Layer Zero sells the four rungs of the hardware ladder that enterprise AI models actually run on — from the desktop card that shows you the honest floor, to the full rack that powers a neighborhood.

Rung 1Desktop Card

NVIDIA GeForce RTX 5090

The entry point, and mostly a reality check. At 32 GB it's the cheapest way to get GPU memory, but no single card holds a big model, and desktop cards can't pool. It's here so you can see the honest floor: a pile of these looks cheap but can't cluster to run the models in this store. Useful for understanding why you need datacenter hardware — rarely the thing you actually buy.

32GB
$4,099/card
VERIFIED
575 W0.5homes
✗ No — no NVLink

No recommended build uses these cards. 11 × 32 GB = 352 GB on paper — but eleven separate islands that can't pool. See the full math on the Flash model page.

Get a quote →
Rung 2Datacenter Card

NVIDIA H200 NVL

The real building block. 141 GB and, crucially, NVLink — so several pool into one larger machine. The à-la-carte card: you buy the number you need and cluster them yourself (3 for Flash, 7 for GLM). The honest cheapest way to run a big model, if you have the expertise to build and run the cluster. The startup's build.

141GB
$35,998/card
VERIFIED
600 W0.5homes
✓ Yes — NVLink, 900 GB/s
ModelCardsTotal memoryTotal pricePower
DeepSeek V4 Flash3423 GB$107,994
VERIFIED
1,800 W → 1.5 homes
GLM 5.27987 GB$251,986
VERIFIED
4,200 W → 3.5 homes
DeepSeek V4 Pro141,974 GB$503,972
VERIFIED
8,400 W → 7 homes
Get a quote →
Rung 3Integrated Server

NVIDIA DGX H200

A pre-built machine using eight H200 SXM modules — a server-only version of the chip you can't buy as loose cards — wired and clustered at the factory into one 1,128 GB unit. You reach for this when you'd rather buy one integrated, supported machine than assemble a stack of cards. More than the equivalent loose cards, but it arrives working. The mid-size company's build.

1,128GB
~$400–500K
VERIFIED EXAMPLE
$467,311Source: Broadberry
10.2 kW8.5homes·2.7EV batteries/day
✓ Yes — NVLink
CPU and RAM come pre-matched to the GPUs — not numbers you choose.
ModelUnitsTotal memoryPricePower
DeepSeek V4 Flash1× DGX H2001,128 GB~$400–500K
VERIFIED EXAMPLE
10.2 kW → 8.5 homes · 2.7 EV batteries/day
GLM 5.21× DGX H2001,128 GB~$400–500K
VERIFIED EXAMPLE
10.2 kW → 8.5 homes · 2.7 EV batteries/day
DeepSeek V4 Pro2× DGX H2002,256 GB~$800K–$1M
VERIFIED EXAMPLE
20.4 kW → 17 homes · 5.4 EV batteries/day
Get a quote →
Rung 4Full Rack

NVIDIA GB300 NVL72

The top — 72 GPUs acting as a single ~20 TB machine, drawing as much power as ~100 homes. For running the largest models at scale, or many models at once. For a single model it's overkill by design. The "when you outgrow servers entirely" ceiling — if you have the facility and budget to match.

20,100GB
~$3.7–6.5M
ESTIMATE
No official NVIDIA list price.
120.0 kW100homes+90 more·32EV batteries/day
✓ Yes — NVLink
CPU and RAM come pre-matched to the GPUs — not numbers you choose.
ModelUnits% of rack usedPricePower
Any model1× GB300 NVL72Flash ~2% · GLM ~4% · Pro ~10%~$3.7–6.5M
ESTIMATE
No official NVIDIA list price.
120 kW → 100 homes · 32 EV batteries/day

For a single model, this rack is overkill. It makes sense when running multiple large models simultaneously or at the highest inference volumes.

Get a quote →