DeepSeek-V4-Pro hardware requirements
Needs at least 4 GPUs at INT4 / FP4 (GB300 NVL72 GPU 288GB). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00/hr.
Top options for DeepSeek-V4-Pro
The strongest part each vendor offers for DeepSeek-V4-Pro at INT4 / FP4 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.
Tokens per second as users grow
Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.
4 x GB300 NVL72 GPU 288GB: peaks at 6.9k tokens per second and still has memory at 1024 users.4 x Instinct MI355X OAM: peaks at 5.2k tokens per second and still has memory at 1024 users.8 x Ascend 950DT: peaks at 2.3k tokens per second and still has memory at 1024 users.16 x Gaudi 2 96GB: peaks at 947 tokens per second and still has memory at 1024 users.8 x TPU v7 192GB: peaks at 2.8k tokens per second and still has memory at 1024 users.
Memory fills at: 4 x GB300 NVL72 GPU 288GB beyond 1,024 users, 4 x Instinct MI355X OAM beyond 1,024 users, 8 x Ascend 950DT beyond 1,024 users, 16 x Gaudi 2 96GB beyond 1,024 users, 8 x TPU v7 192GB beyond 1,024 users. Roofline estimates at INT4 / FP4.
Bare minimum
- Tokens/s
- 764
- Per hour
- $49.68
- Per million tokens
- $18.07
- Memory used
- 81%
- Tokens/s
- 566
- Per hour
- $8.00
- Per million tokens
- $3.93
- Memory used
- 61%
- Tokens/s
- 1.0k
- Per hour
- $8.00
- Per million tokens
- $2.14
- Memory used
- 62%
Configurations
Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.
4 x Instinct MI355X OAM: 1.5k tokens/s per replica at $23.96/hr.4 x GB300 NVL72 GPU 288GB: 1.6k tokens/s per replica at $49.68/hr.8 x GB200 NVL72 GPU 186GB: 2.8k tokens/s per replica at $84.00/hr.4 x Instinct MI350X OAM: 1.4k tokens/s per replica at $24.64/hr.4 x B300 SXM 262GB: 1.6k tokens/s per replica at $19.96/hr.8 x B200 SXM 180GB: 2.6k tokens/s per replica at $32.72/hr.4 x Instinct MI325X OAM: 548 tokens/s per replica at $15.20/hr.8 x Instinct MI300X 192GB: 985 tokens/s per replica at $19.12/hr.8 x H200 SXM 141GB: 801 tokens/s per replica at $19.84/hr.16 x H100 SXM5 80GB: 1.2k tokens/s per replica at $29.28/hr.8 x H200 NVL 141GB: 718 tokens/s per replica at $28.91/hr.16 x H100 NVL 94GB: 1.2k tokens/s per replica at $51.04/hr.16 x H100 PCIe 80GB: 822 tokens/s per replica at $40.00/hr.16 x RTX PRO 6000 Blackwell Workstation Edition: 1.1k tokens/s per replica at $28.80/hr.16 x RTX PRO 6000 Blackwell Server Edition: 980 tokens/s per replica at $11.04/hr.16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition: 1.0k tokens/s per replica at $8.00/hr.16 x Gaudi 2 96GB: 653 tokens/s per replica at $17.60/hr.16 x A100 PCIe 80GB: 486 tokens/s per replica at $21.60/hr.16 x A100 SXM4 80GB: 494 tokens/s per replica at $22.40/hr.
Priced configurations only; 32 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.
