Blog ▸ B200 Cluster Price in 2026: What Buyers Actually Pay by Term, Size and Region
GPU Infrastructure
The second most requested GPU on the network and the one priced above its successor. The B200 band by term, size and region, and the three cases where it is still the right commit.
B200 Cluster Price in 2026: What Buyers Actually Pay by Term, Size and Region
GPUaaS.com Team
GPU Infrastructure
October 14, 2026
No items found.
B200 cluster price on committed terms is $7.00 to $8.00 per GPU-hour in Q4 2026 across the GPUaaS.com network, with the same band holding across regions this month and term length, cluster size and interconnect deciding where inside it a quote lands.
B200 is the second most requested GPU on the network and the one whose price has been most argued about this year, because it now sits above B300 on committed terms. This post gives the band by term, size and region, puts it next to on-demand and spot, explains why it is holding where it is, and sets out when a buyer should still choose it.
Key takeaways
Committed B200 clusters quote at $7.00 to $8.00 per GPU-hour this month. Twelve-month terms land at the bottom, three-month terms at the top
Hyperscaler on-demand medians ran near $14.24 in late September; specialist clouds about $7.88; marketplace spot about $5.89 where listed
Regional variation is minimal this month. North America, Europe and APAC quote inside the same band; lead time is what differs
B300 committed is $5.00 to $6.50, below B200. The B200 band is holding because early fleets were bought at higher unit cost and demand from older specs is sticky
Contiguous InfiniBand capacity at 64 GPUs and above prices toward the top; non-contiguous Ethernet for inference prices toward the bottom
B200 still fits buyers with existing B200 commitments, FP64 requirements, or air-cooled facilities. For most new workloads B300 is the better commit
◆ B200 COMMITTED BAND BY TERM AND SIZE, Q4 2026
Configuration
3-month term
6-month term
12-month term
8 to 32 GPUs, Ethernet, non-contiguous
Toward $8.00
Mid-band
Toward $7.00
8 to 32 GPUs, InfiniBand, contiguous
Top of band
Upper mid-band
Mid-band
64 to 128 GPUs, InfiniBand, contiguous
Top of band
Upper mid-band
Mid-band
256+ GPUs, InfiniBand, contiguous
Rare at this term
Upper mid-band
Mid-band, negotiated
Committed-term band of $7.00 to $8.00 per GPU-hour across the GPUaaS.com network, October 2026, with positions inside the band shown as direction only. Every quote is per request. Contact for commitment terms. *Prices may vary depending on commitment terms.
◆ WHAT MOVES THE QUOTE
Term first, then fabric, then size, and region barely at all
Term length moves a B200 quote more than any other variable. A twelve-month commitment pays for 8,760 hours with no re-fill risk for the provider, and it lands at the bottom of the band. A three-month commitment carries that risk sooner and lands at the top. Six months sits between. This is the same logic as every other part in the October benchmark, and it is why the band is a range rather than a number.
Fabric comes next. Contiguous InfiniBand capacity, where every node shares one fabric and a training job can all-reduce across the whole allocation, prices above the same GPU count scattered across nodes on Ethernet. Inference buyers who do not need the fabric should say so, because it widens the provider pool and lowers the quote. The InfiniBand vs Ethernet post covers when the fabric matters.
Size moves the quote less than buyers expect. Larger allocations get modest per-GPU improvement on long terms, but at 256 GPUs and above the constraint is finding a single provider with that much contiguous B200, and the quote reflects scarcity rather than volume. Region moved almost nothing this month: the benchmark found similar bands across North America, Europe and APAC, with the difference showing up in lead time.
◆ B200 PER GPU-HOUR, FOUR WAYS TO BUY IT, OCTOBER 2026
Channel
Price
What it is
Hyperscaler on-demand
About $14.24 median
Start and stop any time, inside a platform
Neocloud on-demand
About $7.88
Same right, less platform
Marketplace spot
About $5.89 where listed
Interruptible, thin supply
GPUaaS.com committed
$7.00 - $8.00
Dedicated cluster, fixed term, paid every hour
Hyperscaler and neocloud medians from published rate cards, late September 2026, as cited in the GPUaaS.com October benchmark. Spot from published marketplace listings compiled in June 2026. Different products; the comparison is directional.
