Arsen's AI Hardware Store

Every GPU,
explained.

You shouldn't need a PhD to buy a GPU. This store teaches you exactly what you're buying — specs, prices, power draw, and the honest math behind running real frontier AI models.

9
Products
13,824 GB
Max GPU Memory
$2,400
Starting Price
Scroll

The Hardware Line

$2,400 to $6.5 million. Real NVIDIA products, real prices, real power draw. Every number comes with an explanation.

Desktop

GeForce RTX 4090

$2,400
GPU Memory 24 GB GDDR6X
Power Draw 450 W TDP
What does 24 GB mean?

24 GB is enough to run models up to about 20 billion parameters. Think of it as the workspace your GPU has — bigger models need more space. Perfect for developers and researchers testing ideas before scaling up.

Is $2,400 a fair price?

At $2,400, this is the starting gate for serious AI work on your own hardware. You get professional-grade inference at a consumer price.

450W — what does that cost?

450 watts running full time costs roughly $400–$500 per year in electricity. That's about one-third of what an average home draws around the clock.

≈ 0.38 homes of continuous household electricity.
200 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

Desktop cards cannot share VRAM. Each card is isolated — you can only run models that fit in 24 GB.

Request Quote →
Professional Workstation

RTX 6000 Ada

$6,800
GPU Memory 48 GB GDDR6 ECC
Power Draw 300 W TDP
What does 48 GB mean?

48 GB with ECC (error-correcting memory) means fewer silent data errors — critical when you're running inference on customer-facing products. Fits models up to about 40B parameters.

Is $6,800 a fair price?

Professional certification and ECC memory justify the jump from a consumer card. For a business running AI workloads, the reliability is worth it.

300W — what does that cost?

300 watts — more memory than the RTX 4090 but actually draws less power. Efficient design means lower operating costs over time.

≈ 0.25 homes of continuous household electricity.
300 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

Workstation GPUs use NVLink on select boards but still can't pool memory like server GPUs can.

Request Quote →
Data Center Inference

NVIDIA L40S

$14,000
GPU Memory 48 GB GDDR6 ECC
Power Draw 350 W TDP
What does 48 GB mean?

48 GB in a server form factor. The L40S is built specifically for inference — serving model responses to many users at once, not just one researcher.

Is $14,000 a fair price?

At $14,000 you get a card that can handle sustained production traffic. Think of this as the difference between a test kitchen and a restaurant.

350W — what does that cost?

350 watts in a rack means you can stack 4–8 of these in a server without overwhelming your power supply.

≈ 0.29 homes of continuous household electricity.
257 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

Designed for rack deployment. Multiple L40S cards in a server work in parallel on different inference requests, not the same model.

Request Quote →
Data Center Training

NVIDIA A100 80GB

$25,000
GPU Memory 80 GB HBM2e
Power Draw 400 W TDP
What does 80 GB mean?

80 GB of HBM2e (High Bandwidth Memory) — this memory sits right next to the GPU chip, connected by thousands of tiny wires instead of one bus. That means much faster data transfer than GDDR6.

Is $25,000 a fair price?

The A100 was the GPU that trained GPT-3 and early GPT-4 class models. It remains a reliable, proven workhorse for serious training runs.

400W — what does that cost?

400 watts per card — in a full 8-GPU DGX system you're looking at 3.2 kW just for GPUs, plus CPUs and cooling.

≈ 0.33 homes of continuous household electricity.
225 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

A100s connect via NVLink within a server, sharing high-speed memory access. Multiple DGX A100 servers connect via InfiniBand for true distributed training.

Request Quote →
Data Center

NVIDIA H100 PCIe 80GB

$32,000
GPU Memory 80 GB HBM3
Power Draw 350 W TDP
What does 80 GB mean?

80 GB of HBM3 — the next generation of high-bandwidth memory after HBM2e. Roughly 2× the memory bandwidth of the A100, meaning the GPU can access training data much faster.

Is $32,000 a fair price?

The H100 PCIe fits into standard servers without special infrastructure. This is the most common single-card purchase for serious AI teams building their first cluster.

350W — what does that cost?

350 watts is notably efficient for this performance tier. 4 of these in a server draws 1.4 kW from the GPUs alone — manageable in a standard rack.

