On-demand NVIDIA cloud GPUs billed by the hour. No contracts, no caps. Deploy AI in minutes.
- Tweets295
- Following97
- Followers61
- Likes25
ALT Massed Compute graphic titled Three models, three jobs, three picks. Clef labels on an A100 SXM4 at $1.38 an hour, the least expensive card we ran that held the weights, at 270.8 ms per decision. JEV routing a ticket on an RTX PRO 6000 at $2.19 an hour, with 2.1x the decisions per second of the A100. JEV writing on an A100 PCIe at $1.35 an hour, with the most output tokens per dollar. Kolibri writing at scale on an RTX PRO 6000 at $2.19 an hour, at 1,885 output tokens per second across 32 requests.
ALT Massed Compute graphic titled Value, speed, and fit are three different cards. For LFM2.5, the L40S at $0.97 an hour is the best value at 1,110 output tokens per second per dollar. The A100 SXM4 at $1.38 an hour has the most throughput at 1,415 tokens per second. The RTX PRO 6000 at $2.19 an hour has the fastest first token at 101.9 ms. For MiniCPM-V-4.7, the RTX PRO 6000 was the one launched card that finished the test, at 1,019 tokens per second.
ALT Massed Compute graphic titled What a GPU quote is really saying. Cost per GPU-hour is the starting point, and storage, transfer, and idle time shape what you pay. Memory bandwidth often sets real speed. Interconnects decide whether more GPUs add speed. Training is a burst, and inference grows with every user. On-demand fits changing work, and reserved fits steady work.
ALT Benchmark card for MiniCPM-V-4.7-35B-A3B on a single RTX PRO 6000 Blackwell with vLLM, showing 1018.8 output tokens per second at concurrency 32
ALT Benchmark card for LFM2.5-24B-A2B on a single A100 SXM4 with vLLM, showing 1415.3 output tokens per second at concurrency 32
ALT Massed Compute graphic titled How companies get NVIDIA GPUs right now, from CNBC. Five options are listed. Hyperscalers are trusted, with capacity that can run short. Large neoclouds mean upfront payment and a wait for chips. Smaller neoclouds, highlighted, offer faster access and more flexible terms. Hourly rental costs less than buying for early teams. On-premises gives full control plus power, space, and staff. The hourly spot price of an NVIDIA B200 has more than doubled since March, per the Ornn index via CNBC. Sunny Smith, co-founder and CTO, told CNBC customers often commit to capacity when they expect prices to rise.
ALT Massed Compute graphic titled Same 32 GB. Different building. GeForce RTX 5090 is the desktop you sit at, with 32 GB GDDR7, a card you buy and own, for games and local models one person at a time. RTX PRO 4500 is a server you start and stop, with 32 GB on Blackwell at $0.92 an hour, running while you are away, and a second card for $1.84 an hour. For more than 32 GB on one GPU, the RTX PRO 6000 Blackwell has 96 GB at $2.19 an hour.
ALT Massed Compute chart titled Pay for the memory the job uses. 1x RTX PRO 4500 at $0.92 an hour with 32 GB, for one job that fits in 32 GB, highlighted as the pick. 2x RTX PRO 4500 at $1.84 an hour with 64 GB split across two cards, for two jobs or one job that splits cleanly. 1x RTX PRO 6000 at $2.19 an hour with 96 GB in one pool, for one job that needs more than 32 GB on one GPU.
ALT Massed Compute chart for Qwen-Image-2.1 at 1024 by 1024. Headline reads The cheaper hour made the pricier image. RTX A6000 at $0.57 an hour takes 29.5 seconds for $0.0047 per image. L40S at $0.97 an hour takes 16.5 seconds for $0.0045 per image, the lowest cost. RTX PRO 6000 Blackwell at $2.19 an hour takes 9.3 seconds for $0.0056 per image. 2048 by 2048 needs more than 48 GB, and Blackwell finished it in 51 seconds.
