For a small startup
You want to run one capable open model for your product, as cheaply as it can be run without lying about the memory. We pick the friendliest big model and the cheapest cards that honestly hold it — three RTX PRO 6000s giving a 288 GB pool for a 235B model that needs 282 GB.
Recommended build: Qwen3-235B (235B)
What does this mean?
A model's size, counted in billions of adjustable numbers it learned during training. 'Qwen3-235B' means 235 billion parameters. More parameters usually means a smarter model — and, crucially, more GPU memory needed to run it. The 'B' is your first clue to how big a machine you'll need.
When it matters: It's the starting point for every buying decision: parameters tell you the memory, memory tells you the hardware.
What does this mean?
235B parameters × 1 GB = 235 GB, plus 20% working room (47 GB) = 282 GB of GPU memory required. The single most important number. A model's 'weights' — its entire brain — must fit inside the GPU's own memory all at once, or the model simply will not run. If a model needs 282 GB and your cards add up to 240 GB, it doesn't run slowly, it doesn't run at all. This is the number that decides what you can buy.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
What does this mean?
This build gives 288 GB — enough, with 6 GB of headroom to spare.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
What does this mean?
Big models don't fit on one card, so you link several together and their memory adds up. Four 96 GB cards give you 384 GB of usable pool. More GPUs also means more speed for more users at once.
When it matters: When one card isn't big enough for the model, or when you need to serve many users at the same time.
What does this mean?
1,800 watts — about 2 average homes' worth of power, running around the clock.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
What does this mean?
1,800 watts — about 2 average homes' worth of power, running around the clock.
When it matters: When you're deciding whether your building (and budget) can even feed the machine.
What does this mean?
The power bill, turned into money. We assume the machine runs around the clock at an average US electricity price of about $0.17 per kilowatt-hour. Formula: kilowatts × 24 hours × 30 days × $0.17. Your local rate will differ, but this makes 'watts' feel like what it really is: a monthly bill.
When it matters: When comparing builds. A cheap-to-buy machine that burns twice the power can cost more within a couple of years.
What does this mean?
Running 24/7, this build uses about 0.5 full electric-car batteries (~90 kWh each) of energy every day.
When it matters: When you want a gut sense of energy over time, not just an instant number.
What does this mean?
What you pay once, up front, to own the hardware. It does not include the electricity to run it, the cooling, or the room to put it in — those are ongoing.
When it matters: At purchase. But always read it next to the power draw: cheap-to-buy can be expensive-to-run.
This is a cluster — what that means
A cluster is several machines linked with fast networking so their GPU memory and compute pool together and act like one bigger machine.
This build links 3 GPUs so their memory (288 GB) and power pool together and act as one machine.
The honest part: Here's the honest difference. Real datacenter clustering (like an NVL72 rack or DGX nodes joined by NVLink and InfiniBand) connects GPUs with links measured in hundreds of gigabytes to terabytes per second, so 72 GPUs genuinely act like one giant GPU. A 'pile of desktop cards' — say sixteen RTX 5090s in consumer machines wired over ordinary networking — has the memory on paper, but the cards talk to each other so slowly that for one big model they spend more time waiting than working. The memory adds up; the teamwork doesn't. For serving one large model, purpose-built datacenter interconnect is why the expensive racks exist. Desktop cards are wonderful for one card's worth of work — not for pretending to be a datacenter.