≈ 0.29 homes of continuous household electricity.
257 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

PCIe means it connects through a standard CPU server slot. You can put 4–8 in one machine. Unlike the SXM version, these don't use NVLink for GPU-to-GPU communication.

Request Quote →
Data Center

NVIDIA H100 SXM 80GB

$40,000
GPU Memory 80 GB HBM3
Power Draw 700 W TDP
What does 80 GB mean?

Same 80 GB HBM3 as the PCIe version, but the SXM connector enables NVLink — a high-speed bridge that lets multiple GPUs share memory at 900 GB/s. This is why it costs more and draws twice the power.

Is $40,000 a fair price?

The SXM version is bought as part of a DGX or HGX system, not alone. The premium pays for the NVLink fabric that makes multi-GPU model loading possible.

700W — what does that cost?

700 watts — twice the PCIe version. That extra power feeds the NVLink interconnect. Eight of these in a DGX draw 5.6 kW from GPUs alone.

≈ 0.58 homes of continuous household electricity.
128 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

SXM GPUs connect via NVLink inside a DGX server, allowing a single model to span all 8 GPUs as if they were one 640 GB memory pool.

Request Quote →
Enterprise Server

NVIDIA DGX H100

$399,000
GPU Memory 640 GB 8× HBM3 (8× H100 SXM)
Power Draw 10200 W TDP
What does 640 GB mean?

640 GB total across 8 H100 SXM GPUs connected by NVLink. A model can span all 8 GPUs simultaneously — it looks like one 640 GB memory pool to the software. This is real GPU memory sharing, not a workaround.

Is $399,000 a fair price?

At $399,000 you're buying a complete, tested, rack-mounted system: 8 GPUs, dual Xeon CPUs, 2 TB RAM, 30 TB NVMe storage, and InfiniBand networking. Everything configured and certified.

10200W — what does that cost?

10,200 watts — you need a 20-amp, 208V power circuit (like a commercial HVAC unit). This is 8.5× the continuous draw of an average home.

≈ 8.5 homes of continuous household electricity.
8.8 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

Multiple DGX H100s connect via InfiniBand networking for true distributed training. This is real datacenter clustering — not just computers in proximity.

Request Quote →
Enterprise Server

NVIDIA DGX H200

$499,000
GPU Memory 1128 GB 8× HBM3e (8× H200 SXM)
Power Draw 10200 W TDP
What does 1128 GB mean?

1,128 GB across 8 H200 GPUs. The H200 upgrades HBM3 to HBM3e — 76% more memory bandwidth than HBM2e. This is the first single system that can run a 671B parameter model without clustering.

Is $499,000 a fair price?

The $100K premium over the DGX H100 buys you 488 GB more GPU memory and higher bandwidth. If your model needs more than 640 GB, this is your minimum viable single-machine solution.

10200W — what does that cost?

Same 10,200 watts as the DGX H100 — NVIDIA kept power identical while nearly doubling memory. More efficient than the previous generation at this power draw.

≈ 8.5 homes of continuous household electricity.
8.8 hours per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

Two DGX H200s in a cluster give you 2,256 GB of GPU memory — enough to run DeepSeek-V3 with room to spare for two simultaneous models.

Request Quote →
Hyperscale Rack

NVIDIA GB200 NVL72

$6,500,000
GPU Memory 13824 GB 72× HBM3e (72× B200 GPU)
Power Draw 120000 W TDP
What does 13824 GB mean?

13,824 GB — nearly 14 TB — of GPU memory across 72 B200 GPUs and 36 Grace CPUs. This is an entire rack that functions as one computer. You could run every frontier model simultaneously with room to spare.

Is $6,500,000 a fair price?

At $6.5M you're buying a piece of infrastructure, not a component. This is what hyperscalers and national labs buy to train frontier models. The per-FLOP cost is actually lower than any smaller system at this scale.

120000W — what does that cost?

120,000 watts — 120 kilowatts. That is the continuous draw of about 100 average homes. You need a dedicated power substation and purpose-built cooling infrastructure.

≈ 100 homes of continuous household electricity.
0.75 hours (45 min) per EV battery charge on a 90 kWh EV battery.

Can I combine multiple units?

All 72 GPUs are NVLink-connected within the rack. This IS the cluster — no external networking needed. Multiple racks connect via InfiniBand for national-scale training.

Request Quote →

The Models

Three frontier open-source models — MIT and Apache 2.0 licensed. You can run them commercially, free, forever.

