GPU Rent Jumped 15% in 5 Weeks: Rent CHEAP Before the Next Rise 👇
GPU Rent Jumped 15% in 5 Weeks: Rent CHEAP Before the Next Rise 👇 NVIDIA reported on November 17. Revenue hit $115.8 billion for the quarter. That is up 103 percent from a year ago. Data center revenue hit $107.6 billi
GPU Rent Jumped 15% in 5 Weeks: Rent CHEAP Before the Next Rise 👇
NVIDIA reported on November 17. Revenue hit $115.8 billion for the quarter. That is up 103 percent from a year ago. Data center revenue hit $107.6 billion. That line is up 110 percent from a year ago.
The same call carried a split. Token prices fell nearly 60 percent since May. GPU rental prices rose at the same time. Old Hopper cards now re-contract at up to twice their first price. B300 rates rose more than 60 percent since launch.
This guide prices the squeeze. It shows the math for one real job. It gives three ways to earn from it.
Why rents rise while tokens fall
Demand for compute still outruns supply. Vera Rubin earned nearly $20 billion in its first quarter. NVIDIA calls it the fastest product ramp in its history. Next quarter, NVIDIA guides $130 billion in revenue. Its supply commitments hit $390 billion. Most of that money is locked into multi-year memory deals.
Cheaper tokens mean more use. More use means more GPUs. More GPUs mean higher rents. That loop is the squeeze.
The math
One formula runs every GPU bill. GPUs times hours times the rate per GPU hour. That is the whole bill in one line.
| GPU | Lambda | RunPod Community | AWS |
|---|---|---|---|
| H100 SXM, 80 GB | $4.59 | $2.69 | $6.88 |
| H200, 141 GB | not listed | $3.59 | $7.91 |
| B200, 180 GB | $7.69 | $5.98 | $14.24 |
Rates are per GPU hour, on demand, in US dollars. Prices are list prices from the provider pages, checked November 2026. They change often. Lambda raised its H100 rate from $3.99 to $4.59 in five weeks. That is a 15 percent rise.
Take one fine-tune job. It needs 8 H100 cards for 100 hours. On Lambda it costs $3,672. On AWS it costs $5,504. On RunPod Community Cloud it costs $2,152. Same job.
The gap between the dearest and the cheapest is $3,352.
One card running all month costs real money. An H100 at $4.59 an hour costs about $3,351 a month. The same card on RunPod Community costs about $1,964 a month. The gap pays for the card search.
Pick the card by memory, not by name
The model must fit in the card's memory. An 80 GB card holds about twice the model a 40 GB card holds. A 180 GB B200 holds more than four times the 40 GB A100. Pay for the memory you need. Skip the memory you never use.
The H200 sits between the H100 and the B200. It costs $3.59 an hour on RunPod Community. Its 141 GB of memory suits long-context jobs. A B200 rents for about 68 percent more per hour than an H100. One cloud pricing guide puts its training speed at about twice an H100. The cost per job can end up lower on the dearer card.
Path 1: Buy low, sell the job
Rent the cheapest card that runs the job. Quote the customer a flat price for the finished job. Keep the gap between the flat price and the rent. The gap is the business.
The spread is wide right now. The same H100 costs $2.69 an hour on one cloud and $6.88 on another. That is about 2.6 times. A shop that prices the job at the dear rate and rents the cheap card keeps most of the fee.
Start with jobs you already know. Fine-tunes and batch inference both fit. Each job has a fixed GPU count and a fixed hour count. Price it once. Repeat it often.
Path 2: Rent old iron
New cards cost more per hour. Old cards still run most jobs. An A100 with 40 GB costs $1.99 an hour on Lambda. That is less than half the H100 rate. A V100 costs $0.79 an hour.
A model that fits in 40 GB runs fine on the A100. For light jobs, the V100 costs less than a dollar an hour. The newest card is the wrong pick when the old one fits.
Path 3: Lock the rate before the next rise
Rates rose 15 percent in five weeks. A steady workload should lock a rate. Lambda sells 1-Click Clusters for terms of 2 weeks to 1 year. A B200 cluster of 256 cards costs $8.87 per GPU hour.
That rate sits above the $7.69 on-demand rate. The extra money buys a promise. The cluster is yours when you need it. On-demand cards can sell out at the wrong hour.
For terms of one year or more, Lambda quotes by request. Ask for the quote before the next rise. A locked rate turns a moving cost into a fixed one.
What can go wrong
Prices move fast. Lambda raised H100 by 15 percent in five weeks. A margin built on this week can break next month. Re-check the rate before you quote a price.
Community clouds are cheap for a reason. Cards can vanish. Support is thin. A job that must finish on a fixed date needs a safer home.
Reserved clusters cost more per hour than on-demand singles. The promise costs money. Do not pay for a promise you never use.
New supply can cool rents too. Rubin is shipping in volume now. Rents fall when supply catches up. A bet on ever-rising rent can lose.
Verdict
The squeeze is real. Token prices fell. GPU rents rose. The gap between the two is where small shops earn.
Start with one number. Pick one job you run this month. Price it on three provider pages. The cheapest page is your first margin.
Follow along
I test AI tools and build in public:
- GitHub: https://github.com/ovu-lab
- YouTube: https://youtube.com/@oVuCyber
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.