B300 vs H200 vs RTX PRO 6000: which GPU server pays back fastest?

The short version

On a three-year contract at today’s rates, the RTX PRO 6000 and the H200 pay back in a dead heat: about 22 months each. The B300 needs about 27. All three are our estimates, and the arithmetic is below.

Change the contract and the winner changes. The PRO 6000 wins when the customer prepays. The H200 wins when the market turns. The B300 wins nowhere at today’s rates.

The H200 wins on cushion. Its contract rate can fall 34% before payback slips past 36 months. The PRO 6000 can take 30%. The B300, 22%.

We see all three in our deal flow, and we underwrite them the same way: at the node. The 8-GPU server is what gets bought, racked, rented and financed.

Deal terms we assume

Every base-case number rests on one contract structure.

Exhibit 1. Deal terms behind every base-case number

Term	Assumption
Initial contract term	36 months, all three nodes
Customer prepay	None (0%). Straight monthly billing from day one
Billing	Take-or-pay: 100% of GPU-hours billed for the full 36 months
Rate	Fixed flat for the full term. No escalator, no step-down
Rate source vs. term	B300: 36-mo 8-GPU listing. H200: Silicon Data index (no stated term). PRO 6000: 12-mo 8-GPU listing, held for 36 months
After month 36	Not modeled. Payback is measured inside the initial term only

We leave prepay out of the base case because prepayment terms vary widely from deal to deal. But prepay and tenor move payback more than the chip does. Here’s how much.

Exhibit 2. Payback in months, by contract structure

Contract (tenor / prepay)	B300	H200	RTX PRO 6000
3 yr / none	27	22	22
3 yr / 15%	24	19	18
3 yr / 25%	22	15	14
5 yr / none	32	26	27
5 yr / 15%	25	18	17
5 yr / 25%	19	11	7

Payback on capital at risk (capex less prepay), in months, estimated. Prepay is a share of total contract value, paid at signing and credited back evenly against billing. Five-year rates are scaled from the three-year rate using the steps in our B300 tenor ladder (about 15% lower).

Source: GPU Capital estimates.

• 8x B300
• 8x H200
• 8x RTX PRO 6000

Prepay is a share of total contract value, paid at signing and credited back evenly against billing. Five-year rates are scaled from three-year rates using the steps in the B300 tenor ladder (about 15% lower).

Source: GPU Capital estimates from UpStation catalog prices (September 9, 2026), GetDeploying and Silicon Data rates (September 24 to 29, 2026) and one recent quote.

Show data

Prepayment and tenor move payback more than the chip does

Contract	8x B300	8x H200	8x RTX PRO 6000
3 years, no prepay	27	22	22
3 years, 15% prepay	24	19	18
3 years, 25% prepay	22	15	14
5 years, no prepay	32	26	27
5 years, 15% prepay	25	18	17
5 years, 25% prepay	19	11	7

Read down any column: the contract swings payback by 13 to 20 months. Read across any row: the chip swings it by 5 to 12. Without prepay, a longer term is slower, because the lower locked rate costs more than the extra years add. With prepay, a longer term is faster, because the deposit is sized on a bigger contract. On the PRO 6000, a 25% prepay on five years is about 83% of the node’s cost. At that point the customer is financing the box.

The three nodes

Same GPU count. Very different machines.

Exhibit 4. The three 8-GPU nodes

Specification	8× B300 (HGX)	8× H200 (HGX)	8× RTX PRO 6000 Server Ed.
Architecture	Blackwell Ultra	Hopper	Blackwell
Memory / GPU	288 GB HBM3e (about 270 GB usable)	141 GB HBM3e	96 GB GDDR7
Bandwidth / GPU	Up to 8 TB/s	4.8 TB/s	1.6 TB/s
GPU power	About 1,100 W air-cooled, up to about 1,400 W liquid-cooled	Up to 700 W	Up to 600 W
Interconnect	NVLink 5 (1.8 TB/s GPU-GPU)	NVLink 4 (900 GB/s)	PCIe Gen5, no NVLink
Cooling	Air or direct liquid cooling, by server model	Air	Air (passive cards)
Node nameplate power	13.6 kW	9.1 kW	6.8 kW
Node capex (base case)	$572,364	$319,422	$125,000
Capex per GPU (derived)	$71,546	$39,928	$15,625

B300 node power is the 13.6 kW nameplate of the Supermicro build priced here; NVIDIA lists about 14 kW for a DGX B300 system, which would leave B300 payback at about 27 months. Air-cooled B300 servers run the GPUs at about 1,100 W each; liquid-cooled servers run them at up to about 1,400 W.

