◧ Govand's AI Hardware Store

The product line

Three tiers, from a card that fits under your desk to a rack that fills a room. Every number has a plain-words explanation — hover or just read underneath.

Desktop & Workstation

NVIDIA

NVIDIA GeForce RTX 5090

The top consumer card. A great, affordable start for tinkering, smaller models and fine-tuning experiments on one desk.

Price
$1,999

The maker's suggested retail price for one unit - roughly what you pay to buy one.

GPU memory
32 GB GDDR7

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
575 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
GDDR7

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA DGX Spark (GB10 Grace Blackwell)

A tiny desktop AI computer with 128 GB of unified memory - purpose-built to develop and run mid-size models locally.

Price
$3,999

The maker's suggested retail price for one unit - roughly what you pay to buy one.

GPU memory
128 GB LPDDR5X (unified)

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
240 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
LPDDR5X (unified)

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA RTX PRO 5000 Blackwell

A workstation card with 48 GB of ECC memory at a gentle power draw - roomy and reliable for serious desktop work.

Price
$4,569

The maker's suggested retail price for one unit - roughly what you pay to buy one.

GPU memory
48 GB GDDR7

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
300 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
GDDR7

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA RTX PRO 6000 Blackwell (Workstation Edition)

The biggest single desktop card you can buy - 96 GB lets one card hold surprisingly large models.

Price
$8,565

The maker's suggested retail price for one unit - roughly what you pay to buy one.

GPU memory
96 GB GDDR7

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
600 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
GDDR7

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

Server GPUs

NVIDIA

NVIDIA H100 (SXM, 80GB)

The GPU that trained much of today's AI. Still a dependable workhorse; you normally buy several together.

Price
est. $27,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
80 GB HBM3

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
700 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
HBM3

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA H200 (SXM)

An H100 with far more memory - 141 GB per card, the same power draw. The popular choice for serving big models.

Price
est. $32,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
141 GB HBM3e

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
700 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
HBM3e

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA B200 (Blackwell, SXM)

NVIDIA's current-generation datacenter GPU - much more speed than the H200, and a bigger power bill to match.

Price
est. $40,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
180 GB HBM3e

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
1000 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
HBM3e

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

NVIDIA

NVIDIA B300 (Blackwell Ultra, SXM)

The most memory in a single NVIDIA card here - 288 GB. Fits enormous models on the fewest GPUs; needs serious cooling.

Price
est. $50,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
288 GB HBM3e

The fast memory that holds the AI model while it runs. If a model needs more memory than this, it will not fit on one card.

Power draw
1400 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

Memory type
HBM3e

The kind of memory chips. HBM (High-Bandwidth Memory) is the very fast, expensive kind used in datacenter GPUs; GDDR is the cheaper kind in desktop cards; LPDDR5X 'unified' memory is shared by the chip's CPU and GPU together.

Full systems / datacenter racks

NVIDIA

NVIDIA HGX B200 (8-GPU baseboard)

Eight B200 GPUs wired together with fast NVLink - the standard building block of modern AI servers.

Price
est. $390,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
1440 GB HBM3e

The combined fast memory of every GPU inside this one machine. A model has to fit inside this total to run on a single box.

Power draw
10200 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

GPUs inside
8

How many separate GPU chips are built into this one machine.

NVIDIA

NVIDIA DGX B200 (8-GPU system)

NVIDIA's turnkey AI box: eight B200 GPUs, CPUs, networking and cooling in one complete, ready-to-run system.

Price
est. $515,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
1440 GB HBM3e

The combined fast memory of every GPU inside this one machine. A model has to fit inside this total to run on a single box.

Power draw
14300 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

GPUs inside
8

How many separate GPU chips are built into this one machine.

NVIDIA

NVIDIA HGX B300 (8-GPU baseboard)

The Blackwell Ultra building block: eight B300 GPUs and a colossal 2.3 TB of pooled memory in one baseboard.

Price
Request a quote

There is no public price for this - you send us your details and we come back with a custom quote.

GPU memory
2304 GB HBM3e

The combined fast memory of every GPU inside this one machine. A model has to fit inside this total to run on a single box.

Power draw
11200 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

GPUs inside
8

How many separate GPU chips are built into this one machine.

NVIDIA

NVIDIA GB200 NVL72 (full rack)

A whole rack acting as one giant computer: 72 GPUs and 13.4 TB of pooled memory, joined by NVLink at datacenter speed.

Price
est. $3,000,000

An ESTIMATED street price. Datacenter hardware is normally sold by custom quote, so treat this as a rough ballpark, not a fixed price.

GPU memory
13400 GB HBM3e

The combined fast memory of every GPU inside this one machine. A model has to fit inside this total to run on a single box.

Power draw
120000 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

GPUs inside
72

How many separate GPU chips are built into this one machine.

NVIDIA

NVIDIA GB300 NVL72 (Blackwell Ultra rack)

The flagship. Seventy-two Blackwell Ultra GPUs and 20 TB of pooled memory in one rack - the machine that builds frontier AI.

Price
Request a quote

There is no public price for this - you send us your details and we come back with a custom quote.

GPU memory
20480 GB HBM3e

The combined fast memory of every GPU inside this one machine. A model has to fit inside this total to run on a single box.

Power draw
135000 W

How much electricity it draws at full load. This is what drives your power bill and how much cooling the room needs.

GPUs inside
72

How many separate GPU chips are built into this one machine.