| Rent | |||
|---|---|---|---|
| 4 | 4.6k tok/s | Specs → | |
| 4 | 4.2k tok/s | Specs → | |
TPU 8t announced | 8 | ~2.5k tok/s | Specs → |
TPU 8i announced | 4 | ~1.6k tok/s | Specs → |
| 4 | 1.5k tok/s | Rent → | |
| 4 | 1.6k tok/s | Rent → | |
| 8 | 2.8k tok/s | Rent → | |
| 4 | 1.4k tok/s | Rent → | |
| 8 | ~1.6k tok/s | Specs → | |
Instinct MI430X OAM announced | 4 | ~2.9k tok/s | Specs → |
| 4 | 1.6k tok/s | Rent → | |
| 8 | 2.6k tok/s | Rent → | |
| 4 | 2.1k tok/s | Specs → | |
| 8 | 2.5k tok/s | Specs → | |
| 8 | ~4.9k tok/s | Specs → | |
| 4 | ~548 tok/s | Rent → | |
| 8 | ~985 tok/s | Rent → | |
| 16 | ~815 tok/s | Specs → | |
| 4 | 1.2k tok/s | Specs → | |
| 8 | ~801 tok/s | Rent → | |
H100 SXM5 80GB 2 nodes | 16 | ~1.2k tok/s | Rent → |
| 8 | ~622 tok/s | Specs → | |
| 8 | ~718 tok/s | Rent → | |
H100 NVL 94GB 2 nodes | 16 | ~1.2k tok/s | Rent → |
| 8 | ~599 tok/s | Specs → | |
H100 PCIe 80GB 2 nodes | 16 | ~822 tok/s | Rent → |
H800 PCIe 80GB 2 nodes | 16 | ~822 tok/s | Specs → |
| 8 | 1.1k tok/s | Specs → | |
| 16 | 1.1k tok/s | Rent → | |
| 8 | 556 tok/s | Specs → | |
| 16 | ~598 tok/s | Specs → | |
| 16 | 980 tok/s | Rent → | |
| 8 | ~935 tok/s | Specs → | |
TPU v5p 95GB 2 nodes | 16 | ~708 tok/s | Specs → |
| 16 | 1.0k tok/s | Rent → | |
| 8 | ~170 tok/s | Specs → | |
Gaudi 2 96GB 2 nodes | 16 | ~653 tok/s | Rent → |
| 16 | 837 tok/s | Specs → | |
| 16 | ~529 tok/s | Specs → | |
| 8 | ~370 tok/s | Specs → | |
| 8 | ~355 tok/s | Specs → | |
A800 PCIe 80GB 2 nodes | 16 | ~486 tok/s | Specs → |
A800 SXM4 2 nodes | 16 | ~494 tok/s | Specs → |
A100 PCIe 80GB 2 nodes | 16 | ~486 tok/s | Rent → |
A100 SXM4 80GB 2 nodes | 16 | ~494 tok/s | Rent → |
RTX PRO 5000 Blackwell 72GB 2 nodes | 16 | ~727 tok/s | Specs → |
Instinct MI210 PCIe 2 nodes | 16 | ~313 tok/s | Specs → |
| 8 | ~173 tok/s | Specs → | |
H20 96GB 2 nodes | 16 | ~306 tok/s | Specs → |
| 16 | ~160 tok/s | Specs → | |
| 16 | ~215 tok/s | Specs → |
By vendor
At INT4 / FP4 the NVIDIA pick is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 1.0k tokens/s for $8.00/hr; the AMD pick is 4 x Instinct MI355X OAM at about 1.5k tokens/s for $23.96/hr. Per rental dollar NVIDIA delivers 2.11x the tokens of AMD here.
NVIDIA: 130 tokens/s per dollar, 216 tokens/s per kW (16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition).AMD: 61.5 tokens/s per dollar, 263 tokens/s per kW (4 x Instinct MI355X OAM).Huawei: no live price, no power figure (8 x Ascend 950DT).Intel: 37.1 tokens/s per dollar, 68.0 tokens/s per kW (16 x Gaudi 2 96GB).Google: no live price, no power figure (8 x TPU v7 192GB).Other: no live price, 141 tokens/s per kW (8 x Cloud AI 100 Ultra).
Fleet what-if
Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.
| Fleet | Tokens/s |
|---|---|
1,000 x H200 SXM 141GB | 100k tok/s |
104 x GB200 NVL72 GPU 186GB | 36k tok/s |
1,000 x H200 SXM 141GB: 100k tokens/s, $2,480/hr, 700 kW.104 x GB200 NVL72 GPU 186GB: 36k tokens/s, $1,092/hr, no power figure.
H200 SXM 141GB: 100k tokens/s at 1000 GPUs.GB200 NVL72 GPU 186GB: 348k tokens/s at 1000 GPUs.
Size DeepSeek-V4-Pro for a tokens per second target
The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.
Frequently asked questions
How much GPU memory does DeepSeek-V4-Pro need?
The weights take 899 GB at the native INT4 / FP4 precision (3,198 GB at BF16, 1,599 GB at FP8, 899 GB at INT4). Each concurrent stream adds 4.1 KB of KV cache per token: 0.0 GB at 4,096 tokens and 0.1 GB at 32,768 tokens.
What is the bare minimum to run DeepSeek-V4-Pro?
4 x GB300 NVL72 GPU 288GB at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00 per hour.
How many H100s do you need to run DeepSeek-V4-Pro?
16 H100s at INT4 / FP4 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 16 H100s, and 1,000 tokens per second takes 16 H100s across 1 replicas.
How much does DeepSeek-V4-Pro cost per million tokens on an H100?
About $6.69 per million output tokens at INT4 / FP4, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.14.
What precision does DeepSeek-V4-Pro ship in?
The published checkpoint is FP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.
All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. Experts are stored in FP4 and attention in FP8. The KV cache is a single 512-wide latent per layer compressed 4x to 128x on most layers; 4.2 KB per token is an estimate from compress_ratios and the true figure depends on the serving stack. Active parameters computed from the expert geometry.