NVIDIA B300 Price in 2026: Blackwell Ultra Specs, Hourly Rates and Reservations

The B300, NVIDIA's Blackwell Ultra GPU, shows where inference hardware is heading: more memory per GPU and more compute at 4-bit precision, with no gain at the FP8 and FP16 precisions most models still run at. That makes it a specialist's GPU. For long-context reasoning and very large models, it can be the cheapest way to serve. For most other work, a B200 does the same job for less.

What changed from the B200

• Memory. 288 GB of HBM3e per GPU, 50% more than the B200's 192 GB, at the same 8 TB/s.¹
• FP4 compute. 15 petaFLOPS of dense NVFP4 per chip, against 10 for Blackwell. In 8-GPU HGX servers that is 13.5 against 9 per GPU.¹²
• Attention. NVIDIA rates attention-layer compute at up to 2 times Blackwell.¹
• Networking. Each GPU in a DGX B300 gets an 800 Gb/s ConnectX-8 port, twice the 400 Gb/s in a DGX B200.³⁴

FP8 and FP16 throughput did not change. FP64 and INT8 throughput fell sharply.²

B300 specifications compared with the B200, GB200 and H200

B300 and B200 throughput is per GPU in NVIDIA's 8-GPU HGX servers.² GB200 figures are per GPU in a GB200 superchip.⁵ H200 figures are for the H200 SXM.⁶ Dense throughput is half of NVIDIA's sparse figure, except FP4.

Three details matter for sizing:

• Usable memory is about 260 GB per GPU. NVIDIA lists 2.1 TB in total for an 8-GPU DGX B300.³
• FP64 and INT8 are nearly gone. An HGX B300 has 10 TFLOPS of FP64, against 296 on an HGX B200, and INT8 drops from 72 to 3 sparse petaOPS.² Scientific computing and INT8 models belong on a B200 or Hopper GPU.
• Only FP4 is faster. At BF16 or FP8, a B300 matches a B200.

When a B300 is worth it, and when a B200 is cheaper

A B300 pays off when the workload uses its extra memory, FP4 throughput or faster attention:

• Long-context reasoning and large mixture-of-experts models. Every expert must stay in memory, and long contexts make attention a growing share of the time. More memory holds more weights, larger KV caches and more concurrent requests.
• High-concurrency serving. Each active session holds KV cache, so more memory means more sessions per runner before requests queue. If your current GPUs keep up at peak, you probably do not need a B300.
• Large video models and multi-model pipelines. Long video sequences use both the attention gain and the memory, and chained models can stay resident on one GPU.
• FP4-quantized inference. The 1.5 times FP4 gain helps compute-bound prefill more than bandwidth-bound decode. Validate output quality first, because not every model tolerates 4-bit weights.

A B200 is cheaper for bandwidth-bound models that fit in 192 GB. Image generation and small models rarely need 288 GB and run well on far cheaper GPUs such as the L40 or RTX PRO 6000.⁷

What a B300 costs to rent

On fal, the B300 lists at $12.99 per GPU-hour. Here it is next to the rest of NVIDIA's current data-center lineup:⁸

Both prices are per GPU-hour. One B300 second costs about $0.00361 at list price, or about $0.00166 at the "as low as" rate.

The B200 price guide explains hourly, per-second and reserved billing. B300s ship mostly in 8-GPU systems, so tell the fal team whether you need single GPUs or full 8-GPU nodes.⁸

Worked B300 cost examples

The bursty month uses one-eighth of the warm month's GPU time, so it costs one-eighth as much. On a GPU this expensive, scaling to zero between bursts matters as much as the hourly rate, at the cost of cold starts when traffic returns.

B300 or B200 on price

Dollars per GPU-hour, using fal's rates and NVIDIA's per-GPU specs. Memory uses fal's listed VRAM.

• At list price, the B200 is cheaper on every measure. The B300 costs about 1.63 times as much for the same 8 TB/s, so choose it only when the model or its KV cache does not fit on a B200.
• At "as low as" rates, the B300 is the cheapest of the four per FP4 petaFLOPS and per GB, and only about 1.09 times the B200's rate. For large, FP4-ready models at that rate, it is the stronger buy.

Reserving B300 capacity on fal

To reserve B300 capacity, contact fal sales at fal.ai/enterprise with your model, expected traffic and how many GPUs you need around the clock.

At $12.99 list against $5.99 "as low as", the gap is $168 per GPU-day, which pays off for a baseline that keeps B300s busy around the clock. Send peaks above that baseline to Serverless runners, which autoscale and cost nothing while none are running.⁹ keep_alive, min_concurrency and max_concurrency control how many runners stay warm.¹⁰ For where the crossover falls, see dedicated GPUs vs serverless inference.

Buy or rent a B300

B300s are sold mainly inside 8-GPU servers, priced per deal. To compare a quote, renting one GPU for a year (8,760 hours) costs:

• $113,792.40 at list price.
• $52,472.40 at the "as low as" rate.

A purchase adds power, liquid cooling, networking, space and staff, and costs the same whether busy or idle. An 8-GPU DGX B300 draws about 14 kW.³ Owning makes sense mainly if you already run large data centers and can keep B300s busy around the clock.

To start, request Serverless access and follow the Quick Start.

FAQ

How much does an NVIDIA B300 cost?

B300s are sold mainly inside 8-GPU servers, priced per deal. On fal, the B300 is $12.99 per GPU-hour at list price, and as low as $5.99 per GPU-hour.

What is the B300 price per hour?

$12.99 per GPU-hour at list price, or as low as $5.99. Eight GPUs for one hour cost $103.92 at list price.

What is the B300 price compared with the B200?

fal lists the B300 at $12.99 and the B200 at $7.99 per GPU-hour, about 1.63 times. At "as low as" rates, they are $5.99 and $5.49, about 1.09 times.

Can I reserve B300 capacity on fal?

Yes. Contact fal sales at fal.ai/enterprise.

References