Memory math: Each billion parameters needs roughly 1 GB of GPU memory at standard precision. Add 20% on top for working memory during inference. Formula: Parameters (B) × 1 GB + 20% = minimum GPU memory required.
MIT Top 5 open models globally

DeepSeek-V3

by DeepSeek

One of the most capable open-source models in the world. Uses a Mixture-of-Experts design: despite 671B total parameters, only 37B are active during each inference pass — making it far faster than a 671B dense model would be.

671B parameters
671B total parameters (Mixture-of-Experts architecture; 37B parameters active per inference pass)
Memory Calculation

Why total parameters — not active-per-pass? In Mixture-of-Experts models, all expert weights must stay resident in GPU memory. The hardware cannot swap inactive experts in on demand — latency would make inference unusable.

Total parameters 671 billion parameters
Base memory needed × 1 GB per billion = 671 GB
Working room buffer + 20% working room = 134 GB
Minimum required = 805 GB minimum GPU memory
Minimum build to run this
DGX H200
$499,000
1128 GB available > 805 GB needed ✓

The DGX H200 (1,128 GB) is the smallest single machine that fits DeepSeek-V3 with comfortable headroom. A 2× DGX H100 cluster (1,280 GB) is the multi-machine alternative.

Advanced reasoning Code generation Research assistant Long-context analysis
MIT Top agent leaderboard models

GLM-5.2 Max

by Z.ai

Z.ai's top agent-tuned model, built specifically for complex multi-step tasks and tool use. MIT licensed — you own your deployment completely, no usage fees, no API dependency.

130B parameters
~130B parameters — verify exact count at arena.ai/leaderboard/agent
Memory Calculation

Why total parameters — not active-per-pass? In Mixture-of-Experts models, all expert weights must stay resident in GPU memory. The hardware cannot swap inactive experts in on demand — latency would make inference unusable.

Total parameters ~130 billion parameters
Base memory needed × 1 GB per billion = 130 GB
Working room buffer + 20% working room = 26 GB
Minimum required = 156 GB minimum GPU memory
Minimum build to run this
2× H100 PCIe in server
$84,000
160 GB available > 156 GB needed ✓

Two H100 PCIe cards (160 GB total) in a purpose-built server with PCIe bifurcation support. Note: these don't share memory like NVLink — both cards must hold a copy, so you effectively need 2× the model size to run inference across both.

Agent workflows Tool use Multi-step reasoning Automated tasks
Apache 2.0 Top 10 open models globally

Qwen3-235B

by Alibaba / Qwen Team

Alibaba's largest open model. Apache 2.0 license means truly free commercial use — no restrictions, no attribution required. The MoE design activates only 22B parameters per pass, delivering frontier-quality output at practical inference speeds.

235B parameters
235B total parameters (Mixture-of-Experts; 22B active per pass)
Memory Calculation

Why total parameters — not active-per-pass? In Mixture-of-Experts models, all expert weights must stay resident in GPU memory. The hardware cannot swap inactive experts in on demand — latency would make inference unusable.

Total parameters 235 billion parameters
Base memory needed × 1 GB per billion = 235 GB
Working room buffer + 20% working room = 47 GB
Minimum required = 282 GB minimum GPU memory
Minimum build to run this
4× H100 PCIe in server
$148,000
320 GB available > 282 GB needed ✓

Four H100 PCIe cards (320 GB total) in a 4U server chassis. This is the minimum practical setup — enough headroom above the 282 GB requirement for smooth operation and future quantization experiments.

Multilingual tasks Code generation Scientific reasoning Business automation

Help Me Choose

Two honest paths. Pick the one that fits your situation.

The Startup Build

You want to ship a product powered by a real frontier model. You don't need multi-tenant scale yet — just reliable inference for your users.

