B300 Price Factors: What Moves a Reserved Capacity Quote

Picture a procurement lead at a mid-sized AI lab who has two reserved B300 quotes on her desk. One looks simpler and the other has twice as many line items. Her CFO wants to know why they differ, and her infrastructure team wants to know which one will actually deliver a working cluster on a predictable schedule. Neither question can be answered by comparing headline numbers. The answer sits in the assumptions each vendor made about configuration, fabric, tenancy, acceptance and support.

This guide is for procurement and infrastructure buyers who are evaluating a reserved B300 capacity quote and need to understand what moves the B300 price. It contains no rates or dollar figures on purpose. Market prices shift, and a number without its assumptions is useless in a negotiation. What lasts is knowing which variables drive cost, which ones you can influence and which parts of a quote are firm. If you are working through an H100 deal at the same time, the parallel breakdown in our guide to H100 reserved capacity pricing factors uses the same framework.

1. Start with the configuration baseline

Every reserved B300 quote rests on a configuration baseline. If two vendors quote different baselines, you are comparing two different products. Before you look at any commercial term, write down exactly what hardware each quote assumes.

The common unit is the B300 SXM GPU with 288 GB of memory, packaged eight per node in an HGX B300 server. At cluster scale, Pacific works from a Pod 32 reference: 32 HGX B300 nodes with a dedicated, configurable fabric. A reference design like this gives both sides a shared starting point. Each change you request from it, such as more nodes, a different storage tier or a modified network layout, becomes a visible line item that you can discuss.

Check each quote against this list:

• GPU model and memory per GPU, stated explicitly as B300 SXM 288 GB
• GPUs per node and total node count
• Whether the quote is built from a named reference pod or a custom design
• CPU, system memory and local NVMe per node
• Shared storage capacity and tier, if included
• Any spare nodes or hot spares held in the cluster

A quote that leaves out storage or spares will look leaner. It will also leave you buying those items later, often on a tighter schedule and with less negotiating room.

2. Single-tenant bare metal versus shared capacity

Tenancy model is one of the biggest structural drivers of a reserved B300 quote. A single-tenant bare-metal cluster means the nodes, the fabric and the operating environment are dedicated to you. Nobody else's jobs share your interconnect, and nothing sits between your workload and the hardware.

That isolation has a cost basis. The provider cannot resell idle cycles on your nodes, and the fabric is built and tuned for one customer. In return, you get predictable performance, a cleaner security story for sensitive data and full control over drivers, schedulers and images.

Shared or virtualized capacity spreads infrastructure cost across tenants, which can change the economics. It also introduces variables you need to price in yourself: noisy neighbors on shared links, limits on low-level access and less certainty about which physical hardware you land on. When comparing quotes, confirm in writing which model each one assumes. Pacific's reserved B300 offer is a single-tenant bare-metal cluster, so the comparison should be made against other dedicated options.

3. Fabric and interconnect choices

For multi-node training, the fabric matters as much as the GPUs. It is also where quotes most often diverge without saying so. A dedicated, configurable fabric lets you shape the network around your workload, and every choice in that design carries cost.

Questions that move the fabric line of a quote:

• Compute fabric technology and per-node bandwidth for GPU-to-GPU traffic across nodes
• Topology, including whether the design is non-blocking at the scale you plan to run
• Whether storage and management traffic run on separate networks or share the compute fabric
• Optics, cabling and switch redundancy included in the scope
• Room to expand the fabric if you add nodes later in the term
• Who owns fabric configuration changes after handover

A configurable fabric gives you leverage. If your workload is mostly inference or smaller fine-tuning jobs, you may not need the same fabric design as a team training large models across the full pod. Describe your actual job profile early in the conversation so the design, and the quote, reflect what you will run.

4. Acceptance evidence: factory, site and integration

Acceptance testing is easy to overlook when reading a quote, and it has a real effect on both cost and risk. Producing structured, witnessed evidence at several stages takes time, equipment and people. A quote that includes it will carry that work. A quote that skips it moves the risk onto you.

Pacific's reserved B300 offer includes witnessed acceptance evidence at three points:

• Factory acceptance, where hardware is tested before it ships
• Site acceptance, where the installed system is verified in place
• Integration acceptance, where the cluster is checked as a working whole, including fabric and storage

When you compare quotes, ask each vendor what tests run at each stage, who can witness them, what the pass criteria are and what happens to the schedule if a test fails. Our post on acceptance criteria for reserved B300 clusters covers how to write criteria your team can actually verify.

5. Site readiness and the delivery window

Where the cluster runs affects cost as much as what it is made of. Power availability, cooling design, placement and site interfaces all feed into the quote and into the delivery window.

