NVIDIA GB200 Price in 2026: What a GPU, a Tray and an NVL72 Rack Cost

For a decade, the unit of AI compute was the 8-GPU server. Models that outgrew it were split across servers over InfiniBand, and every token paid for that hop. The GB200 NVL72 moves the boundary to the rack: 72 GPUs share one NVLink fabric and behave, for scheduling purposes, like one very large GPU. So the useful question is not what one GB200 costs. It is what it costs to keep a model inside one NVLink domain.

What a GB200 is: superchip, tray and rack

The name covers three levels of hardware:

• Superchip. One Grace CPU and two Blackwell GPUs on one module, joined by NVLink-C2C at 900 GB/s. The CPU and GPUs share a coherent memory space, so the GPUs can address the CPU's LPDDR5X as well as their own HBM.¹²
• Compute tray. Two superchips: two Grace CPUs and four GPUs, liquid-cooled, with 1.7 TB of fast memory.²
• NVL72 rack. 18 compute trays and nine NVLink switch trays: 36 Grace CPUs and 72 GPUs in one NVLink domain with 130 TB/s of GPU-to-GPU bandwidth.¹²³

GB200 hardware ships in trays and racks, and rented capacity is carved from them.

GB200 specifications

Superchip and rack columns are NVIDIA's NVL72 specifications. Per-GPU values divide the rack figures by 72 (CPU memory by two GPUs per superchip). Dense throughput is half of NVIDIA's sparse figure.¹ A rack holds 30.2 TB of fast memory in total, HBM and LPDDR5X combined.³

Per GPU, a GB200 is close to a B200: the same 8 TB/s and about 11% more dense FP4 compute.⁴ The difference is what surrounds the GPU.

When the NVLink domain pays off

Inside an NVL72, each GPU talks to the other 71 at 1.8 TB/s. Between racks, each GPU gets one 400 Gb/s InfiniBand port.³ Counting both directions, NVLink is roughly 18 times faster. Workloads that move a lot of data between GPUs on every token benefit most:

• Large mixture-of-experts models. Expert parallelism sends each token to the GPUs that hold its experts, then gathers the results. When all experts sit in one NVLink domain, that all-to-all never touches the slower network. NVIDIA reports 30 times faster real-time inference than H100 on a 1.8-trillion-parameter MoE model, a result it credits to FP4 and the 72-GPU NVLink domain.¹²
• Models larger than one server. A 2-trillion-parameter model at FP8 needs about 2 TB for weights alone. That is more than an 8-GPU B200 server holds (1.4 TB) and nearly all of an 8-GPU B300 server (2.1 TB).⁴ On an NVL72 it uses about 15% of the rack's HBM, leaving the rest for KV cache and concurrent requests.
• Memory spill to the CPU. The 900 GB/s coherent link lets KV cache or embeddings overflow into Grace memory without a PCIe round trip.²

A B200 in an 8-GPU server is the better buy when the model fits in 1.4 TB with room for its KV cache. That covers most 70B to 400B-class LLMs, nearly all image and video models, and any job that runs as independent replicas. Those workloads gain nothing from the larger domain, and the B200 costs 20% less per GPU-hour at list price. The B200 price guide covers that GPU.

What a GB200 costs to rent

On fal, the GB200 lists at $9.99 per GPU-hour, and as low as $5.89 per GPU-hour.⁵ It sits between the B200 and the B300 in NVIDIA's Blackwell lineup:

Prices and VRAM are from fal.ai/pricing. One GB200 GPU-second costs about $0.00278 at list, or $0.00164 at the "as low as" rate.

What each GB200 unit costs per hour

At the per-GPU rate, the hardware levels above price out as follows:

Only a full rack gives a model the 72-GPU NVLink domain. Across a rack, the gap between list and the "as low as" rate is $295.20 per rack-hour, or about $212,500 over 30 days. For when holding capacity beats paying per second, see dedicated GPUs vs serverless inference.

GB200 vs B200 vs B300: what the premium buys

Premium over a B200, per GPU. NVIDIA's per-GPU specs from 8-GPU HGX systems and the NVL72; fal's rates.¹⁴⁵

The B300 premium buys a bigger GPU. The GB200 premium buys a bigger domain. At "as low as" rates the two premiums are almost the same, so the choice follows the model: memory per GPU for long context on one server, NVLink reach for models that span many GPUs.

Cost to keep one terabyte of HBM inside one NVLink domain, per hour. The largest domain on each platform, divided by its HBM.¹⁴⁵

Per terabyte, the GB200 rack costs about the same as a B200 server at "as low as" rates, and about 18% more at list. What changes is the ceiling: a B200 server stops at 1.4 TB, and the NVL72 goes to 13.4 TB before any traffic crosses InfiniBand. For the B300's long-context case, see reserving NVIDIA B300 capacity.

Buy or rent at rack scale

GB200 racks are priced per deal. In May 2024, HSBC analysts estimated about $3 million for an NVL72 rack, $1.8 million for an NVL36 and $60,000 to $70,000 per superchip.⁶ At that rack price, the same money rents a full rack for:

• 4,171 hours (about 174 days) at $719.28 per hour, list price.
• 7,074 hours (about 295 days) at $424.08 per hour, the "as low as" rate.

The purchase price is only the start. An NVL72 draws about 120 kW, or roughly 2,880 kWh a day, through eight power shelves.⁷ Its GPUs and CPUs are liquid-cooled through rack manifolds and cold plates, so the facility needs liquid cooling at the rack.⁷ Add InfiniBand between racks, spares and staff. Owning makes sense if you already operate liquid-cooled data centers and can keep all 72 GPUs busy for years. The H200 price guide runs the same buy-or-rent test for an air-cooled GPU.

Getting GB200 capacity on fal

GB200 capacity is arranged with the fal team. Contact sales at fal.ai/enterprise with your model, its memory footprint and how many GPUs it needs to share one NVLink domain. To run your own models on fal, request Serverless access.

FAQ

How much does an NVIDIA GB200 cost?

GB200 hardware is priced per deal. Analysts estimated about $3 million for an NVL72 rack and $60,000 to $70,000 per superchip in May 2024.⁶ On fal, a GB200 rents for $9.99 per GPU-hour at list, or as low as $5.89.

What is the GB200 price per hour?

$9.99 per GPU-hour at list price on fal, about $0.00278 per second. It goes as low as $5.89 per GPU-hour.

What does a GB200 NVL72 rack cost per hour?

72 GPUs at $9.99 is $719.28 per hour at list price, or $424.08 at the "as low as" rate.

GB200 vs B200: which should I choose?

Choose a B200 if the model and its KV cache fit in an 8-GPU server, because it costs 20% less per GPU-hour at list. Choose GB200 when a model or its experts need more GPUs than one server holds and must stay on NVLink.

References