Recommended model
Qwen3-235B
Apache 2.0 means zero licensing cost. The MoE design means faster inference per dollar. At 282 GB, it's the largest truly open model you can run on a single-machine solution.
Recommended hardware
4× NVIDIA H100 PCIe 80GB in a 4U server chassis
Four H100 PCIe cards gives you 320 GB — just enough headroom above the 282 GB minimum. This is the cheapest multi-GPU server config that can actually run Qwen3-235B as a single loaded model.
4× NVIDIA H100 PCIe 80GB $128,000
4U server chassis (e.g., Supermicro AS-4124GO) $12,000
2× AMD EPYC or Intel Xeon CPUs $6,000
512 GB DDR5 server RAM $3,000
4 TB NVMe storage (2× 2TB) $1,500
10GbE networking + rack installation $2,000
Total Build Cost $152,500
1400W
≈ 1.2 homes of continuous household electricity
≈ 64 hours per EV battery charge
Honest note: This setup runs Qwen3-235B but the four H100 PCIe cards don't have NVLink — they can't share memory natively. You'll run the model split across cards using model parallelism (tensor parallel inference), which works well with the right software (vLLM, TGI). For true shared-memory multi-GPU, you'd need a DGX system.
Request This Build →

GPU Clusters

When one machine isn't enough — and when more machines don't help.

What is a cluster?

A cluster is two or more machines connected by high-speed networking — typically InfiniBand at 200–400 Gbit/s — managed by software as a single system. A model too large for one machine spans all of them, using all their combined memory.

Real Datacenter Clustering

Hardware: DGX H100, DGX H200, GB200 NVL72

Connection: NVLink (GPU-to-GPU within a machine) + InfiniBand 200 Gbit/s (machine-to-machine)

What you get: GPU memory actually pools. A 671B model can span two DGX H100s as one 1,280 GB machine. The software (NCCL) handles this transparently.

Honest bottom line: This is the only way to run models larger than one machine's memory. It requires datacenter infrastructure — dedicated power feeds, cooling, InfiniBand fabric switches.

Piling Up Desktop Cards

Hardware: Multiple RTX 4090s or similar consumer cards

Connection: PCIe bus only — no NVLink, no InfiniBand

What you get: Each card has its own isolated 24 GB. You can run different models on different cards, or parallelize separate inference requests — but you cannot load a single model that needs 48 GB across two cards.

Honest bottom line: Desktop multi-GPU is great for parallel workloads. It is not a substitute for NVLink when you need to run one large model across multiple GPUs.

Interactive Cluster Builder
DGX H100 #1
H1
H2
H3
H4
H5
H6
H7
H8
640 GB · 10.2 kW
Combined Memory
640 GB
Combined Power
10,200 W
Combined Cost
$399,000

2× DGX H100 Cluster

Minimum for DeepSeek-V3
1,280 GB
Combined GPU Memory
2 × 640 GB — fits DeepSeek-V3's 805 GB requirement
20,400 W
Combined Power
≈ 17 homes — requires dedicated datacenter power
$798,000
Combined Cost
2 × $399,000 + ~$20K InfiniBand switch

Connected via InfiniBand 200 Gbit/s. NVLink handles GPU-to-GPU within each machine; InfiniBand handles machine-to-machine. The model sees all 16 GPUs as one memory pool.

4× DGX H100 Cluster

Production Scale
2,560 GB
Combined GPU Memory
4 × 640 GB — run DeepSeek-V3 + Qwen3-235B simultaneously
40,800 W
Combined Power
≈ 34 homes — dedicated 480V three-phase feed required
$1,596,000
Combined Cost
4 × $399,000 + ~$50K full InfiniBand fabric switch

At this scale: A/B test two models in production, keep one system reserved for batch training, serve live traffic on the others.

Power Calculator

Select any product and see its power draw expressed in terms you can actually feel. Reference: 1 home ≈ 1,200W continuous. 1 EV battery ≈ 90 kWh.

Select a build
Continuous Draw
450
watts
Home Equivalent
0.38 homes
running continuously at 1,200W avg
EV Battery Duration
200 hours on 90 kWh EV
one full charge would power this for
300W (min) 120,000W (max)
450 watts
Logarithmic scale — 300W to 120,000W
💡
Why power matters when buying

Hardware cost is one-time. Electricity is forever. A GB200 NVL72 rack at 120 kW costs roughly $840,000/year in electricity at $0.08/kWh — that's 13% of the rack's purchase price, annually. Always model 3-year total cost of ownership, not just sticker price.

🔌
What power infrastructure you actually need

RTX 4090: standard 120V outlet. DGX H100: dedicated 20A 208V circuit (like commercial HVAC). GB200 NVL72 rack: its own transformer, cooling towers, and fire suppression. Budget for the room, not just the machine.

Request a Quote

Tell us what you're interested in. No payment, no account. We'll follow up with pricing and availability.