How does GB200 cloud pricing compare to H200, and at what utilization does the newer chip become cheaper per token?

Teams weighing GB200 against H200 see an hourly rate roughly three times higher and often stop there, concluding H200 is the economical choice. For low or uneven workloads that is usually right, but it misses where the newer chip flips. GB200 costs about three times H200 per GPU-hour, so it only becomes cheaper per token once utilization is high enough that its throughput advantage outpaces the price gap. This guide compares the two on price and fit, then shows where the utilization break-even sits.

GB200 and H200 sit at different price points and jobs

The two are not the same product at different rates; they target different workload shapes. H200 is a Hopper-generation GPU with large high-bandwidth memory, well suited to long-context and high-memory inference on a single powerful card. GB200 NVL72 is a liquid-cooled, rack-scale Blackwell system that links 72 GPUs over NVLink into one tightly coupled domain for large-scale training and high-density serving.

In our current published listings, H200 is listed at from $2.60 per GPU-hour and GB200 NVL72 at from $8.00 per GPU-hour, so GB200 starts at roughly a 3x hourly premium. That gap is the number GB200's throughput has to overcome before it can be cheaper per token. Read on the rate card alone, H200 always looks cheaper; read on delivered cost per token, the answer depends on how busy the hardware stays.

Why utilization is the variable that decides

Cost per million tokens is all-in hourly cost divided by the tokens the GPU actually delivers per hour. Utilization sets the denominator, so it moves the per-token cost more than the sticker does. At low utilization a GB200 rack bills its full premium while much of its capacity sits idle, so the per-token cost is high and H200 wins easily. As utilization climbs and the rack runs at high batch across its interconnected GPUs, it processes far more tokens per hour, and the fixed premium spreads across enough work to fall below H200 per token.

The break-even is the utilization where GB200's throughput advantage over H200 grows large enough to offset its roughly 3x price premium; below that point H200 is cheaper per token, above it GB200 is. Because the premium here is larger than the roughly 2x gap between B200 and H100, GB200 needs a higher, more sustained utilization before it pays off. A rack that cannot stay busy will lose the per-token comparison even though its peak throughput is higher.

Reading the break-even for your workload

Use utilization bands to reason about which chip is cheaper per token, then confirm with measured throughput on your model.

Sustained utilization	Cheaper per token	Why
Low, bursty, or unproven	H200 (from $2.60/hr)	GB200 premium bills against idle rack capacity
Moderate, single-card fit	H200	Workload does not fill the interconnected rack
High, large-batch, rack-scale	GB200 (from $8.00/hr)	Throughput spreads the 3x premium below H200 per token
Near-continuous frontier serving	GB200	Rack stays busy enough to clear break-even comfortably

One boundary to keep clear: comparing GB200 and H200 on hourly rate answers a different question than comparing them per token. H200 is the cheaper hourly card and the cheaper per-token card at low or moderate utilization; GB200 only becomes cheaper per token once sustained, high-batch utilization pushes its throughput advantage past the roughly 3x price gap. If you cannot forecast utilization above that threshold, H200 is the safer economic choice.

Getting a per-token comparison you can verify on GMI

Once you know utilization drives the decision, the practical step is measuring throughput on real capacity at your real batch size. We are an AI-native inference cloud that publishes dedicated NVIDIA GPU list pricing and provides both H200 and GB200 NVL72 capacity, so you can benchmark the same workload on each and compare delivered cost per token directly.

Verify the H200 from $2.60 and GB200 NVL72 from $8.00 per GPU-hour starting rates on our GPU infrastructure (https://www.gmicloud.ai/en/gpus) and pricing (https://www.gmicloud.ai/en/pricing), then run your model at production batch to get tokens per second on each. Because the break-even depends on staying busy, our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved, single-tenant capacity with warm, weights-preloaded serving that keeps utilization high, so the per-token math reflects delivered cost rather than idle-inflated hours. Start in our console (https://console.gmicloud.ai) to measure both before committing to either.

Let utilization, not the rate card, pick the chip

If you compare GB200 and H200 on the hourly rate, H200 wins every time and you will pass on GB200 even where it would be cheaper per token. Run the real calculation: forecast your sustained utilization, measure tokens per hour on both at production batch, and compare delivered cost per million tokens. H200 is cheaper per token at low or moderate utilization, and GB200 becomes cheaper only once high, rack-scale utilization pushes its throughput past the roughly 3x price premium. Size utilization first, then let the per-token number choose the chip.