AI chip · NVIDIA
NVIDIA GB200 Grace Blackwell Superchip
Two B200 GPUs + one Grace CPU on a single board via NVLink-C2C. Base unit of the GB200 NVL72 rack. The flagship 2024-2025 AI-training package.
Deployed in 11 named data centers · Part of 1 rack design.
Market position
GB200 is the Blackwell Superchip: 2 B200 GPUs coherently packaged with 1 Grace CPU over 900 GB/s NVLink-C2C. Sold as the NVL72 rack (72 GPUs + 36 Graces). This is the current top-of-stack training chip: Stargate, Colossus 2, Meta Hyperion, HUMAIN all committed to it.
Efficiency and power
How much work you get per watt and per dollar, and how much power a full rack draws.
- Perf per watt
- 3.33 FP8 TFLOPS/W9,000 TFLOPS ÷ 2700 W = 3.33
- Rack power
- 132 kW per NVIDIA GB200 NVL72 (~3667 W per chip)132 kW ÷ 36 chips (whole-system, includes CPU, memory, NICs, cooling)
- Cooling
- waterDepends on the rack design
What fits in 384 GB
Which open-source LLMs run on one NVIDIA GB200 Grace Blackwell Superchip, by precision. Weights only: add roughly 15% headroom for real serving. If a model does not fit at FP16, try INT8 or INT4 (smaller quality trade-off than most people expect).
| Model | Params | FP16 | INT8 | INT4 |
|---|---|---|---|---|
| Llama 3.1 8B dense | 8 B | 16 GB ✓ | 8 GB ✓ | 4 GB ✓ |
| Llama 3.1 70B dense | 70 B | 140 GB ✓ | 70 GB ✓ | 35 GB ✓ |
| Llama 3.1 405B dense | 405 B | 810 GB ✗ | 405 GB ✗ | 203 GB ✓ |
| Llama 3.3 70B dense | 70 B | 140 GB ✓ | 70 GB ✓ | 35 GB ✓ |
| DeepSeek V3 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✓ |
| DeepSeek R1 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✓ |
| Qwen 2.5 7B dense | 7 B | 14 GB ✓ | 7 GB ✓ | 4 GB ✓ |
| Qwen 2.5 72B dense | 72 B | 144 GB ✓ | 72 GB ✓ | 36 GB ✓ |
| Mixtral 8x7B MoE (46.7B total, 12.9B active per token) | 46.7 B | 93 GB ✓ | 47 GB ✓ | 23 GB ✓ |
| Mixtral 8x22B MoE (141B total, 39B active per token) | 141 B | 282 GB ✓ | 141 GB ✓ | 71 GB ✓ |
| Gemma 2 27B dense | 27 B | 54 GB ✓ | 27 GB ✓ | 14 GB ✓ |
| Command R+ dense | 104 B | 208 GB ✓ | 104 GB ✓ | 52 GB ✓ |
| Kimi K2 MoE (1T total, 32B active per token) | 1000 B | 2000 GB ✗ | 1000 GB ✗ | 500 GB ✗ |
Math: FP16 = params × 2 bytes; INT8 = params × 1 byte; INT4 = params × 0.5 bytes. MoE models sum every expert (full weights on disk), not the per-token active subset.
Common questions
How much does NVIDIA GB200 Grace Blackwell Superchip cost?
NVIDIA GB200 Grace Blackwell Superchip doesn't have a public launch list price. Vendors like this one usually price through direct sales rather than a public sheet. See list price per TFLOP for chips that do publish.
How much memory does NVIDIA GB200 Grace Blackwell Superchip have?
384 GB of HBM3e, running at 16,000 GB/s. Straight from the vendor datasheet. See chips with the most memory for context.
How much power does one NVIDIA GB200 Grace Blackwell Superchip draw?
2700 W at the chip. A full server draws more once you add CPU, memory, networking and cooling: see the rack power number above. Compare to other chips on perf-per-watt.
Which data centers use NVIDIA GB200 Grace Blackwell Superchip?
11 named data centers run them, including Meta Prometheus, Cologix Johnstown Campus and Fermi America HyperGrid (Project Matador), and 8 more. See who has the most NVIDIA GB200 Grace Blackwell Superchip for the ranked list.
Which open-source LLMs fit on one NVIDIA GB200 Grace Blackwell Superchip?
