The hardware
Real NVIDIA products, desktop card to datacenter rack. Read every number next to its plain-words explanation — open "What does this mean?" under any figure.
Desktop card
NVIDIA GeForce RTX 5090
The entry point. A single desktop card for experimenting and small models.
Price
$1,999
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.
GPU memory
32 GB
What does this mean?
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.
Power draw
575 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
32 GB GDDR7
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
Slots into a normal desktop PC — your own CPU/RAM.
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
1
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.
Reality check: MSRP $1,999 at launch (Jan 2025), 32 GB, 575 W TDP. Great for one person; far too small for a 100B+ model on its own.
Workstation card
NVIDIA RTX PRO 6000 Blackwell
A professional card with lots of memory — the honest cheapest way to reach big-model memory by ganging a few together.
Price
$8,500
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.
GPU memory
96 GB
What does this mean?
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.
Power draw
600 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
96 GB GDDR7 ECC
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
Goes in a workstation or server chassis.
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
1
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.
Reality check: ~$8,500 street price, 96 GB, 600 W. Three of these give a 288 GB pool — enough to honestly run a 235B model (which needs 282 GB).
Datacenter GPU
NVIDIA H100 (80 GB SXM)
The workhorse datacenter GPU of the last generation. Sold as part of a server.
Price
$28,000
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.
GPU memory
80 GB
What does this mean?
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.
Power draw
700 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
80 GB HBM3
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
Installed in a datacenter server (e.g., an 8-GPU HGX node).
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
1
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.
Reality check: ~$25,000–$30,000 per GPU, 80 GB, 700 W. Eight of them = 640 GB in one node.
Datacenter GPU
NVIDIA H200 (141 GB)
More memory than the H100 at the same power — the sweet spot for very large models.
Price
$35,000
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.
GPU memory
141 GB
What does this mean?
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.
Power draw
700 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
141 GB HBM3e
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
Installed in a datacenter server (e.g., an 8-GPU HGX/DGX node).
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
1
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.
Reality check: ~$30,000–$40,000 per GPU, 141 GB, 700 W — same power as the H100 but 76% more memory.
Server node
NVIDIA DGX H200 (8× H200 node)
A complete, ready-to-run server: eight H200s wired together with NVLink. One box, 1,128 GB pooled.
Price
$400,000
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.
GPU memory
1,128 GB
What does this mean?
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.
Power draw
10,200 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
8 × 141 GB HBM3e (1,128 GB pooled via NVLink)
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
Includes dual server CPUs and 2 TB of system RAM — a turnkey machine.
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
8
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.
Reality check: ~$400,000+ for a complete DGX/HGX H200 node. ~10.2 kW at the wall (GPUs + CPUs + cooling). The eight GPUs share memory over NVLink, so they truly act as one.
Datacenter rack
NVIDIA GB300 NVL72 (Blackwell Ultra rack)
The top of the shop. A full rack: 72 Blackwell Ultra GPUs acting as one enormous machine.
Price
$3,500,000
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.
GPU memory
20,700 GB
What does this mean?
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.
Power draw
135,000 W
What does this mean?
How much electricity the machine burns while it's on. It's a running cost (your power bill) and a design problem (all that power turns into heat you have to cool). A watt is a watt — it shows up on your bill every second the machine runs.
When it matters: Every day you own it. Two builds can cost the same to buy but very different amounts to run.
Memory type
72 × 288 GB HBM3e (~20.7 TB pooled)
What does this mean?
This is the kind and amount of the GPU's own memory — the pool your model's weights must fit inside.
When it matters: Always. It's the first thing to check for any model. Bigger model = more GPU memory, no exceptions.
CPU / chassis
36 Grace CPUs, NVIDIA networking, liquid-cooled — a rack-scale system.
What does this mean?
The general-purpose brain that runs the operating system and feeds data to the GPUs. For running an AI model, the CPU mostly directs traffic — the heavy thinking happens on the GPU. A strong CPU helps, but it is not what decides whether a model runs.
When it matters: It matters for data loading and juggling many requests — but it is NOT the number that decides if a model fits. Don't overpay here.
GPUs in this unit
72
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.
Reality check: ~$3.5M per rack, 72 GPUs, ~20.7 TB memory, 135 kW. That 135 kW is about 112 homes' worth of power, running non-stop.