Answers

Questions people actually ask

Hardware, billing, storage and serverless. The docs go deeper on every one of these; this is the short version.

Platform

The basics

What is Meshive?: Meshive is a task-first GPU cloud for AI training, inference, rendering, and serverless model serving. It provides instant access to high-performance NVIDIA GPUs, including RTX 5090, RTX PRO 5000, and RTX PRO 6000, at up to 70% lower cost than traditional hyperscalers, billed in fixed one-minute intervals, with zero cluster provisioning.
What GPU models are available on Meshive?: Meshive supports a broad range of NVIDIA GPUs, including the RTX 5090 (32GB VRAM), RTX PRO 5000 (48GB VRAM), and RTX PRO 6000 (96GB VRAM) for LLM fine-tuning, vLLM inference, and ComfyUI rendering.
How does Meshive reduce GPU cloud costs by up to 70%?: Meshive operates an automated distributed compute marketplace connecting verified high-performance hardware hosts worldwide. Automated driver installation, health checks, and Kubernetes clustering eliminate traditional data center overhead, passing direct cost savings to users.
How does Meshive Serverless Inference work?: Meshive Serverless provides OpenAI-compatible HTTPS endpoints for LLM text, image, and video generation. You can deploy models with one click, autoscale within a replica range you set, and stream or receive async webhook callbacks directly to your S3, Cloudflare R2, or MinIO storage.

Pods and billing

How is a pod billed?: In fixed one-minute intervals while it runs, with no minimum and no monthly commitment. Rates are reviewed and re-adjusted every two weeks, and each line item — GPU, vCPU, RAM, storage — is visible in the create flow before you confirm.
What does spot mean here?: A discounted rate on the same hardware, with the difference that the pod may be reclaimed and reallocated to a different node. Use it for interruption-tolerant work such as checkpointed training or batch processing, and keep services on demand. Spot pods cannot use local volumes, since those would be lost on reallocation.
Can I move a pod to a different GPU?: Not in place — hardware is pinned to the machine the pod was placed on. Save your image or data to a volume, then create a new pod with the GPU you want and mount the same data.
Is there free credit?: Sign-up promotional credit covers serverless inference. GPU pods run on paid credit.

Storage and data

What happens to my data when a pod stops?: Anything on a persistent volume survives. Local volumes live on the pod's host machine; network volumes live on a storage node and can be mounted by any pod in the workspace. Everything outside a volume goes with the pod.
Can my data be encrypted?: Yes, on either kind of volume, free of charge, chosen at creation. The key is never kept on the host machine, so an encrypted volume is unreadable off a disk that has been lost, resold or physically accessed. Expect roughly 3–5% throughput cost.
Am I billed for storage while the pod is stopped?: Yes. Storage is billed while the data exists, because those bytes still occupy a physical disk. Delete volumes you no longer need — deleting a pod can remove its local volumes in the same step.

Serverless

Do I pick the GPU for a serverless deployment?: No. You set a price cap per replica-hour and a replica range; the cheapest GPU that fits under the cap is assigned, and re-assigned later if a cheaper fit appears. Every candidate's hourly price is shown before you confirm.
Can I serve a model that isn't in the catalog?: Yes — register any Hugging Face repo and the metadata, VRAM estimate and quantization are derived for you. Custom image and video registration is temporarily paused while the generation engine moves to ComfyUI; official image and video models deploy normally.
Is the API really OpenAI-compatible?: Yes — chat, completions and embeddings work with the OpenAI SDKs by changing the base URL and the key. Image and video add a Meshive async path for jobs longer than the 30-second sync budget.
Can I rent out my own GPUs?: Yes. Hosting is automated end to end — drivers, Kubernetes, health checks — and a host machine exposes four ports, not four hundred. Fees start at 7.5%.