NVIDIA B300 vs H200 vs B200 vs H100 and GB300 NVL72 — GPU Comparison for AI Infrastructure
For new AI clusters in 2026 the choice comes down to three platforms: B300 servers, GB300 NVL72 racks, and H200 where a lower entry cost matters. B200 is being replaced by B300 in new builds; H100 is supplied on request for existing clusters.
At a glance
| Platform | Status | Memory per GPU | Best for | Cooling |
|---|---|---|---|---|
| B300 Blackwell Ultra, 8-GPU server | Current | 288 GB | New training and inference clusters | Liquid at cluster density; air-cooled versions exist |
| GB300 NVL72 72 × B300, one rack | Current | 288 GB ~20 TB per rack | Frontier training, largest models at scale | Liquid only, ~132 kW per rack |
| H200 Hopper, 8-GPU server | Previous | 141 GB | Large-model inference at lower entry cost | Air, 20–35 kW per rack |
| B200 Blackwell, 8-GPU server | Superseded | 192 GB | Projects already qualified on B200 | Liquid at cluster density |
| H100 Hopper, 8-GPU server | On request | 80 GB | Expanding existing H100 clusters | Air |
Specifications side by side
| B300 | H200 | B200 | H100 | |
|---|---|---|---|---|
| Architecture | Blackwell Ultra | Hopper | Blackwell | Hopper |
| Memory | 288 GB HBM3e | 141 GB HBM3e | 192 GB HBM3e | 80 GB HBM3 |
| Memory bandwidth | 8 TB/s | 4.8 TB/s | 8 TB/s | 3.35 TB/s |
| FP4, dense | 13.5 PFLOPS | — | 9 PFLOPS | — |
| FP8, with sparsity | 9,000 TFLOPS | 3,958 TFLOPS | 9,000 TFLOPS | 3,958 TFLOPS |
| FP64 | ~1.4 TFLOPS | 34 TFLOPS | 40 TFLOPS | 34 TFLOPS |
| NVLink | 5.0, 1,800 GB/s | 4.0, 900 GB/s | 5.0, 1,800 GB/s | 4.0, 900 GB/s |
| 8-GPU server power | ~14.5 kW | ~10.2 kW | ~14.3 kW | ~10.2 kW |
| Indicative server cost | $550k–750k | $300k–400k | $420k–560k | $260k–360k |
- B300 vs B200: same FP8 rate, but 50% more memory and 50% more dense FP4. B300 drops double precision, so it is an AI part, not an HPC one.
- H200 vs H100: the same chip. H200 only adds memory (+76%) and bandwidth (+43%), so it helps inference, not compute-bound training.
- Blackwell vs Hopper: FP4 support, twice the NVLink bandwidth, an on-die RAS engine for large clusters, and confidential computing.
- FP4 figures from NVIDIA's DGX B300 (108 PFLOPS across eight GPUs) and DGX B200 ratings. Costs are indicative project-order ranges for September 2026; see AI server cost.
GB300 NVL72: rack-scale Blackwell Ultra
GB300 NVL72 is not a server but a rack sold and accepted as one system.
| Inside | 72 B300 GPUs and 36 NVIDIA Grace CPUs |
|---|---|
| GPU memory | About 20 TB HBM3e (72 × 288 GB) |
| NVLink | One NVLink 5.0 domain across all 72 GPUs, 130 TB/s |
| Compute | Roughly 1 exaflop dense FP4; about 100 TFLOPS FP64 |
| Power | About 132 kW per rack; HPE advises up to 192 kW of busway |
| Cooling | Direct liquid cooling only |
| Predecessor | GB200 NVL72 — the same design with B200 GPUs |
Why it matters: all 72 GPUs share one NVLink domain, so model-parallel work that would otherwise cross InfiniBand between servers stays on NVLink. That is what trillion-parameter training and high-concurrency serving of the largest reasoning models need.
The trade-off is the site: about 132 kW per rack, liquid cooling designed in from the start, and heavy floor loading. For most enterprise and neocloud projects, HGX B300 servers are simpler to host and to scale in steps. Power planning for both is covered in GPU cluster power requirements.
Which one to choose
Choose B300 when
- You are building a new cluster to run for three years or more
- You serve the largest models: 400B+, long context, reasoning
- Throughput per rack decides the training timeline
- Cost per unit of work matters more than cost per GPU
Choose GB300 NVL72 when
- You train models above a trillion parameters
- You serve very large reasoning models at high concurrency
- Your site, or a partner hall, supports 120–140 kW liquid-cooled racks
Choose H200 when
- You need 70B–405B inference at a lower entry cost
- Your site is air-cooled at 20–35 kW per rack
- You are expanding an existing Hopper cluster
B200 or H100 only when
- B200: the project is already qualified on it, or it can be supplied sooner than B300
- H100: you are expanding an existing H100 cluster without mixing generations
What does a server on each GPU cost? Indicative, September 2026, eight-GPU HGX servers ordered for a project:
- B300: about $550k–750k — the current platform for new clusters
- GB300 NVL72: quoted per project as a complete rack
- H200: about $300k–400k — previous generation, still a working choice for inference
- B200: about $420k–560k
- H100: about $260k–360k, on request
Why quotes differ, full cluster costs, running cost by region and a worked own-vs-rent table: AI server cost in 2026 →
Training vs inference: which GPU for which job
| Workload | Best choice | Lower-cost option | Why |
|---|---|---|---|
| Pre-training | B300; GB300 NVL72 for the largest models | — | H200 adds nothing over H100 here: FP8 compute is identical |
| Inference serving | B300 | H200 | B300: 288 GB and FP4 for throughput. H200: 70B-class models on one GPU, air-cooled |
| Fine-tuning | B300 for full fine-tuning of 70B+ | H200 for LoRA / QLoRA at 7B–70B | Memory decides; several parallel jobs favor B300 |
PCIe option: H200 NVL is a PCIe card with the same 141 GB. A 70B+ model that needed two 80 GB cards fits on one, in standard air-cooled racks. It suits inference, not training, which needs SXM with NVLink. See NVLink vs PCIe.