Of 13 open-source LLMs we track, 9 fit at FP16, 9 at INT8, and 12 at INT4 (for example Llama 3.1 8B, Llama 3.1 70B and Llama 3.1 405B). Full table above with each model's memory need.
Key facts
- Class
- Compute SoC (AI accelerator)
- Designer
- NVIDIA
- Safety-critical?
- No (data-center inference / training)
- Record as of
- 2026-09-27
- Most recent source
- 2025-07-14 (across 12 sources on this page)
Specifications (8 fields, click to expand)
Straight from the vendor datasheet. Dense throughput shown first; sparse (2:4) numbers in parentheses where the vendor publishes them. Full datasheet linked below.
- Process node
- TSMC 4NP
- TDP
- 2700 W
- Memory
- 384 GB HBM3e
- Memory bandwidth
- 16,000 GB/s
- FP8 (dense)
- 9,000 TFLOPS (18,000 sparse)
- Form factor
- Superchip (2xB200 + 1xGrace)
- Announced
- 2024-03-18
- Released
- 2024-12-01
Source: vendor datasheet
Generation
Foundry & process
- Fabbed at
- TSMC (all chips TSMC makes →)
- Process node
- TSMC 4NP
Rack designs that use it
Whole-rack systems buyers order in bulk (NVL72, HGX, TPU pods) that ship with NVIDIA GB200 Grace Blackwell Superchip inside.
- ×36NVIDIA GB200 NVL72rack
Compare with
Benchmarks
Published performance numbers by workload. Vendor datasheet figures where noted; independent measurements otherwise.
| Workload | Value | Unit | Source | As of |
|---|---|---|---|---|
| peak tflops fp4 dense(Two B200 GPUs per superchip) | 10000 | TFLOPS | vendor | 2024-03-01 |
Data centers running NVIDIA GB200 Grace Blackwell Superchip
Who has the most? →Named data-center campuses with NVIDIA GB200 Grace Blackwell Superchip on site. Counts shown where the operator has published them; other rows are described in general terms.
- Meta Prometheus Data Center (New Albany)as of 2025-07-14partially energizedreported
NVIDIA Blackwell training capacity at the 1 GW Prometheus site
- Cologix Johnstown Campusas of 2025-06-01announcedinferred
NVIDIA Blackwell capacity for AI-training tenants at the Johnstown campus
Multi-tenant hyperscale campus; NVIDIA Blackwell tenants inferred from CoreWeave-generation posture.
- Fermi America HyperGrid (Project Matador)as of 2025-06-01under constructionreported
Planned Blackwell capacity for the Fermi HyperGrid campus (2 GW target, four planned SMRs)
- HUMAIN (Riyadh & Dammam flagship data centers)as of 2025-05-13under constructionreported
Up to 500,000 Blackwell GPUs committed across the HUMAIN Riyadh/Dammam campuses (multi-year)
- Homer City Energy Campusas of 2025-04-01under constructioninferred
Planned 4.5 GW gas-fed AI campus targeting NVIDIA Blackwell tenants
Announced as an AI-first campus. Chip mix not yet published; Blackwell inferred from AI-training positioning.
- Stargate Abileneas of 2025-01-21partially energizedreported
GB200 NVL72 racks part of the Blackwell buildout
- Microsoft Cheyenne Data Center Campusas of 2025-01-01energizedinferred
Blackwell-generation training capacity at Cheyenne
Site published as Fairwater-class; specific chip inferred from Microsoft's Blackwell-first stance for new build
- Microsoft Fairwater Atlanta / QTS Fayetteville Campusas of 2025-01-01partially energizedreported
Second Fairwater campus targeting Blackwell-generation compute
- Microsoft Fairwater Data Center (Mount Pleasant)as of 2025-01-01partially energizedreported
Blackwell-generation AI datacenter (Fairwater flagship); Microsoft's next-generation closed-loop cooled AI DC design
- Project Kilby (Chevron / Microsoft)as of 2025-01-01announcedinferred
Microsoft-anchored AI capacity at the Chevron gas-turbine-fed campus
Inferred from Microsoft's Fairwater-generation Blackwell posture; site is a Chevron-Microsoft build for AI training
- DataVolt NEOM Oxagon Net-Zero AI Factoryas of 2024-09-01announcedreported
1.5 GW zero-carbon AI factory targeting NVIDIA Blackwell tenants
Sources
- GB200 NVL72Nvidia
Adoption rows appear as we confirm each chip-in-robot pairing from a public source. See every chip we track for the full catalog or the NVIDIA page.