GB300 NVL72 vs GB200 NVL72: The Real Price Gap (2026)

Every spec in this comparison comes from NVIDIA's own GB300 NVL72 and GB200 NVL72 specification tables, checked 16 August 2026. That matters here more than usual, because the two most-repeated claims about this comparison are both wrong.

All cloud prices in this article are indicative as of 16 Aug 2026 and fluctuate with GPU availability. Check current GPU pricing for live rates.

For the full GB200 NVL72 architecture and rack spec table, see the GB200 NVL72 guide. If you're still deciding whether you need rack-scale at all versus a standalone GPU, the H200 vs B200 vs GB200 comparison covers that fork first.

Spheron carries both racks in inventory right now. If you want to reserve one, fill out the form on the GB200 NVL72 or GB300 NVL72 rental page and our team confirms availability, region, and pricing within a business day.

Quick Answer

GB300 NVL72 swaps B200 GPUs for B300 Blackwell Ultra: 288 GB HBM3e per GPU instead of 186 GB in-rack, 20 TB of rack GPU memory instead of 13.4 TB, and 1.5x the dense FP4. Sparse FP4 is unchanged. FP64 and INT8 fall off a cliff. Availability is quote-only everywhere; GB200 is the only NVL72 rack with a published hourly rate.

Both racks share NVLink 5 at 130 TB/s, the same Grace CPU count and core count, and liquid cooling. The differences are memory, dense FP4, networking, and what NVIDIA gave up to get there.

The Two Things Everyone Gets Wrong

Wrong claim 1: "GB300 delivers 2.16 exaflops of FP4"

It does not, and that number appears nowhere in NVIDIA's specifications. NVIDIA's GB300 NVL72 table lists FP4 Tensor Core performance as 1,440 PFLOPS with sparsity and 1,080 PFLOPS dense. GB200 NVL72 lists 1,440 PFLOPS sparse and 720 PFLOPS dense.

So the sparse figure is identical between the two racks. The real generational gain is on dense FP4: 720 to 1,080 PFLOPS, exactly the "1.5x more dense FP4 Tensor Core FLOPS" NVIDIA claims in its own GB300 overview copy. The 2.16 exaflops figure appears to come from doubling the dense number, or from applying a 1.5x multiplier to the sparse number that NVIDIA never applied. Earlier versions of this page repeated it. It has been corrected here.

This matters practically. If your serving stack uses structured sparsity, moving from GB200 to GB300 buys you memory and attention performance, not peak FP4 throughput. If your stack runs dense FP4, which most production inference does, you get the full 1.5x.

Wrong claim 2: "GB200 NVL72 has 192 GB per GPU"

A standalone B200 is 192 GB. Inside the NVL72 rack it is not. NVIDIA's GB200 NVL72 table gives the Grace Blackwell Superchip 372 GB of HBM3e across its two GPUs, and the full rack 13.4 TB. Divide either way and you get 186 GB per GPU.

It is a 3% difference and it rarely changes a decision, but it propagates into every capacity calculation built on top of it, and 13.4 TB is the number that actually constrains what fits in a rack.

The Thing Nobody Mentions: GB300 Is Much Worse at FP64

This is the most consequential difference in the comparison and it is almost absent from the coverage.

Blackwell Ultra is an inference-optimised part. NVIDIA spent the transistor budget on FP4 and attention throughput and took it out of double precision and INT8. For LLM inference that is the correct trade and you will never notice. For anything else it is severe:

• Traditional HPC and scientific simulation. Computational fluid dynamics, molecular dynamics, climate modelling, finite element analysis and most linear algebra workloads run in FP64. A GB200 NVL72 rack is roughly 29x faster at FP64 than a GB300 NVL72 rack. Choosing GB300 for those workloads would be a serious and expensive mistake.
• INT8 inference. Older quantised serving paths that target INT8 rather than FP4 or FP8 lose 96% of their peak throughput. If you have not migrated your quantisation to FP8 or FP4, GB300 will underperform badly.
• Mixed AI and HPC clusters. If a single cluster serves both a research computing group and an inference team, GB200 is the far safer common denominator.