Source: NVIDIA HGX platform, NVIDIA HGX reference architecture (H200/B300 memory and bandwidth), NVIDIA H200, NVIDIA RTX PRO 6000 Blackwell Server Edition datasheet (PDF). B300 power and cooling per NVIDIA DGX B300, Supermicro and Dell (air-cooled) and Thunder Compute (liquid-cooled). B300 and H200 capex and all nameplate power: UpStation published catalog (Supermicro builds, prices as of September 9, 2026). RTX PRO 6000 capex: a recent quote in our deal flow.

B300 and H200 are single-node list prices. Volume buys come in lower. Jarvislabs puts an H200 at $30K to $40K per GPU to buy (August 2026), so $319K sits at the top of that range.

The PRO 6000 is where list and market part ways. The same catalog lists a Tier 1 node at $191,888 and a Tier 2 build at $163,910. We were recently quoted $125K. We use the quote, and show what list does to payback further down.

The HGX figures leave out fabric (800G for B300, 400G for H200), storage and install, so real HGX payback runs longer. The PRO 6000 is a PCIe inference box. No fabric needed.

Rental rates: use contract, not on-demand

Headlines quote on-demand. Lenders get paid on contract. The gap is wide.

Exhibit 5. Rental rates by node, dollars per GPU-hour, late September 2026

$/GPU-hr (late September 2026)	B300	H200	RTX PRO 6000
On-demand median	$7.89	$4.48	$2.19
Reserved median (all terms)	$5.97	$3.79	$1.83
Long-tenor 8-GPU listings	$4.58 and $5.49 (36-mo)	$4.47 (36-mo, SXM)	$1.39 (12-mo)
Other data points	Reserved floor $4.25	Silicon Data index $3.10; 1× 36-mo from $2.57	1-mo 8× reserved $1.27
Base case used	$4.58	$3.10	$1.39
Base ÷ on-demand	0.58×	0.69×	0.63×

Sources, all GetDeploying unless noted, pulled September 24 to 29, 2026: B300 listings; B300 on-demand median; B300 reserved and PRO 6000 medians; H200 listings and medians; RTX PRO 6000 listings; Silicon Data H200 index.

For B300 we take the low end of published 36-month 8-GPU listings. Our own tenor ladder has a three-year lock at $4.80, so $4.58 is the low side of where contracts actually print.

For the PRO 6000, 12 months is the longest published 8-GPU tenor we found. A true three-year lock would likely price lower.

Direction matters too. The AIMultiple index shows its latest-generation group median going from $2.12 in October 2024 to $4.72 in September 2026. Rates are firm. That’s the risk: firm rates are when people underwrite the top of the cycle.

Operating costs

Exhibit 6. Operating cost inputs, per node

Input	Assumption	Basis
Hours / month	730	Calendar
Billed utilization	100%	Take-or-pay contract: billed whether the GPUs run or not
Colo (all-in, per node)	B300 $3,816; H200 $2,552; PRO 6000 $1,906 per month	UpStation power & colo: $225/kW-mo, 3-yr term, billed on provisioned kW (nameplate ÷ 0.8)
Ops & support reserve	5% of revenue	Estimate: RMA, remote hands, monitoring
Excluded	Fabric, storage, financing, taxes, operator fees, residual value	Kept out to isolate node-level payback

The payback math

Simple payback: node capex divided by monthly net cash. Every output below is an estimate built from the inputs above.

Exhibit 7. Payback per node, per month

Per node, per month	8× B300	8× H200	8× RTX PRO 6000
Rate ($/GPU-hr)	$4.58	$3.10	$1.39
Revenue (8 × rate × 730 × 100%)	$26,747	$18,104	$8,118
Colo	($3,816)	($2,552)	($1,906)
Ops reserve (5%)	($1,337)	($905)	($406)
Net cash flow	$21,594	$14,647	$5,806
Capex	$572,364	$319,422	$125,000
Simple payback	27 mo	22 mo	22 mo
Annual net cash ÷ capex	45%	55%	56%
Rate needed for 36-mo payback	$3.55	$2.06	$0.97
Cushion vs. base rate	−22%	−34%	−30%

Source: GPU Capital estimates from the inputs in Exhibits 1, 4, 5 and 6.

Source: GPU Capital estimates from UpStation catalog prices (September 9, 2026), GetDeploying and Silicon Data rates (September 24 to 29, 2026) and one recent quote.

