Best Value GPUs: Price per VRAM and Memory Bandwidth

How much memory or bandwidth your money buys.

Aug. 31, 2026 (updated)

We compare widely listed cloud GPUs using cross-provider median on-demand prices for the week of August 31, 2026. Memory value is price per GB of VRAM; bandwidth value is price per TB/s of memory bandwidth.

Plotting the market...

The L40 and the L40S match on both measured specs: 48 GB of VRAM and 0.86 TB/s of memory bandwidth, but the L40's median market price is 34% lower.

Value leaders

Looking for GPU Market rate Value
Lowest memory cost $0.55 /GPU/hr $0.0115 / GB-hour
Lowest memory cost, 80 GB+ $3.09 /GPU/hr $0.0121 / GB-hour
Lowest bandwidth cost $0.19 /GPU/hr $0.42 / TB/s-hour
Lowest bandwidth cost, 2 TB/s+ $3.09 /GPU/hr $0.52 / TB/s-hour

Best value for VRAM

Cost per GB-hour is the median hourly price divided by VRAM. It measures how cheaply you can rent memory capacity, not how fast the card uses it.

GPU VRAM Median price /hr Cost /GB-hour
48 GB $0.55 $0.0115
16 GB $0.19 $0.0119
256 GB $3.09 $0.0121
192 GB $2.99 $0.0156
48 GB $0.77 $0.0160
48 GB $0.85 $0.0177
48 GB $0.99 $0.0205
80 GB $1.76 $0.0220
32 GB $0.72 $0.0225
96 GB $2.20 $0.0229

The H100 is the most widely listed card in this ranking, and its memory cost is 44% above the market median. That premium is not explained by memory capacity alone.

The MI325X is expensive by the hour but sits below the market median on both measured resource costs.

Best value for memory bandwidth

Cost per TB/s-hour is the median hourly price divided by published memory bandwidth. It measures bandwidth per rental dollar, not end-to-end inference speed.

GPU Memory bandwidth Median price /hr Cost /TB/s-hour
0.45 TB/s $0.19 $0.42
6.00 TB/s $3.09 $0.52
5.30 TB/s $2.99 $0.56
0.77 TB/s $0.55 $0.72
7.70 TB/s $6.01 $0.78
0.90 TB/s $0.72 $0.80
0.96 TB/s $0.85 $0.89
1.94 TB/s $1.76 $0.91
4.80 TB/s $4.40 $0.92
8.00 TB/s $7.89 $0.99

The B200 has a high hourly rate but its bandwidth cost is well below the market median.

The L4 looks inexpensive by the hour, but its bandwidth cost is well above the market median.

Where memory and bandwidth value disagree

Some GPUs are much better value for one resource than the other. These are the largest splits in this week's market.

GPU Better value for Memory cost vs median Bandwidth cost vs median
Bandwidth +16% −28%
Capacity −20% +27%
Bandwidth +44% −9%
Capacity −18% +20%
Bandwidth +8% −16%

Heads up: These are medians across providers, not a rate you can book: the cheapest listing sits below them, and a provider's own page may quote a different figure, for example on tax, region or a promotion. Value figures divide that median by advertised VRAM or published bandwidth. Verify before provisioning. More on how we price.

What these metrics don't measure

These metrics cover VRAM and theoretical memory bandwidth only. They do not measure compute throughput, interconnect, software compatibility, provider reliability or actual application performance. Use these metrics to compare memory and bandwidth value, then check the GPU model pages for the rest.

Frequently asked questions

Which cloud GPU offers the cheapest VRAM?

The Nvidia RTX A6000, at $0.0115 per GB-hour: a median rate of $0.55 per GPU per hour over 48 GB, across 13 providers in the week of August 31, 2026. That is the cheapest memory among the datacenter and workstation GPUs ranked here.

Which GPU offers the cheapest memory bandwidth?

The Nvidia A4000, at $0.42 per TB/s-hour. The lowest cost per unit is not the highest bandwidth: it publishes 0.45 TB/s, against 8.0 TB/s for the B300, the highest-bandwidth card in this ranking. Cost per TB/s-hour is what bandwidth costs to rent, not how fast your workload will run.

Is the H100 good value?

On the two resources this page measures, separately. Its memory cost is 44% above the market median at $0.0415 per GB-hour; its bandwidth cost is 9% below the market median at $0.99 per TB/s-hour. These metrics do not cover compute throughput, interconnect, software ecosystem or provider reliability.

Why do the memory and bandwidth rankings disagree?

Because they price different physical resources. A GPU can offer cheap capacity and relatively expensive bandwidth, or the reverse, depending on how much memory its architecture pairs with how much bandwidth. The table above shows the largest current disagreements as each card's percentage above or below the market median on both metrics. The B200 shows the widest gap between the two this week.

Are datacenter GPUs better value than consumer GPUs?

On raw memory per dollar, often not: the Nvidia RTX 3060 is currently 2.3x cheaper per gigabyte than the RTX A6000, the cheapest memory among the datacenter and workstation GPUs ranked here. Consumer and datacenter cards are different products, though: features, support terms and hosting vary by card and by provider, so this is not comparing apples to apples. Consumer cards can be added to the chart with its class filter.

Methodology

Market rate. The cross-provider median on-demand price for the week of August 31, 2026, the most recent full week of data: each provider's own median first, then the median across providers, normalized to USD per GPU per hour. Only GPUs listed by at least 3 providers that week are ranked. The rankings above cover datacenter and workstation GPUs. Consumer cards can be added to the chart with the class filter.

Memory value. The hourly market rate divided by advertised VRAM in gigabytes, in dollars per GB-hour.

Bandwidth value. The hourly market rate divided by published memory bandwidth in TB/s, in dollars per TB/s-hour.

Relative to median. Percentage comparisons use the median cost per unit across the datacenter and workstation GPUs ranked here, for the same week. Negative means cheaper than the market median; positive means more expensive.

For the lowest price you can pay today rather than the weekly median, see the cheapest cloud GPUs guide; for how prices move over time, the GPU Price Index; for every card, provider and configuration we track, the cloud GPU index.

Back to top ↑