If you run double precision, stop reading the FP4 comparisons. GB200 NVL72 is the correct rack, and it is also the cheaper one.

Memory: Where GB300 Genuinely Wins

Each B300 carries 288 GB of HBM3e against the B200's 186 GB in-rack, a 55% increase per GPU. Across the rack that is 20 TB of GPU memory versus 13.4 TB, and 37 TB of fast memory versus roughly 30 TB once Grace LPDDR5X is included.

The extra headroom matters for models that barely fit or do not fit in 13.4 TB today. DeepSeek R1 at 671B parameters needs roughly 700-750 GB in FP8 for weights and runtime buffers, which a GB200 rack swallows easily with room for large KV caches. Where GB300's memory earns its keep is trillion-parameter dense or MoE models, and production serving of very long context at high concurrency, where KV cache alone can consume hundreds of gigabytes per rack.

GB300 also doubles scale-out networking, from ConnectX-7 at 400 Gb/s per GPU to ConnectX-8 at 800 Gb/s. That only matters once you are clustering multiple racks; inside a single rack both use the same 130 TB/s NVLink 5 fabric.

For the single-GPU breakdown of what separates the two chips, see the B300 Blackwell Ultra guide and the B200 complete guide. For why FP4 throughput translates into real cost savings, and where it does not, see the FP4 quantization on Blackwell guide.

Power: nobody publishes it

NVIDIA publishes no rack power figure for either system. Operator and analyst reporting generally places GB200 NVL72 around 120 kW and GB300 NVL72 around 135-140 kW, but these are not vendor specifications and they vary with configuration and cooling design. Both require direct liquid cooling; air cooling was never viable at either density. The practical point stands regardless of the exact number: a facility provisioned for GB200 racks does not automatically have headroom for GB300 racks, which is one reason GB300 deployment has stayed concentrated among operators who control their own power and cooling. For more on power as the actual constraint on GPU availability, see AI data center power constraints in 2026.

Who Is Actually Renting These in 2026

We checked provider pricing pages directly on 16 August 2026. The result is stark: exactly one NVL72 rack has a published on-demand hourly rate anywhere, and it is GB200.

Sources: CoreWeave pricing, Nebius pricing, Together AI GB300 NVL72, Lambda, Azure.

Be sceptical of GB300 hourly rates you find on aggregator sites. We chased several widely-cited figures and none of them survived a check against the provider's own pricing page. A related trap: some aggregators quote HGX B300 and HGX B200 rates, which are real, and present them as NVL72 rack rates. Those are 8-GPU boards, not 72-GPU racks, and the numbers are not comparable.

Azure's own GB300 pricing breakdown covers exactly what is confirmed on cost versus what remains sales-gated. For CoreWeave's full rate card and contract model, see CoreWeave GPU pricing 2026.

GB200 still has the availability edge

GB200 NVL72 has been shipping over a year longer, is available across more providers, and is the only one you can price without a sales call. If your timeline is measured in weeks rather than quarters, that is a real advantage independent of the spec sheet.

Rack Price: The Estimates Genuinely Disagree

Neither NVIDIA nor the cloud providers publish rack pricing, and unlike most spec questions this one does not resolve cleanly.

Those two GB300 estimates are roughly 1.7x apart. That is not a rounding difference, it is a disagreement about what the product costs. The Loop Capital figure implies a modest 10-30% premium over GB200, consistent with a mid-cycle refresh. The Tom's Hardware figure implies close to double, and describes an inference-optimised configuration.

We are not going to pretend to resolve it. What you can safely conclude:

1. GB300 costs meaningfully more than GB200. Every estimate agrees on direction.
2. The premium is somewhere between 10% and 100%, and anyone quoting a precise figure is over-claiming.
3. Divided by 72 GPUs, GB200 works out to roughly $43,000 per GPU bare or $54,000 all-in. GB300 lands between roughly $51,000 and $90,000 per GPU depending on which estimate you believe.