Dense B300 deployments depend on verified power and a workable placement plan. Pacific develops factory-produced AI infrastructure for unused or stalled power, and its module reference includes direct-to-chip liquid cooling, networking, storage and bare-metal compute. Modules can land on a prepared pad, at a colocation location or at a customer site, followed by site commissioning and integration. Each path carries different preparation work, and that work shows up in the quote.

The delivery window is a targeted deployment timeline confirmed site by site. It is never a guaranteed date. Treat any vendor that offers a firm date before site facts are verified with caution. Ask what has been confirmed about the power feed, the connection point and site access, and which items still depend on utility, landlord or engineering sign-off. A power letter, an access route and a drawing each answer a different question, and a strong quote tells you which of those it already has.

6. Support scope and the acceptance boundary

Two quotes for the same hardware can differ a lot because of what happens after handover. Support scope defines who does what once the cluster is running, and the acceptance boundary defines the exact point where responsibility changes hands.

Pin down these items before you compare commercial terms:

• What the provider supports: hardware, fabric, firmware, base OS, drivers
• What you own: job scheduling, containers, application stack, data
• Response and repair expectations for failed nodes, GPUs or links
• How spare parts and replacement nodes are handled during the term
• The exact acceptance event that starts the reserved term
• How partial acceptance works if some nodes pass and others need rework

The acceptance boundary is especially important for reserved capacity, because it decides when your commitment begins. A clearly defined boundary protects you from paying for capacity that has been delivered but does not yet work as a cluster.

7. Confirmed versus indicative: reading the quote

The most useful skill when reading any GPU quote is separating what has been confirmed from what is indicative. Indicative figures help with budgeting. Confirmed terms are what you can sign against.

On Pacific's GPU compute catalog of dedicated bare-metal nodes, B300 SXM is listed by request, with pricing by request. Where preview rates appear for other GPU types, they are indicative until Pacific confirms the allocation, and they are not a live inventory feed. Review revalidates capacity, the current quote and terms before you commit. Placing a paid order does not mean the resource is ready. Availability is confirmed when an order is accepted.

For a reserved B300 cluster, the path is more structured. Configuration and the delivery window are confirmed through a project review and a written offer. Commercial terms are discussed privately on a qualified call. That sequence keeps hardware scope, site facts and money as separate threads, so each can be checked before you commit.

When reading any written offer, sort each line into one of three buckets:

• Confirmed: configuration, node count, fabric design and acceptance stages written into the offer
• Targeted: delivery window, dependent on site-by-site confirmation
• Open: anything still waiting on utility, landlord, engineering or commercial sign-off

If you cannot tell which bucket a line belongs in, ask. A vendor who can answer that question clearly is usually one who can deliver on it. For the broader picture of how cloud GPU pricing works across rental models and providers, see Pacific's guide to general GPU cloud pricing cost drivers. This post stays focused on what shapes a reserved B300 quote specifically.

8. Build your comparison sheet

Once you understand the drivers, put every quote into the same format. A single comparison sheet makes gaps obvious and gives your finance team something they can audit.

• Configuration baseline and any deviations from a reference pod
• Tenancy model, with single-tenant bare metal stated explicitly or excluded
• Fabric design, bandwidth and expansion path
• Acceptance stages, witnesses and pass criteria
• Site path: prepared pad, colocation or customer site
• Delivery window status: targeted, with the site facts it depends on
• Support scope and the event that starts the term
• Which items are confirmed, targeted or still open

With that sheet filled in, the B300 price conversation becomes far more productive. You are negotiating over defined scope, and both sides can see which levers change the number.

If you are scoping a reserved B300 cluster and want the configuration, fabric and acceptance plan worked out against your workload, Pacific's reserved B300 offer sets out the Pod 32 reference and the project review process. To start a project review with Pacific Intelligent Technologies, Inc., book a qualified call to discuss your reserved B300 requirements.

FAQ

Why doesn't Pacific publish a B300 price?

Reserved B300 clusters depend on configuration, fabric design, site path, acceptance scope and support terms, and each of those varies by project. Any published figure would be indicative at best. Pacific confirms configuration and the delivery window through a project review and written offer, and discusses commercial terms privately on a qualified call.

What is the Pod 32 reference?

Pod 32 is Pacific's reference configuration for reserved B300 capacity: 32 HGX B300 nodes, each with eight B300 SXM GPUs at 288 GB, connected by a dedicated, configurable fabric. It gives buyers a defined baseline, so any change from it can be scoped and quoted clearly.

Is the delivery date in a B300 written offer guaranteed?

No. The deployment timeline is a target confirmed site by site, because it depends on verified power, placement and site interfaces. The written offer confirms the configuration and delivery window, and your comparison sheet should record which site facts that window depends on.