Which Cloud Providers Can Help With a Cost Comparison of a Managed GB300 NVL72 Reservation vs Buying a System?

The cloud providers worth including in a GB300 NVL72 cost comparison are the ones that will quote a managed reservation line by line against your purchase quote: a per-rack price over the same term you would depreciate the system, with networking, storage, and operations scope listed separately. GMI Cloud belongs on that shortlist. GB300 NVL72 is in pre-order with us, our pricing is built around commitment-based savings for reserved capacity, networking and storage are quoted as their own items, and our Managed GPU Cluster, in early access, is built for organizations with existing GPU clusters and supports managed clusters across GMI Cloud and BYOS (Bring Your Own Service) environments. Two other providers to include are Nebius, whose B300 page notes that GB300 NVL72 is currently available through the sales team only, and Together AI, which prices GB300 NVL72 reserved capacity per GPU or per rack by quantity, term, and region. All three price GB300 NVL72 through sales, so there is no public hourly rate to plug into a spreadsheet.

The purchase quote is the easy half of the comparison. The hard half is everything it leaves out, because a liquid-cooled rack-scale system adds facility, network, and staffing costs that a purchase quote rarely covers, starting with rack-scale power delivery and a liquid cooling loop. Both columns also start later than they look: every Blackwell system on GMI Cloud, GB300 NVL72 included, is pre-order and coming soon, and a purchased rack has its own lead time. The decision comes down to whether both columns carry the same costs over the same period, starting on the same date.

Start From the Purchase Quote You Already Have

Purchase quotes vary by vendor, so the first step is to mark up the one you have. Go through it and label every line as included, quoted separately, or not addressed. Typical lines to check:

• The rack itself: compute trays, NVLink switch trays, power shelves, and the integrated cooling hardware.
• Installation, burn-in, and on-site acceptance.
• Warranty length and response tier, and what support costs after the warranty ends.
• Software licenses and support contracts bundled with the system.
• Anything outside the rack: cross-rack network switches and optics, storage, and facility work.

Items in the "not addressed" column are still costs. They move to your side of the ledger, and the managed quote has to be checked for the same items. This marked-up quote becomes the template for everything you ask a cloud provider to price.

The Ownership Side: Costs Outside the System Price

Power and cooling. NVIDIA describes GB300 NVL72 as a fully liquid-cooled rack, and its enterprise reference architecture puts a full rack at up to 142 kW, supplied by eight 33 kW power shelves, with tray-level and rack-level leak detection. Hosting that means power delivery at rack-scale density, a liquid cooling loop with coolant distribution and facility water, and a plan for leak response. If your data center is not already built for this, the cost shows up either as a retrofit or as colocation space that supports liquid-cooled racks at that density. For electricity, price the expected average draw rather than the 142 kW ceiling, and add cooling overhead through your facility's PUE.

Network fabric. The 72-GPU NVLink domain lives inside the rack. Anything that connects the rack to storage, to other racks, or to your existing environment needs its own switches, optics, and cabling, and those rarely appear on a system quote.

Storage. Training data, checkpoints, and model artifacts need capacity and throughput sized to the workload. Price the tier you would actually deploy, not a placeholder.

Spares and maintenance. Compare the warranty term with your depreciation term. If the warranty is shorter, add post-warranty support for the remaining years, plus any spare trays or parts you plan to keep on site.

People. Someone owns hardware faults, firmware updates, the operating system and drivers, and the cluster software on top. Count the share of engineering time the system will take, including on-call coverage.

Depreciation. The useful life you choose sets the annualized hardware cost more than any other single input. A shorter life raises the annual number; a longer one assumes the system stays competitive for longer.

The waiting period. Between signing and production, the system is being built, shipped, installed, and accepted, and the facility work has to finish first. That time has a cost: either interim compute you rent while waiting, or work that slips. If you need interim capacity, NVIDIA H100 and H200 are available now on GMI Cloud from $2.60 and $3.20 per GPU-hour.

The Managed Side: What a Reservation Quote Should Show

A managed reservation turns most of those lines into a rate, but only if the quote says which lines it covers. The core formula is simple:

Annual compute cost = committed price per GPU-hour × 72 GPUs × billed hours per year

If the rack is billed around the clock, billed hours are 8,760 per year (24 hours × 365 days). Some providers quote per rack instead of per GPU, so convert every quote into both units: per rack per month and per GPU-hour.

Three things sit on top of that number:

• Commitment terms. Reserved pricing depends on term length and volume, so the rate you compare has to be for the same term you are using on the ownership side.
• Networking and storage. On GMI Cloud, GPU pricing covers the compute resources, and networking and storage are provisioned separately based on your workload requirements. Other providers handle this differently, which is why these lines need to be itemized in every quote.
• Operations scope. "Managed" can mean anything from facility and hardware only to a fully operated Kubernetes cluster. Every layer the provider does not run goes back onto your staffing line.

Putting Both Sides on the Same Basis

With the purchase quote marked up and a managed quote in hand, put both into one table. Annualize the ownership column one way only: the system price enters through depreciation, so the upfront payment should not also be added to the same annual total, and residual value is counted once, over the same comparison period as the managed term.