Show data

The H200 and RTX PRO 6000 pay back in about 22 months, the B300 in 27

Node	Months to payback
8x H200 (HGX)	22 months
8x RTX PRO 6000	22 months
8x B300 (HGX)	27 months

The B300 earns 1.5× the monthly cash of an H200 on 1.8× the capital. That’s why it finishes last. The H200 gets there in under 22 months with the widest margin for error. For a lender, that counts more than the headline.

The PRO 6000 ties the H200, but only at $125K. At the $163,910 Tier 2 list, payback stretches to about 28 months. At $191,888 Tier 1, about 33 (est.). On this node, the purchase price is the whole trade.

What breaks each case

Three stress tests. One input moves, the rest hold at base.

Rate compression (one-time cut to the contract rate)

This is a single re-price, not an annual decline. Annual decay is the third table.

Exhibit 9. Payback after a one-time cut to the contract rate, months

Payback (months, est.)	Base	−10%	−20%	−30%
B300	27	30	35	41
H200	22	25	29	34
RTX PRO 6000	22	25	29	36

Source: GPU Capital estimates.

Dots mark the cut at which each node’s payback reaches the 36-month term. A 10-point drop in billed utilization moves payback exactly as a 10% rate cut does, because costs are fixed.

Source: GPU Capital estimates from UpStation catalog prices (September 9, 2026), GetDeploying and Silicon Data rates (September 24 to 29, 2026) and one recent quote.

Show data

The H200 has the widest cushion against a cut in the contract rate

One-time cut to the contract rate	8x B300	8x H200	8x RTX PRO 6000
0% cut	27	22	22
10% cut	30	25	25
20% cut	35	29	29
30% cut	41	34	36
40% cut	50	41	46

Utilization drop (rate held at base)

Exhibit 11. Payback at lower billed utilization, months

Payback (months, est.)	100%	90%	80%	70%
B300	27	30	35	41
H200	22	25	29	34
RTX PRO 6000	22	25	29	36

Source: GPU Capital estimates.

Same numbers, and not by accident. With costs fixed, a ten-point drop in billed hours does exactly what a 10% rate cut does. On a take-or-pay contract, utilization risk only shows up if the customer stops paying, or when the contract rolls and the node has to be re-rented.

Rate decays every year from day one

Exhibit 12. Payback when the rate falls every year, months

Payback (months, est.)	0%/yr	10%/yr	20%/yr	30%/yr
B300	27	31	40	Never
H200	22	25	30	46
RTX PRO 6000	22	25	31	Never

Source: GPU Capital estimates.

“Never” means cumulative cash never catches up to capex on that decay path. This is the case for short contracts that re-price as the market falls.

Where each node fails

B300 fails on the rate you lock. The 36-month breakeven is $3.55. The lowest reserved B300 listing today is $4.25, so there’s room, but the least of the three.

It’s also the most concentrated bet: $572K per node before a single optic. And Vera Rubin is already shipping. The H100’s posted rate fell from more than $7 in early 2024 to under $3 this year (GPU Appraisals, our affiliated appraisal practice). Expect the B300 to walk the same path, faster. At 30% a year it never pays back.

H200 fails on duration. Breakeven is $2.06. The cheapest published 36-month H200 commit is $2.57 (GetDeploying), 25% above that line. It’s the only node that pays back at all under 30% annual decay. Its risk is the roll. When the contract ends, it re-rents two generations behind the frontier.

RTX PRO 6000 fails on purchase price. At $125K it’s as robust as the H200 on a one-time cut. At list, it isn’t. And the spot market already trades close to its $0.97 breakeven: 1-month 8-GPU reservations list at $1.27, and single-GPU reserved floors go lower. An uncontracted PRO 6000 is exposed.

Buying at $125K is getting harder. Thunder Compute reports NVIDIA’s list price for the card at $16,000, up from an $8,565 launch MSRP, on GDDR7 shortages.

What we take from this

There’s no single winner. The PRO 6000 pays back fastest whenever the customer prepays. The H200 pays back fastest in every stress case. The B300 is last in every scenario we ran.

The H200 is the node we’d underwrite first today. Near-fastest payback, the widest rate cushion, and the only node that still pays back at 30% annual decay. The risk sits after the contract, not during it.

The RTX PRO 6000 matches it, but only when bought well below list. At $125K it’s a strong node. At $190K it’s a three-year trade with little room. Price the purchase before you price the rental.

The B300 pays back last and has the thinnest cushion. It needs a long contract near today’s rates and a customer strong enough to hold it through Rubin.

The rule we apply across all three: underwrite the contract and the price you paid, not the chip.

And payback is one lens, not the answer. Residual value, fabric capex and counterparty credit all move the result. That’s why GPU Capital screens nodes on the GPU Capital Yield Score alongside payback.