
for the purpose of sending you updates about Norman Supercomputers, including:
- Announcements about product availability and pre-orders.
- Specs and pricing updates as our models are finalized.
- Important changes to our offerings or policies.
2. No sale or third-party sharing
We do not sell, rent, or trade your contact information with any
third parties. We also do not share your email address or other details with
marketing lists, brokers, or advertisers.
3. Opt-out anytime
Every email we send will include a simple unsubscribe link. Clicking that link
will remove you from our mailing list. You may also contact us directly to request
removal.
4. Data retention
We retain your mailing list information only as long as it is needed for the
purposes described above or until you unsubscribe.
5. Security
We use reasonable technical and organizational measures to keep your contact
information secure and separate from any production systems.
As a small, family-run company, we take your trust seriously. If you ever have questions or concerns about how your information is used, please reach out and we’ll address it directly.
This policy applies specifically to our mailing list. A fuller site-wide privacy policy will be published alongside our product launch.