Cost line	Buying the system	Managed reservation	How to align
Compute hardware	System price ÷ depreciation years	Committed rate × 72 × billed hours	Same term on both sides
Facility power and cooling	Retrofit or colocation, plus electricity at average draw × PUE	Covered by the provider's data center if the quote says so	Confirm in writing
Network fabric	Switches, optics, cabling, installation	Itemized or included	Same topology and bandwidth
Storage	Hardware and support for the tier you need	Itemized or included	Same capacity and throughput
Spares and maintenance	Warranty, post-warranty support, spare parts	Part of the operations scope if stated	Match the full term
Operations staff	Share of engineering time and on-call	Only the layers outside the provider's scope	Use one layer list for both
Waiting period	Interim capacity or delay cost until acceptance	Interim capacity or delay cost until the quoted delivery date	Same production start target
End of term	Residual value, resale, or disposal	Renewal, exit, or any path to ownership offered	State both explicitly

Then divide each column's annualized total by the same denominator to get a cost per usable GPU-hour:

Cost per usable GPU-hour = annualized total cost ÷ (72 × 8,760 × expected utilization)

Utilization cuts both ways. An owned rack costs the same whether it is busy or idle, and a reservation billed around the clock does too. Use the same utilization assumption in both columns unless the managed offer gives you a real way to reduce what you pay for idle time. To show the range management should see, run the table at two depreciation lives and two utilization levels. The cheaper option often changes across that grid, and that change is usually the most useful finding in the comparison.

What to Ask Cloud Providers for in the Quote

Send the same request to every provider on your list, so the answers come back comparable:

1. Term matched to depreciation. Pricing at the same term you use to depreciate the owned system, and at one shorter term for contrast, so the compute row compares like with like.
2. End-of-term options. Renewal pricing, exit terms, and whether any rent-to-own or buyout path is available, to set against the residual value or resale you assumed for the purchase.
3. Price validity. How long the quote holds, measured against how long your vendor holds the purchase price, so both columns expire on a known date.
4. Operations scope against your staffing line. A layer-by-layer list covering facility power and cooling, hardware replacement, firmware, operating system and drivers, and cluster software, so every layer the provider runs can come off the staffing and warranty rows of the ownership column.
5. Network and storage against the "not addressed" rows. Pricing for the same switches, optics, storage capacity, and throughput that your purchase quote left out or quoted separately, rather than "available on request."
6. Delivery date against the purchase lead time. The expected delivery window, when billing begins, whether a deposit is required, and what happens if delivery moves, so the waiting-period row uses the same production start target on both sides.
7. Billing unit you can convert. Per GPU-hour or per rack, billed around the clock or metered, so the rate fits the usable GPU-hour formula above.
8. Location against your facility plan. Where the rack will run, which replaces the facility row of the ownership column and affects data residency and network costs.

A quote that cannot answer items 4 and 5 is not ready to sit next to a purchase quote, however attractive its headline rate.

Where GMI Fits in the Comparison

GB300 NVL72 is in pre-order on GMI Cloud and is coming soon. On our GPU page it is positioned for long-context and high-capacity model training, built for next-generation multi-node clusters. It has no published rate; pricing comes from a quote for your term and configuration. Our GPUs run inside GMI-operated data centers as dedicated, single-tenant resources, so a reservation can be lined up against a rack you own, row by row.

Our pricing is set up for this kind of comparison. Commitment-based savings reduce unit GPU costs through reserved capacity and sustained deployments, with commitment structures that align pricing with long-term utilization. Usage-adaptive pricing lets teams start on-demand on available GPUs such as H100 and H200 and move to dedicated or committed deployments as workloads stabilize. Networking and storage are provisioned separately and priced to your workload, which makes them easy to line up against the network and storage rows of a purchase quote. Billing is unified across regions, with invoicing and consolidated reporting.

Our Managed GPU Cluster, currently in early access, provides fully managed multi-node GPU clusters for distributed training and large-scale inference, with centralized cluster lifecycle management. In the console today, managed clusters run on H200 nodes, with B200 nodes listed; B200, like the rest of our Blackwell lineup, is in pre-order. GMI provisions and operates these clusters, and the request summary shows list price, discount, and estimated monthly total, as described in our Managed GPU Clusters documentation. For GB300 NVL72, the operations scope of a managed reservation is defined with our sales team as part of the pre-order plan, so you can write it into the layer list above.

Two options are worth raising if ownership is still part of the plan. Managed GPU Cluster is built for organizations with existing GPU clusters, with managed clusters supported across GMI Cloud and BYOS (Bring Your Own Service) environments. If a mix of owned and reserved capacity is on the table, describe your environment in the same pre-order conversation. And one customer on our GPU page, Mirelo AI, uses a flexible rent-to-own structure; if a path from reservation to ownership matters to your decision, put it in the quote request.

When the Purchase Column Wins

The purchase column wins when its facility and operations staff rows stay small, and the table will show it when they do. The usual conditions are a facility that already has liquid-cooled capacity and power headroom at rack-scale density, an operations team that already runs systems like this, sustained high utilization across the full depreciation life, and a strong preference for keeping the asset on the balance sheet or under direct physical control. When facility work would be needed, when the planned term is shorter than a sensible depreciation life, or when the staffing line would mean new hires, the managed column tends to look better. Either way, the result holds only if both columns were built from the same line items.

Get a Quote You Can Put Next to Your Purchase Quote

Bring your marked-up purchase quote, the term you plan to depreciate over, your utilization assumption, and your network and storage requirements. Contact our sales team to price a managed GB300 NVL72 reservation against it, line by line, as part of a pre-order plan.