Lambda Labs review

It does not. Lambda Labs lists No charge for ingress or egress, so pulling checkpoints, datasets or model weights out of the cloud adds no separate line to the invoice. That matters for training runs, where the transfer back after a job can otherwise cost more than the compute that produced the file.

Usage is counted per minute, and the published pricing model is hourly. A job that starts and stops still pays for the interval it used, which is the clearest difference from providers that bill each second of a run.

Yes. The documented billing models are hourly and reserved, so teams can start on hourly pricing for experiments and move to a reserved term once a workload is steady enough to commit to. Both models are described in the public cloud billing documentation.

The public cloud images are linux. That covers container builds, model training and serving stacks that are Linux-first, and it narrows the choice for anyone who depends on Windows instances or licensed desktop software and needs a broader image list.

The listed server type is vps, so the catalogue is focused on virtual machines rather than a mix of instance types. Buyers who want ownership of dedicated bare metal hardware should weigh a specialist provider instead of assuming this range covers it.

Lambda Labs publishes SOC 2 Type II. That gives a vendor review something concrete to point at, which helps teams that must show an audited standard to their own compliance function before a purchase clears.

Nodes are connected with NVIDIA Quantum-2 InfiniBand. For distributed training that detail is decisive, because exchanges between accelerators sit on the critical path, and a slower link turns paid compute hours into time spent waiting on data.

Lambda Labs meters per minute, while Salad bills per second and PoolCompute bills per second. Fine-grained metering favours quick bursts and short experiments; Lambda Labs trades that for No charge for ingress or egress once data starts moving.