The B200 costs more per hour. Sometimes it costs less overall. Here's the framework for knowing which is true for your job.
B200 on-demand rates run $4.99–$6.19/hr per GPU versus $3.50–$4.54/hr for H200, but the B200 is worth the premium only when your workload is compute-bound, FP4-eligible, or large enough that faster completion beats a lower hourly rate.
Key takeaways
Every team evaluating B200 GPU rental against H200 runs into the same problem: the B200 costs more per hour, but the spec sheet says it's faster. Neither number tells you what you'll actually pay. The real comparison is total workload cost: how much you spend to finish the job, not the sticker price on the invoice.
This framework walks through where the B200 premium pays off, where H200 remains the better buy, and how to run the math yourself before committing to either GPU for your next training run or inference deployment.
The headline numbers are straightforward. H200 SXM wholesale on-demand pricing runs $3.50 to $4.54 per GPU per hour. B200 SXM lands at $4.99 to $6.19 per GPU per hour. That's a premium of roughly 40-60% depending on provider and region. Hyperscaler pricing from AWS, Azure, and GCP runs 2-3x higher for both GPUs, so the gap compounds fast if you're comparing hyperscale quotes instead of wholesale rates.
B200 SXM wholesale on-demand rates sit 40-60% above H200 SXM at most providers as of May 2026.
The H200 is a memory refresh on the same Hopper compute die as the H100: same TFLOPS, more VRAM, more bandwidth. The B200 is a different architecture entirely: dual-die Blackwell design, native FP4 precision, and roughly 2.3x the FP8 throughput of H200. That architectural gap is why the price premium exists, and it's also why a flat hourly comparison misses the point.
The B200 premium pays for itself in specific, identifiable situations. Outside of these, H200 or even H100 usually wins on cost.
Right for
Wrong for
For a 405B parameter model at FP8 precision, the VRAM difference alone changes the math: H200 needs 4 GPUs to hold the model, while B200's 192 GB lets the same model run on 3. That's not just a cost saving on GPU-hours. It removes an entire node's worth of NVLink coordination overhead.
Fine-tuning is the clearest compute-bound case. B200 completes a 70B LoRA fine-tune roughly 2.2x faster than H200, based on MLPerf Training v4.1 benchmarks. On a job that takes 800 GPU-hours on H200 at $4/hr ($3,200 total), the same job finishes on B200 in about 364 hours, and even at the higher per-hour rate, total cost comes out around $2,000. The B200 wins on total cost despite costing more per hour.
B200 cost per hour is the wrong number to anchor on. What matters is total workload cost: hourly rate multiplied by the hours the job actually takes to finish. A GPU that costs 50% more per hour but finishes in half the time is cheaper, not more expensive.
Rate x Hours
Total workload cost, not hourly price, decides which GPU is actually cheaper for a given job.
packet.ai ROI framework
Use this sequence to model total cost for your own workload:
Estimate GPU-hours on H200
Use your own benchmarks or provider quotes for the model size and batch settings you plan to run.
Apply the B200 speedup factor
For training, expect roughly 2x. For FP4-optimized inference on 70B+ models, expect 3-4x. Use conservative estimates until you've benchmarked your own model.
Multiply each by its hourly rate
H200 hours x H200 rate versus B200 hours x B200 rate. Compare the two totals directly.
Factor in GPU count for large models
If VRAM headroom lets B200 run the job on fewer GPUs, multiply savings across the whole cluster, not just per-card.
On inference workloads the gap is larger than training. SemiAnalysis InferenceX benchmarks put B200 at roughly 3x lower cost-per-token than H200 on large models running FP4, and cost-per-million-tokens on B200 can land near $0.17 versus $0.50 on H200 for comparable Llama-class models. That gap only shows up once your serving stack actually runs FP4. TensorRT-LLM 0.16+ and recent vLLM builds support it, but older stacks won't see the benefit without a quantization pass first.
Note
The B200 speedup numbers assume FP4 quantization or NVLink 5 multi-GPU scaling. If your team hasn't set up an FP4 quantization pipeline yet, the realistic near-term gain is closer to the FP8-to-FP8 compute ratio: meaningful, but well short of the 3-4x figures vendors advertise for optimized deployments.