Server platforms available from Haink
| GPU | Platforms |
|---|---|
| B300 | Eight-GPU HGX B300 systems, including Supermicro ARS-821GL-NHR (B300 configuration) and NVIDIA DGX B300; GB300 NVL72 racks |
| H200 | Supermicro SYS-821GE-TNHR, Dell PowerEdge XE9680, HPE Cray XD670 (8 GPU); H200 NVL PCIe servers with 2–10 GPUs |
| B200 | Supermicro ARS-821GL-NHR (8 GPU), NVIDIA DGX B200 |
| H100 | The same Hopper platforms as H200, on request |
Where Haink delivers and hosts these systems
- Partner data centers in the United States, El Salvador, Hong Kong, Malaysia, Indonesia, the Philippines, Japan and Australia, several of them liquid-cooling ready for B300 and GB300 NVL72. Hosting is optional: GPU cluster hosting with a Tier 1 partner. Capacity and dates by location: AI server cost.
- Export rules: every destination has its own regime. Advanced data-center GPUs for Hong Kong or Mainland China require a US export licence, and every order goes through end-user screening.
- Regional notes: United States, Hong Kong, Singapore, Dubai, Mainland China. For El Salvador, Malaysia, Indonesia, the Philippines, Japan and Australia, capacity, cooling and dates are in the partner data-center table.
Choosing between B300, GB300 NVL72 and H200 for a real project?
Tell us the workload, where it will run and when. We come back with the platform that fits, a costed bill of materials, capacity at a partner data center and a realistic timeline.
Related Resources
- AI Server Supplier
- GPU Infrastructure
- Enterprise AI Infrastructure
- AI Workstation for Small Teams
- Supermicro GPU Servers
- Dell AI Servers
- HPE AI Servers
- All IT Hardware Brands
- AI Server & GPU Procurement Process
Frequently Asked Questions
What is NVIDIA B300 and how does it differ from B200?
NVIDIA B300 (Blackwell Ultra) is the current evolution of Blackwell. It raises GPU memory from 192 GB to 288 GB of HBM3e and dense FP4 compute from 9 to 13.5 PFLOPS per GPU, while keeping B200's FP8 rate and giving up most double-precision performance. For new AI clusters in 2026 it is the forward platform; B200 is largely superseded in new builds.
What is GB300 NVL72?
GB300 NVL72 is NVIDIA's rack-scale Blackwell Ultra system: 72 B300 GPUs and 36 Grace CPUs in one liquid-cooled rack, joined in a single NVLink 5.0 domain with 130 TB/s of bandwidth and about 20 TB of HBM3e. It draws about 132 kW per rack and is built for frontier-scale training and serving the largest models. It succeeds the GB200 NVL72, which uses B200 GPUs.
B300 or H200: which should I choose?
For a new cluster that will run for several years, B300: twice the memory per GPU, FP4 support and several times the inference throughput, at roughly 1.8 times the cost per GPU. H200 is the lower-cost entry for large-model inference, runs in air-cooled racks at 20–35 kW, and suits expansions of existing Hopper clusters.
Do B300 and GB300 NVL72 need liquid cooling?
GB300 NVL72 is liquid-cooled only, at about 132 kW per rack. Eight-GPU B300 servers are offered in air- and liquid-cooled versions at about 14.5 kW each; at cluster density, four servers per rack, liquid cooling is the norm. H200 servers can run air-cooled.
What is the difference between H100 and H200?
H100 and H200 use the same Hopper die with the same compute. H200 has 141 GB of HBM3e against H100's 80 GB, 76% more memory and 43% more bandwidth, so it helps memory-bound work such as large-model inference and does nothing for compute-bound training.
Should I still buy H100?
Only to expand an existing H100 cluster without mixing generations. For new projects, H200 is the lower-cost Hopper option and B300 the current generation.
Which GPU is best for running local LLMs on a small team server?
For small teams, a workstation with one or two RTX PRO 6000 or RTX 6000 Ada GPUs, or an NVIDIA DGX Spark, is more practical than data-center GPUs, which need rack servers, high-density power and, for B300 at scale, liquid cooling. See the AI Workstation page.