GB300 NVL72 vs GB200 NVL72: Pricing & Availability (2026)

CoreWeave powered on the first GB300 NVL72 in July 2025. By December, AWS had it generally available too. Less than a year in, the question isn't whether you can get Blackwell Ultra, it's whether you should pay for it over GB200. This post has the rack price gap, the availability map, and the math for deciding.

All cloud prices in this article are indicative as of 12 Jul 2026 and can fluctuate over time based on 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.

GB300 NVL72 vs GB200 NVL72: Quick Answer

GB300 NVL72 swaps in B300 (Blackwell Ultra) GPUs for B200 GPUs: 288 GB HBM3e per GPU instead of 192 GB, and 2.16 exaflops FP4 per rack instead of 1.44. Analyst estimates put the rack itself at a 10-30% price premium over GB200, not double. Availability still runs through CoreWeave, Azure, and AWS enterprise sales first; GB200 remains the more broadly available rack in 2026.

Both racks share NVLink 5, the Grace CPU pairing, and liquid cooling. The differences are memory, compute, power, and how far down the supply chain each one has reached.

Specs, Memory, and Power: What Changed From GB200 to GB300

GB300 is not a new architecture. It's the same NVL72 rack design, Grace CPUs, and NVLink 5 fabric as GB200, with the GPU swapped from B200 to B300. For the single-GPU breakdown of what separates those two chips, see the B300 Blackwell Ultra guide and the B200 complete guide.

Memory: 288 GB vs 192 GB HBM3e per GPU

Each B300 GPU carries 288 GB of HBM3e against the B200's 192 GB, a 50% increase per GPU. Across a full 72-GPU rack, CoreWeave puts that at up to 21 TB of GPU memory, versus roughly 13.4 TB on GB200 NVL72, a figure it describes as 1.5x more than the prior generation. AWS's own comparison of P6e-GB300 to P6e-GB200 UltraServers confirms the same 1.5x GPU memory jump at the instance level.

The extra headroom matters most for models that barely fit or don't 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, or production inference serving very long context windows at high concurrency, where KV cache alone can eat hundreds of gigabytes per rack.

Compute: 2.16 vs 1.44 Exaflops FP4

AWS states P6e-GB300 delivers 1.5x the FP4 compute of P6e-GB200 without sparsity. Applied to GB200 NVL72's well-documented 1.44 exaflops FP4 (with sparsity) figure, that puts GB300 NVL72 at 2.16 exaflops with sparsity, a number Spheron's own GB300 rack specifications confirm directly. At the chip level, that's the B300's 15 dense petaFLOPS FP4 against the B200's 9.

That compute gain compounds with FP4 quantization support in current serving stacks. For the mechanics of why FP4 throughput translates into real cost savings (and where it doesn't), see the FP4 quantization on Blackwell guide.

Power and Cooling: Why GB300 Runs Hotter

A single B300 GPU runs at 1,400W TDP, 40% above the B200's 1,000W. At rack scale, that pushes GB300 NVL72 to roughly 132 kW of power draw against GB200's ~120 kW, per Spheron's own rack specifications. Both racks require direct liquid cooling; air cooling was never viable at either density. What changes with GB300 is the margin: a data center already stretched to support 120 kW racks doesn't automatically have 132 kW of spare capacity per rack, which is one more reason GB300 deployment has stayed concentrated at operators who control their own power and cooling buildout. For more on how power, not chip supply, is increasingly the actual constraint on GPU availability, see AI data center power constraints in 2026.

Networking scales too: GB300 racks ship with ConnectX-8 NICs at 800 Gb/s per port, up from ConnectX-7's 400 Gb/s on most GB200 deployments, which matters once you're clustering multiple racks together.

Cloud Availability: Who Is Actually Renting GB300 NVL72 in 2026

GB300 NVL72 reached cloud availability faster than most rack-scale launches, but "available" still means an enterprise sales conversation at every major provider, not a self-serve instance you spin up in a console.

CoreWeave: First to Deploy (July 2025), First to Sell

CoreWeave announced the first industry deployment of GB300 NVL72 on July 3, 2025, ahead of every other cloud provider. The company claims up to a 10x boost in user responsiveness and a 5x improvement in throughput per watt over prior-generation hardware, alongside the 21 TB memory-per-rack figure. Being first gave CoreWeave a run of large enterprise contracts before AWS or Azure had GB300 capacity to sell, and it remains the provider with the longest GB300 track record heading into H2 2026.

Azure and AWS: The Hyperscaler Rollout

Microsoft Azure announced the industry's first supercomputer-scale production GB300 NVL72 cluster on October 9, 2025, built for OpenAI: more than 4,608 Blackwell Ultra GPUs on Quantum-X800 InfiniBand, with roughly 92.1 exaflops of FP4 inference across the full cluster and 37 TB of fast memory per VM. That's a dedicated deployment for one customer at a scale most teams will never approach, but it confirmed GB300 could run at supercomputer scale in production, not just in a lab.

AWS followed with general availability of EC2 P6e-GB300 UltraServers on December 2, 2025, close to 20 TB of GPU memory per UltraServer and the 1.5x memory and compute gains over P6e-GB200 described above. Access still runs through an AWS sales representative rather than the standard EC2 console flow.

Where GB200 NVL72 Still Has the Availability Edge

GB200 NVL72 has been shipping for over a year longer and is available across more providers with fewer strings attached. CoreWeave, Oracle Cloud, Azure, and Google Cloud all list GB200 capacity, and CoreWeave and Oracle publish indicative on-demand rates rather than requiring a quote for every deployment: roughly $10.50 to $27 per GPU-hour across CoreWeave, Oracle, and Azure, per our GB200 rack spec breakdown. If your timeline is measured in weeks rather than quarters, GB200's wider bench of providers is a real advantage independent of price.

Rack Price: Is GB300 Really Double the Cost of GB200

No. Analyst estimates put the GB300 premium closer to 10-30% over GB200, not double, though neither NVIDIA nor the cloud providers publish an official number.

SemiAnalysis estimated a GB200 NVL72 rack at $3.1M bare and $3.9M all-in once networking and storage are included. Loop Capital analyst Ananda Baruah, evaluating Apple's reported Blackwell Ultra order, put GB300 NVL72 at $3.7M to $4.0M per rack. Line those up and the gap is real but modest: GB300's low estimate sits roughly 19% above GB200's bare-rack figure, and its high estimate is within a few percent of GB200's all-in figure.

The two estimates come from different analysts using different methodologies and don't specify identical scope (bare rack versus fully networked), so treat the comparison as directional rather than exact. What it rules out is the "double the price" assumption that circulated when GB300 first launched. The premium is closer to what you'd expect from a mid-cycle refresh than a generational leap in cost.

Cost Per GPU and Cost Per Token: Is the Blackwell Ultra Premium Worth It

Divide the rack estimates by 72 GPUs and the picture sharpens. GB200 comes out to roughly $43,000 per GPU bare or $54,000 all-in. GB300 comes out to roughly $51,000 to $56,000 per GPU. At the high end, that's a similar per-GPU price to GB200's all-in figure, for 50% more memory and 50% more FP4 compute per GPU. Spec-for-spec, GB300 is arguably the better deal even where the absolute rack price is higher.

Whether that translates to lower cost per token depends entirely on whether your workload is compute-bound or memory-bandwidth-bound. The single-GPU math is easier to reason about and scales up 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 memory bandwidth, not compute, is the bottleneck. The same logic holds at rack scale: workloads that saturate FP4 compute across all 72 GPUs see the fuller benefit of GB300's premium; workloads that are memory-bandwidth-bound or don't come close to filling a GB200 rack's 13.4 TB won't see much of it.

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

Decision Framework: Which Rack Should You Actually Rent

Work through these in order.

1. Does your model plus KV cache fit in 13.4 TB? If yes, GB200 NVL72 is very likely the better economic call. You get the same NVLink fabric and Grace CPU pairing at a lower rack price and wider provider availability.
2. Is your workload genuinely compute-bound at FP4? Large-batch inference and dense or MoE training at 200B+ parameters are where GB300's 2.16 exaflops shows up as real throughput, not just a spec sheet number. If your bottleneck is memory bandwidth rather than compute, the B300 advantage compresses.
3. Can your timeline tolerate a hyperscaler sales cycle? GB300 access at CoreWeave, Azure, and AWS still runs through enterprise procurement. If you need capacity in weeks, GB200's wider bench of providers with published rates is the more reliable path.
4. Do you actually need a full rack? Most teams running 7B-70B models don't. A non-rack B200 or B300 cluster clears the bar for the vast majority of production inference and fine-tuning workloads at a fraction of rack-scale cost. Spheron's live on-demand B200 rate sits at $9.36/hr per GPU as of 12 Jul 2026, with spot from $5.34/hr, no rack commitment required.

If none of the first three 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 requirements, the same variables that drive every rack-scale quote industry-wide; submit your GPU count and timeline and the team confirms availability and pricing within a business day.

For teams that decide a full rack isn't the right call yet, B200 and B300 are available today with per-minute billing and no long-term commitment. B200 is on-demand from $9.36/hr per GPU with spot from $5.34/hr; B300 rates shift with inventory, so check the live rate before you commit. For multi-node setup once you've 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 is the architecture that eventually replaces both of these racks; the Vera Rubin NVL72 cloud availability guide covers where that stands heading into H2 2026.

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 →