Run through these questions before choosing between B200 GPU rental and H200 for your next deployment:
If you answered yes to the VRAM or FP4 questions, run the total-cost math from the previous section before committing. If you answered no to most of these, H200 or H100 will almost always come out cheaper for the same job.
On-demand pricing is only one billing model, and it's usually the most expensive one. Reserved and dedicated contracts change which GPU wins the ROI comparison.
packet.ai B200 pricing by billing model (per GPU per hour)
On packet.ai, Dynamic B200 pricing runs $3.75/GPU-hr on shared, scheduler-isolated infrastructure with no commitment. Dedicated B200 access, a single-tenant card reserved exclusively for one account with a 99.99% SLA, runs $5.90/GPU-hr. Unlike some competitors, the dedicated tier here costs more than the shared tier since it trades a lower rate for guaranteed isolation rather than a term commitment. Across the broader market, reserved 36-month B200 contracts have dropped to roughly $2.25/GPU/hr, and spot pricing starts near $2.12/hr for checkpoint-friendly batch jobs, though those figures reflect market-wide reserved and spot listings rather than packet.ai's own tiers.
The tradeoff is flexibility. Spot capacity from other providers can be reclaimed mid-job, which is fine for checkpointed batch training but risky for anything serving live traffic. Long-term reserved contracts elsewhere lock in a lower rate but commit you to a term, usually 6 to 36 months, before you've validated the workload at scale. packet.ai's Dynamic tier carries no commitment risk and its Dedicated tier adds SLA-backed isolation without a term lock-in, which matters most during early evaluation or unpredictable traffic.
Virtualized B200 instances work for most training and inference jobs, but bare metal access matters in specific cases: when you need full control over NVLink topology for multi-node scaling, when your serving stack requires kernel-level tuning that a hypervisor layer interferes with, or when consistent low-latency performance is a hard requirement rather than a nice-to-have.
Watch out
B200 hardware remains supply-constrained industry-wide, with an estimated multi-million-unit order backlog as of mid-2026. Bare metal capacity is the tightest tier. Confirm availability before committing a project timeline to a specific start date.
Bare metal also matters for teams planning mixed H200 and B200 clusters. Mixed-hardware clusters incur a small throughput penalty, typically 3-5%, as schedulers balance uneven GPU performance across nodes. Bare metal access lets you isolate B200 nodes for compute-bound jobs while keeping H200 nodes on memory-bound inference, avoiding that penalty entirely, once H200 capacity is available.
packet.ai prices B200 GPU rental from $3.75/hr Dynamic ($5.90/hr Dedicated), well below wholesale on-demand rates and the $10.00-$14.24/hr hyperscaler quotes from AWS, Azure, and GCP for the same silicon. H200 is coming soon to packet.ai. It is not available for rental yet, but is open for waitlist signup with a listed Dedicated rate of $2.49/hr once it launches.
Note
H200 on packet.ai is listed as coming soon. Join the waitlist for early access. Until it launches, teams needing H200 today should use another provider or plan around packet.ai's currently available B200 and H100 capacity.
The difference comes down to what you're paying for. Hyperscaler B200 pricing bundles in margin, egress fees, and infrastructure overhead unrelated to the GPU itself. Marketplace aggregators list a wide spread, from spot listings as low as $2.69/hr to premium dedicated tiers over $16/hr, because they're pooling inventory from dozens of providers with different reliability guarantees. packet.ai prices Dynamic and Dedicated B200 capacity directly, without the quota friction or multi-week waitlists common on hyperscale platforms, so the rate you see is close to the rate you pay at scale.
Teams planning to mix memory-bound and compute-bound capacity can pair H200 on packet.ai for inference serving with B200 for training or FP4-optimized batch jobs once H200 launches from waitlist. In the meantime, B200 and H100 are both live today. Browse current cluster availability to see live capacity.
Last reviewed: July 15, 2026. Compare live B200 and H200 rates and browse available clusters on packet.ai.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →
