Vast.ai
A cloud GPU compute rental platform offering spot and on-demand instances across multiple GPU models, primarily targeting AI and machine learning developers and researchers.
Timeline (last 90 days)
The author recommends first running a benchmark on Vast.ai to measure tokens/second, choosing on-demand instances to avoid being preempted, syncing code with git archive (noting that committed secrets will be packaged), and destroying the instance immediately after use to avoid disk billing.
Renting an RTX 5090 on Vast.ai ($0.584 per hour), using bf16 and torch.compile, Keera's pretraining throughput was about 234,000 tokens/second, the 2B-token training took 2 hours 23 minutes, and the validation loss was 1.501.
In GPU Index's price data snapshot, the lowest price for H100 is $1.37/hr (Vast.ai, spot instance), for H200 is $2.62/hr (CoreWeave, spot), for A100 is $1.15/hr (Paperspace, three-year contract), for L40S is $0.79/hr (RunPod, community cloud), and for B200 is $4.26/hr (CoreWeave, spot).
In the professional GPU rental and hosting space, in addition to the three major cloud providers AWS, GCP, and Azure, there are emerging players such as CoreWeave, Lambda Labs, RunPod, and Vast.ai.
RunPod and Vast.ai aggregate distributed computing power through a market-based competitive pricing model, offering price flexibility but relatively weaker stability.
Lambda, CoreWeave, RunPod, and Vast.ai are emerging cloud platforms focused on AI compute.
Vast.ai uses an aggregation model that unifies compute from individual mining rigs and small data centers scattered around the world into a marketplace, where buyers can bid by the hour.
Computing matchmaking platforms such as Vast.ai and RunPod offer the lowest prices because what is being rented is essentially other people's idle machines.
Typical players in the offtake network space include GPU cloud platforms such as CoreWeave and Crusoe, as well as P2P compute markets such as Vast.ai and RunPod.
Vast.ai is a compute-sharing marketplace with lower prices but slightly inferior stability.
4 more timeline events
All Facts (11)
Renting an RTX 5090 on Vast.ai ($0.584 per hour), using bf16 and torch.compile, Keera's pretraining throughput was about 234,000 tokens/second, the 2B-token training took 2 hours 23 minutes, and the validation loss was 1.501.
50%UnverifiedThe author recommends first running a benchmark on Vast.ai to measure tokens/second, choosing on-demand instances to avoid being preempted, syncing code with git archive (noting that committed secrets will be packaged), and destroying the instance immediately after use to avoid disk billing.
50%UnverifiedIn GPU Index's price data snapshot, the lowest price for H100 is $1.37/hr (Vast.ai, spot instance), for H200 is $2.62/hr (CoreWeave, spot), for A100 is $1.15/hr (Paperspace, three-year contract), for L40S is $0.79/hr (RunPod, community cloud), and for B200 is $4.26/hr (CoreWeave, spot).
50%UnverifiedRunPod and Vast.ai aggregate distributed computing power through a market-based competitive pricing model, offering price flexibility but relatively weaker stability.
50%UnverifiedIn the professional GPU rental and hosting space, in addition to the three major cloud providers AWS, GCP, and Azure, there are emerging players such as CoreWeave, Lambda Labs, RunPod, and Vast.ai.
50%UnverifiedVast.ai uses an aggregation model that unifies compute from individual mining rigs and small data centers scattered around the world into a marketplace, where buyers can bid by the hour.
50%UnverifiedLambda, CoreWeave, RunPod, and Vast.ai are emerging cloud platforms focused on AI compute.
50%UnverifiedComputing matchmaking platforms such as Vast.ai and RunPod offer the lowest prices because what is being rented is essentially other people's idle machines.
50%UnverifiedTypical players in the offtake network space include GPU cloud platforms such as CoreWeave and Crusoe, as well as P2P compute markets such as Vast.ai and RunPod.
50%UnverifiedVast.ai is a compute-sharing marketplace with lower prices but slightly inferior stability.
50%UnverifiedCrofAI claimed to run Kimi K3 on RTX Pro 6000 units rented from Vast, but even with Q2_K quantization the model requires approximately 802 GiB of VRAM, while the largest 8-GPU RTX PRO 6000 machine available on Vast has only 765 GiB of VRAM.
50%