Maria's AI Hardware Company
Real NVIDIA hardware for running real open-source AI models — from a single workstation card to a full datacenter rack. Every number explained in plain words.
Help me choose the right build →Hardware We Sell
Six products, from a $2,499 desktop card to a $3,000,000 datacenter rack. Every spec explained so you know what it means for your workload.
NVIDIA GeForce RTX 4090
This is the entry point to serious AI hardware — roughly the price of a used car. It is the highest-end card you can buy off the shelf for a gaming or workstation PC.
24 GB of video memory means the card can hold AI model weights for models up to about 24 billion parameters. Useful for smaller open-source models, but far too small for the three models listed on this page.
450 W is about as much as four bright incandescent light bulbs running at once, or roughly 38% of what an average home uses continuously.
Best for: Small language models (up to ~24B parameters), image generation, video editing, game development.
Request a QuoteNVIDIA RTX 6000 Ada Generation
Professional tier — used by scientists, architects, and AI researchers who need reliability at a workstation desk. ECC memory automatically catches and corrects memory errors, which matters for long-running AI jobs where a single bad bit could silently corrupt results.
48 GB is double the RTX 4090's capacity. You can fit models with up to about 48 billion parameters in memory at once, making it suitable for medium-sized open-source models.
At 300 W, this professional card actually uses less power than the consumer RTX 4090 — NVIDIA optimized it for sustained all-day workstation use rather than short gaming bursts.
Best for: Medium language models (up to ~48B parameters), scientific computing, professional visualization.
Request a QuoteNVIDIA H100 PCIe 80GB
Entering datacenter territory: one card costs more than most cars. The price reflects purpose-built 24/7 reliability, error-correcting memory, and roughly 3× the AI throughput of consumer cards per watt.
80 GB of HBM2e (High-Bandwidth Memory) — more than triple the RTX 4090. When you connect several H100s together in a cluster, their memories pool together so a model can span multiple cards.
350 W is similar to the professional RTX 6000 Ada, but this card delivers far more AI performance per watt for the matrix calculations transformer models use.
Best for: Large language model inference, fine-tuning, building multi-GPU clusters for very large models.
Request a QuoteNVIDIA H100 SXM5 80GB
The SXM version costs more because the SXM socket lets multiple GPUs share data via NVLink — a special high-speed highway between cards — at 900 GB/s per GPU. When a model is spread across several cards, that speed cuts the time each card waits for data from the others.
Same 80 GB as the PCIe version, but with HBM3 memory that moves data 67% faster. When a model is split across many cards, this extra speed translates directly into faster responses.
700 W — double the PCIe version. The SXM form factor requires a special server chassis with high-power delivery and liquid-assisted cooling. It is not for a standard server rack; it lives inside the DGX H100 system below.
Best for: Large AI model training and inference where maximum inter-GPU bandwidth is critical — housed inside DGX H100 systems.
Request a QuoteNVIDIA DGX H100
A complete, ready-to-run AI server — eight H100 SXM GPUs, the chassis, power supplies, cooling, and networking all in one tested box. Roughly what a small house costs in some U.S. cities.
8 × 80 GB = 640 GB total. The eight GPUs are connected via NVLink at 900 GB/s each, so from the software's point of view they act as one large 640 GB memory pool — not eight separate cards.
10,200 W — about 8.5× what an entire average American home draws continuously. It needs a dedicated power circuit and datacenter-grade cooling; it cannot run in a normal office.
Best for: Running or fine-tuning large AI models up to ~640B parameters; on-premise AI infrastructure for serious organizations.
Request a QuoteNVIDIA GB200 NVL72 Rack
Three million dollars for a single rack. This is what the world's leading AI labs and cloud providers buy — the current ceiling of publicly available AI hardware. Companies typically sign orders for dozens of racks at a time.
72 Blackwell B200 GPUs × 192 GB each = 13,824 GB — nearly 14 terabytes of GPU memory. The largest AI models in the world fit here with room to spare; all three models on this page could run simultaneously.
120,000 W — 120 kilowatts, or about as much as 100 average homes running at once. A single rack requires its own dedicated power feed and liquid cooling infrastructure; it cannot share a building's normal electrical system.
Best for: Frontier AI model training, massive-scale inference, datacenter AI infrastructure for cloud providers and top AI labs.
Request a QuoteOpen-Source Models We Support
Three frontier open-source models, each with MIT license — free for commercial use. Below you'll see exactly how much hardware each one needs to run, with the memory math shown step by step.
What "active parameters" means: All three models below use a Mixture of Experts (MoE) architecture — only a fraction of parameters activate for each question. This makes inference more efficient, but all parameters still need to be loaded into GPU memory so the right experts are available on demand.
Memory formula used on this page: Total parameters (billions) × 1 GB + 20% safety buffer. This assumes INT8 quantization (1 byte per parameter), which is the standard way to run large models efficiently without too much quality loss.
MiMo V2.5 Pro
Xiaomi's flagship open model. Despite having 1.02 trillion total parameters, only 42 billion activate per inference pass — the MoE architecture routes each question to a small subset of 'expert' sub-networks, making it efficient to run while retaining the knowledge of a much larger model.
Parameter Counts
| Total parameters | 1,020 billion |
| Active per inference pass | 42 billion (only these fire per question, but all must be loaded) |
| Context window | 1 million tokens |
Memory Calculation
Minimum Hardware to Run This Model
Request a QuoteGLM 5.1
Zhipu AI's reasoning-focused open model, built on a 754B MoE architecture. Designed for complex reasoning and software engineering, it achieves 94.6% of Claude Opus 4.6's coding benchmark — a remarkable result for an open-source model trained without NVIDIA hardware.
Parameter Counts
| Total parameters | 754 billion |
| Active per inference pass | 40 billion (only these fire per question, but all must be loaded) |
| Context window | 200,000 tokens |
Memory Calculation
Minimum Hardware to Run This Model
Request a QuoteDeepSeek V4 Pro
DeepSeek's most powerful open model. With 1.6 trillion total parameters and a 1-million-token context window, it rivals the best closed-source frontier models on most benchmarks — and is free to use commercially under the MIT license.
Parameter Counts
| Total parameters | 1,600 billion |
| Active per inference pass | 49 billion (only these fire per question, but all must be loaded) |
| Context window | 1 million tokens |