At the low GB300 estimate you are paying a similar per-GPU price to GB200 all-in for 55% more memory and 50% more dense FP4, which is a good deal. At the high estimate you are not. Since neither number is verifiable, treat rack purchase economics as an input you have to source from an actual quote.

Cost Per Token: Where the Premium Pays Off

Whether the premium translates to lower cost per token depends on whether your workload is compute-bound or memory-bandwidth-bound. Note that memory bandwidth is unchanged between the two racks at up to 576 TB/s, so GB300's advantage is capacity and dense compute, not bandwidth.

The single-GPU math scales cleanly: the B300 vs B200 cost-per-token breakdown found roughly 26% lower cost per million tokens on B300 for large-batch Llama 3.3 70B inference, and a smaller gap for smaller models where bandwidth rather than compute is the bottleneck. The same logic holds at rack scale. Workloads that saturate dense FP4 across all 72 GPUs see the fuller benefit; workloads that are bandwidth-bound, or that do not come close to filling 13.4 TB, will not.

For a broader view of where rack-scale sits against every other option, the GPU cloud pricing comparison covers the non-rack alternatives.

Decision Framework

Work through these in order.

1. Do you need FP64 or INT8? If yes, choose GB200 NVL72 and stop here. GB300 is roughly 29x slower at FP64 and 30x slower at INT8. This single question decides more cases than any other and is the one most comparisons omit.
2. Does your model plus KV cache fit in 13.4 TB? If yes, GB200 is very likely the better economic call. Same NVLink fabric, same Grace pairing, lower price, published hourly rate, wider availability.
3. Is your workload dense-FP4 compute-bound? Large-batch inference and dense or MoE training at 200B+ parameters is where GB300's 1,080 PFLOPS dense shows up as real throughput. If you rely on structured sparsity, note that both racks are identical at 1,440 PFLOPS sparse.
4. Can your timeline tolerate a sales cycle? GB300 is quote-only everywhere. If you need capacity in weeks, GB200's published rates are the more reliable path.
5. Do you actually need a full rack? Most teams running 7B-70B models do not. A non-rack B200 or B300 cluster clears the bar for the vast majority of production inference and fine-tuning at a fraction of rack-scale cost. Watch for node-minimum policies when shopping non-rack quotes: our Genesis Cloud pricing breakdown covers a provider whose H100 nodes are 8-GPU-only, which turns a competitive headline rate into a much bigger bill if you only need one or two cards.

If none of the first four push you toward GB300, default to GB200 or skip rack-scale entirely.

Getting Access on Spheron

Spheron keeps both GB200 NVL72 and GB300 NVL72 open for reservation outside the standard hyperscaler sales process, sized from a single 8-GPU node up to a full rack. Pricing depends on quantity, commitment length, region, and networking, the same variables that drive every rack-scale quote industry-wide.

For teams that decide a full rack is not the right call, B200 and B300 are available today with per-minute billing and no commitment. Live rates on 16 Aug 2026: B300 SXM6 at $9.08/hr on-demand and $5.81/hr spot, B200 SXM6 at $9.36/hr on-demand and $5.34/hr spot. Note that B300 currently prices below B200 on-demand while carrying 50% more memory, so check the live rate rather than assuming the newer part costs more. Aggregating across multiple data center partners is what makes that kind of inversion visible in the first place.

Pricing fluctuates based on GPU availability. The prices above are based on 16 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

For multi-node setup once you have picked a configuration, the distributed training guide walks through PyTorch DDP and DeepSpeed on multi-GPU clusters.

Looking further out, NVIDIA's Vera Rubin NVL72 eventually replaces both racks. The Vera Rubin NVL72 cloud availability guide covers where that stands, and our Vera Rubin NVL72 specs guide has the full per-GPU breakdown, including why the "H300" many people search for does not exist. Outside the NVIDIA ecosystem, Huawei's rack-scale answer is covered in our Ascend 950 vs B300 and B200 comparison.

Most teams don't need to choose between GB300 and GB200 blind. Get a quote sized to your actual GPU count and workload before committing to either rack.

Reserve GB300 NVL72 on Spheron →