Are there reserved or spot pricing options for GB300 GPUs, and how much can I save versus on-demand rates?

Teams budgeting for next-generation Blackwell often go looking for a GB300 spot market or a reserved discount the way they would for an H100. The catch is that GB300 sits at a different point in the supply cycle, so the usual on-demand-versus-spot math does not apply yet. GB300 NVL72 is largely a pre-order product with no public on-demand rate, which means there is no on-demand baseline to discount from and no meaningful spot market, so the only real savings lever today is a reserved or committed capacity agreement. This guide shows why the three billing models behave differently for GB300 and what to negotiate before you plan a budget around it.

Why spot pricing barely exists for rack-scale Blackwell

Spot pricing works when a provider has surplus capacity it would rather rent cheaply than leave idle. GB300 is the opposite situation: it is a scarce, newly shipping, liquid-cooled NVL72 rack in high demand, so there is no surplus pool to auction. That is a structural reason, not a temporary one.

A single H100 can be spun up and torn down in seconds, which is what makes a spot market liquid. A GB300 NVL72 is an interconnected rack domain that needs significant power and cooling and gets provisioned as a whole allocation, so providers hold it under reservation rather than release it to a volatile spot tier. For rack-scale Blackwell like GB300, on-demand and reserved capacity are the realistic billing models, and true preemptible spot instances are rare to nonexistent. If a quote advertises GB300 "spot" pricing, confirm whether it is genuinely preemptible or just a marketing label on short-notice allocation.

GB300 is pre-order, so there is no on-demand rate to save against

The more important fact for budgeting is where GB300 is in the release cycle. In our current published listings, GB200 NVL72 is listed at from $8.00 per GPU-hour and marked Available Now, while GB300 NVL72 is shown as pre-order with no published per-GPU-hour rate. When a chip is pre-order, the on-demand price is not yet public, so a "how much can I save versus on-demand" question has no numeric answer until that baseline is set.

This changes the shape of the decision. For an available chip, you compare an on-demand rate against a reserved rate and compute a discount. For a pre-order chip like GB300, you are negotiating an entry position, and the reserved commitment is often what secures both the capacity and the rate. The saving is not a published percentage off a rate card; it is the unit cost you lock in by committing early, measured later against whatever on-demand price the market settles on.

Reserved, on-demand, and spot compared for GB300

Read each model against how it actually applies to a pre-order rack chip, not against how it works for a mature single-GPU SKU.

Pricing model	Availability for GB300 today	Typical savings lever
Reserved / committed	Primary path; pre-order allocation	Lower unit rate for sustained commitment
On-demand	Not yet published (pre-order)	No baseline set to discount from
Spot / preemptible	Rare to nonexistent at rack scale	No surplus pool to auction

The pattern holds across most clouds carrying rack-scale Blackwell: the reserved column is where the real money moves, on-demand is a future baseline, and spot is mostly absent. That is why a GB300 budget built on an expected spot discount is planning around a market that does not exist yet.

What GMI's pricing model means for GB300 savings

Once you accept that the lever is a commitment rather than a spot auction, the practical work is choosing the right commitment level. We are an AI-native GPU and inference cloud that lists dedicated NVIDIA GPU pricing and offers both on-demand and reserved capacity plans, which is the structure that fits a pre-order chip like GB300.

We offer commitment-based savings, where reserved capacity and sustained deployment lower the effective per-GPU cost, alongside usage-adaptive pricing that lets you start on demand on available chips and move to committed capacity as workloads stabilize. For GB300 specifically, treat the pre-order label as a reservation: ask for the expected delivery window, the committed unit rate, and the deposit and cancel terms in the same conversation, since those are what actually determine your cost. Verify the GB200 from $8.00 per GPU-hour Available Now listing and the GB300 pre-order status on our GPU infrastructure (https://www.gmicloud.ai/en/gpus) and pricing (https://www.gmicloud.ai/en/pricing), and re-check them at order time because Blackwell pricing and stock move quickly. When the workload is sustained production inference rather than a one-off training run, our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved GPU capacity with single-tenant isolation and warm, weights-preloaded serving, so you hold rack-scale capacity for real traffic instead of chasing spot stock that is not there. Start allocation and commit-pricing conversations in our console (https://console.gmicloud.ai) or contact our sales team.

Budget GB300 on a commitment, not on a spot discount

If you build a GB300 budget expecting a spot market or a published on-demand discount, you are planning around pricing structures that do not yet exist for this chip. The honest read is simpler: GB300 is pre-order, so there is no on-demand baseline and no liquid spot tier, and the saving comes from the reserved commitment you negotiate at order. Get the delivery window, the committed unit rate, and the cancel terms in writing, verify the live pricing page when the on-demand rate does publish, and compare your locked reserved cost against that number once it exists rather than against a discount you assumed today.