$14.24
the hyperscaler on-demand median for a B200 in late September, against $7.00 to $8.00 committed. The same GPU, at roughly half, for a buyer who can name a term
GPUaaS.com October 2026 benchmark
◆ WHY THE BAND IS WHERE IT IS
Above B300, and holding
The B200 band sits $1.50 to $2.00 above B300 this month, which is the reverse of what a generation gap usually looks like. The B300 vs B200 post covers the reasons in detail; in short, B200 fleets were bought earlier at higher unit cost, their owners are pricing to recover it, and a steady stream of requests written against B200 specs from the first half of the year keeps the band from falling.
That will not hold indefinitely. As B300 allocation keeps landing and buyers revisit older specs, B200 providers will have to choose between the band and the fill rate. For a buyer the practical implication is that a B200 quote at the bottom of the band on a twelve-month term is a reasonable deal today, and a B200 quote at the top of the band on a short term is paying a premium for a part that is no longer the best on the market.
The two-price market explains why the committed band and spot can both be right at once, and the H200 vs B200 comparison is the right read for a buyer deciding whether to skip B200 downward rather than upward.
◆ WHEN TO COMMIT TO B200
Three cases, and how to specify it
B200 is the right commit when a buyer already holds a B200 contract with months to run and the migration cost to B300 outweighs the band gap; when the workload needs FP64, which B300 has effectively dropped; or when the capacity is for owned-style deployment in an air-cooled facility where B200's 1,000W envelope sites more easily. For a new inference or training workload with none of those constraints, B300 at $5.00 to $6.50 with 288GB is the better commit.
To specify it: give the term first, since it moves the quote most. Give the GPU count in multiples of eight. Say whether the allocation must be contiguous on InfiniBand or whether Ethernet is acceptable. Name a region only if residency requires it, since the band does not move with region this month but the provider pool does. Quotes come back within a few hours from 20+ vetted partners with no buyer fees, and the contract runs between the provider, GPUaaS.com and the buyer.
For single B200 GPUs rather than an 8-GPU node, packet.ai handles self-serve access, and the cluster sizing post covers how many B200s a given job actually needs.
Get a B200 cluster quote inside this month's band.
Submit a spec and quotes come back within a few hours, sourced across 20+ vetted partners. No buyer fees. 8-GPU node minimum.
$7.00 to $8.00 per GPU-hour on committed terms through the GPUaaS.com network in October 2026, with twelve-month terms at the bottom of the band and three-month terms at the top. Hyperscaler on-demand medians ran near $14.24 and neocloud on-demand about $7.88 in late September.
B200 fleets were bought earlier at higher unit cost and their owners are pricing to recover it, while B300 allocation is arriving and being quoted rather than held. Requests still written against B200 specs from earlier in the year keep the band from falling. It is unlikely to hold as buyers revisit those specs.
Very little this month. The $7.00 to $8.00 band held across North America, Europe and APAC in the October benchmark. What differs by region is lead time and the size of the provider pool, with North America deepest and European residency-constrained requests slowest to fill.
Term length first: twelve months lands lowest. Then fabric: contiguous InfiniBand prices above non-contiguous Ethernet. Then size, modestly, with scarcity rather than volume setting the price above 256 GPUs. Region barely moves it this month.
B300 for most new workloads: more memory, more FP4 and a lower committed band this month. B200 when an existing B200 contract has months to run, when the workload needs FP64, or when an air-cooled facility cannot site B300 at density.
One 8-GPU HGX node through GPUaaS.com, on short-term or long-term commitments. Single B200 GPUs are handled by packet.ai on a self-serve basis. Capacity can be sourced in any region except China and Taiwan.
