B300 Node Price: What an 8x HGX B300 Node Actually Costs in Q4 2026

◆ WHAT AN 8-GPU B300 NODE ACTUALLY QUOTES AT

Published listings and reseller quotes, September to October 2026, excluding VAT. A standalone B300 module runs roughly $40,000 to $60,000.

◆ WHY THE SPREAD IS THAT WIDE

The GPUs are the smaller half of the bill

Eight B300 modules at roughly $53,000 each come to about $424,000. A node listed at $795,000 is therefore carrying around $370,000 of everything that is not the GPU: chassis, CPUs, NVLink fabric, NICs, power delivery, liquid cooling, integration and support.

At the $300,000 anchor the arithmetic implies either volume pricing on the modules or a thinner platform around them. Both exist, and they are not the same product even though the GPU count matches.

Comparing node quotes without comparing what surrounds the GPUs is the most common error in these models. Interconnect tier alone changes throughput on multi-node training enough to matter, which is covered in InfiniBand vs Ethernet.

◆ THE LINE THAT GETS LEFT OUT

14.5 kW, every hour it runs

An HGX B300 node draws roughly 14.5 kW. Hosted in colocation at 70% utilization, that works out to about $50,000 to $145,000 a year once energy, contracted power, cooling, support and remote hands are counted. Staff and software licences sit on top.

Contracted power runs $130 to $380 per kW per month depending on the facility, which alone accounts for $23,000 to $66,000 of that annual figure. Facilities built for 8 to 12 kW racks cannot take a Blackwell Ultra node without work.

Over three years the run cost adds $150,000 to $435,000 to the purchase price. On the cheapest node that is more than the hardware cost again.

◆ WHAT OWNING COSTS PER GPU-HOUR, THREE YEARS AT 70% UTILIZATION

Purchase price plus three years of run cost, divided by 8 GPUs across 26,280 hours at 70% utilization. Excludes financing, staff, software and residual value.

◆ AGAINST THE RENTAL BAND

Owning sits inside the range, not below it

Committed B300 cluster capacity is quoting at $5.00 to $6.50 per GPU-hour this month. Set that against the owned range of $3.06 to $8.05 and the comparison stops being obvious.

Owning wins clearly at the low end, where a node bought near $300,000 and run in a cheap facility lands around $3.06. It loses clearly at the high end, where a $750,000 node in an expensive facility reaches $8.05 and renting is cheaper outright. In the middle, which is where most real purchases sit, the two overlap.

And that owned figure assumes 70% utilization sustained for three years. At 40% the same node costs roughly 1.75 times more per useful GPU-hour, which is usually what decides it.

◆ THE DELIVERY GAP

Months before the first hour runs

Current-generation Blackwell nodes run 16 to 26 weeks from order to delivery, and one integrator listing in September quoted November delivery against a September order. Rented capacity provisions in minutes.

That gap does not appear in a cost-per-GPU-hour comparison, but it decides whether the hardware arrives in time to matter. A model trained six months from now on owned hardware competes against models trained next week on rented capacity, and the cost advantage has to be large enough to justify the wait.

◆ TERM LENGTH MOVES THE RENTAL RATE

The ladder is visible in published rate cards

One provider publishes B300 at $7.49 on demand, $6.49 on a twelve-month term, $5.99 at twenty-four months and $5.49 at thirty-six. Another lists $4.65 at thirty-six months and $4.10 at forty-eight.

That ladder is the clearest illustration of what a committed band actually prices. The provider is selling certainty, and the discount scales with how much of it the buyer hands over.

It also means a quoted rental rate without a term attached is not a number. Across 33 tracked providers B300 spans $3.49 to $7.49, and most of that spread is term rather than provider quality.

◆ MOST B300 IS NOT ON A RATE CARD

Contact sales is the default

Several of the largest specialist providers list HGX B300 and GB300 as contact sales rather than publishing a rate. That is not evasion, it reflects that most Blackwell Ultra capacity is still allocated through conversations rather than self-serve.

The practical consequence for a buyer is that published rates are a floor on information rather than a ceiling on price. A spec taken to several providers will usually produce numbers that no page displays.

◆ THE COSTS BUY-VS-RENT MODELS LEAVE OUT

Financing, staff and residual value move the number further than people expect

The $3.06 to $8.05 per GPU-hour range above excludes three things that most owners actually pay. Financing is the largest. A node bought on a loan or lease at anything above a token interest rate adds real cost on top of the purchase price, and a $600,000 node financed over three years at a typical equipment-loan rate can add $60,000 to $90,000 in interest alone, which moves the owned figure up by roughly $0.75 to $1.10 per GPU-hour at the utilisation assumed above.

Staff is the second. Someone has to rack the node, keep the firmware current, monitor it, and be the person who gets paged when a GPU throws an error at 2am. For a single node that cost is proportionally large; for an organisation already running infrastructure at scale it is closer to free, since the incremental node adds little to an existing team's workload. That is the main reason the real-world answer to buy-versus-rent often splits by how much infrastructure an organisation already operates, not by the raw arithmetic on one node.

Residual value cuts the other way and is usually ignored entirely. A three-year-old B300 node is not worthless at the end of the depreciation schedule; it still has resale or internal-redeployment value, and the resale value post covers what current-generation GPUs have actually fetched on the secondary market. Crediting even a conservative resale figure back against the purchase price narrows the owned range by several tens of cents per GPU-hour, which is often enough to tip a borderline case toward buying.

◆ A WORKED EXAMPLE

One node, three scenarios, three different answers

A lab buying a single $550,000 HGX B300 node for a sustained research programme, with an existing data centre team and 70% utilisation expected over three years, lands near $4.76 to $6.69 per GPU-hour once run cost is added, before financing or staff overhead, which are close to zero for them since the team already exists. Against a committed rental band of $5.00 to $6.50, owning is a reasonable bet and gets better the longer the programme runs past three years, since the node keeps working after the model fully depreciates.

A startup buying the same node with no existing infrastructure team, financing the purchase, and running it at a more realistic 40% utilisation while the product finds its workload, pays the same $4.76 to $6.69 run-cost range but stretches it across far fewer useful hours, pushing the effective figure past $8 to $11 per GPU-hour once financing and a part-time hire are added. For that buyer, renting at $5.00 to $6.50 committed is simply cheaper, and it avoids the 16 to 26 week wait for delivery on top.

A sovereign or regulated buyer who must keep the hardware in a specific jurisdiction for compliance reasons is not really choosing on cost at all; residency requirements can rule out rental if no provider holds documented capacity in that jurisdiction, in which case owning is the only option regardless of what the arithmetic says. The India pricing post and the regional demand report cover how residency narrows the rental option specifically.

The question worth answering before any of this is utilization. A node sustained above 70% for three years makes ownership defensible at the right purchase price. Below that, the fixed cost spreads across fewer useful hours and renting wins on arithmetic rather than preference.

Current committed bands across every tier are published in the October price benchmark.

◆ FAQ

Frequently asked questions