Arsen's AI Hardware Store
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.
$2,400 to $6.5 million. Real NVIDIA products, real prices, real power draw. Every number comes with an explanation.
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.
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.
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.
Desktop cards cannot share VRAM. Each card is isolated — you can only run models that fit in 24 GB.
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.
Professional certification and ECC memory justify the jump from a consumer card. For a business running AI workloads, the reliability is worth it.
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.
Workstation GPUs use NVLink on select boards but still can't pool memory like server GPUs can.
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.
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.
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.
Designed for rack deployment. Multiple L40S cards in a server work in parallel on different inference requests, not the same model.
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.
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.
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.
A100s connect via NVLink within a server, sharing high-speed memory access. Multiple DGX A100 servers connect via InfiniBand for true distributed training.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Multiple DGX H100s connect via InfiniBand networking for true distributed training. This is real datacenter clustering — not just computers in proximity.
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.
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.
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.
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.
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.
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.
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.
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.
Three frontier open-source models — MIT and Apache 2.0 licensed. You can run them commercially, free, forever.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Two honest paths. Pick the one that fits your situation.
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.
When one machine isn't enough — and when more machines don't help.
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.
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.
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.
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.
At this scale: A/B test two models in production, keep one system reserved for batch training, serve live traffic on the others.
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.
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.
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.
Tell us what you're interested in. No payment, no account. We'll follow up with pricing and availability.