NVIDIA B300 vs B200 vs H100: Rental Price and VRAM Guide

Decision first: H100 is the practical starting point when hourly budget and provider choice matter most. B200 is the more interesting upgrade when 80GB per GPU forces awkward model sharding: at this snapshot it offers 2.25 times the memory of H100 for about 1.91 times the lowest hourly rate. B300 should not win merely because it is newer. Its 288GB per GPU is valuable when memory density changes the cluster design, but the current tracked entry price is $4.99/hour and 7 providers are present in the comparable data.

The quote sheet on July 26

We pulled the latest comparable on-demand rows from ComputeUnion before writing this guide. The lowest observed rates were $1.4678/hour for H100, $2.80/hour for B200, and $4.99/hour for B300. H100 had 19 tracked providers, B200 had seven, and B300 had one. Those counts matter: a one-provider price is a useful market signal, but it is not yet a deep market.

GPU	Official memory	Lowest tracked rate	Tracked providers	Observed price per GB-hour
H100 SXM 80GB	80GB HBM3	$1.4678/hr	19	$0.0183
B200 SXM 180GB	180GB HBM3e	$2.80/hr	7	$0.0156
B300 SXM 288GB	288GB HBM3e	$4.99/hr	7	$0.0173

Price per GB-hour is included only as a memory-capacity lens. It is not a performance benchmark. A workload can still run faster or cheaper on another GPU because kernel support, precision, batch size, interconnect, and utilization change the result.

What a day and a month actually cost

A low hourly number becomes easier to judge after multiplying it by time. The table assumes one GPU running continuously for 24 hours or 720 hours. It excludes storage, CPU, networking, egress, taxes, reserved-use discounts, and multi-GPU node minimums.

GPU	24 hours	30 days (720h)	Extra spend versus H100/month
H100 SXM 80GB	$35.23	$1056.82	Baseline
B200 SXM 180GB	$67.20	$2016.00	+$959.18
B300 SXM 288GB	$119.76	$3592.80	+$2535.98

The B200 premium is easier to defend when 180GB avoids splitting a model over multiple 80GB GPUs. The B300 premium needs a stronger argument: 288GB must remove enough cards, communication, engineering work, or failed runs to recover more than $4,000 per GPU-month versus the current H100 floor.

Three decisions that look similar but are not

1. “The model fits” is not the same as “the job fits”

Weights fitting in memory only establishes a lower bound. Production inference also needs KV cache, runtime buffers, batching headroom, and room for the serving framework. Training and fine-tuning add optimizer states, gradients, activations, checkpoints, and communication buffers. If an 80GB plan is already close to the limit on paper, it is usually too close in production.

2. Fewer GPUs do not automatically mean a cheaper node

Moving from H100 to B200 or B300 can reduce the card count for a memory-bound model, but rental platforms may sell Blackwell systems as full nodes or require a minimum allocation. Compare the billable unit before comparing the GPU label. A four-GPU minimum can erase the saving from fitting the model on fewer cards.

3. The cheapest listing is not always the safest production choice

The H100 floor in this snapshot comes from a low-availability listing, while other H100 offers sit higher. A production team should compare inventory, region, interconnect, storage, support, interruption policy, and replacement capacity. ComputeUnion keeps all current provider rows on the GPU detail pages so the lowest price is not presented as the only answer.

Which GPU should you test first?

Workload condition	Start with	Reason to test the next tier
Model and runtime fit comfortably inside 80GB	H100	Move only if Blackwell throughput or power efficiency improves cost per accepted job in your benchmark.
80GB forces tensor parallelism or repeated out-of-memory tuning	B200	180GB may reduce sharding and communication while keeping a broader market than B300.
Very large inference, long-context cache, or training state needs maximum memory density	B300	288GB can simplify topology, but validate node minimums and real availability before committing.
Early proof of concept with uncertain utilization	H100 or the cheapest available compatible GPU	Do not pay a generation premium while the workload is idle or still changing.

A procurement test that prevents an expensive guess

1. Run the same model revision, precision, framework, batch, and sequence lengths on each candidate.
2. Record accepted jobs per hour, peak memory, tokens or samples per second, retries, and engineering time.
3. Use the actual platform billable unit: single GPU, complete node, reserved block, or spot instance.
4. Calculate cost per accepted job. Hourly price alone cannot show failed runs or idle memory.
5. Repeat the test with the expected production concurrency, not only a one-request demo.

Conclusion

At the current ComputeUnion snapshot, H100 is the supply-and-price baseline, B200 has the strongest memory-per-dollar story, and B300 is a specialized memory-density purchase rather than a default upgrade. The right answer can change when inventory and provider coverage change, so use this article for the decision method and the linked GPU pages for the live quote.

Sources and method

Memory sizes and system characteristics come from the NVIDIA HGX AI Factory reference architecture, the NVIDIA MIG supported GPU table, and the NVIDIA H100 product page. Rental prices and provider counts are a ComputeUnion on-demand market snapshot checked on 2026-07-26. They are not fixed quotes or NVIDIA list prices. See the ComputeUnion methodology for how price observations are handled.