Vast.ai

Vast operates the software layer — search, containers, billing, the CLI and Python SDK — but the hardware belongs to hosts. Those hosts range from individual operators running a handful of consumer cards to professional facilities; Vast's pricing page describes supply spanning more than forty data centres and 68 GPU types, from an RTX 3060 up to a B200. Prices are set by hosts competing for renters, not by Vast, which is the whole mechanism behind the discount and also the reason quoted rates decay quickly.

The practical consequence is that "the price of an H100 on Vast" is not a number, it is a distribution. On 24 August 2026, public single-GPU on-demand listings for an H100 SXM ranged from $1.34/hr to well over $3/hr depending on host, location and verification status, with the cheapest offer coming from an unverified host in Moldova carrying a reliability score around 0.95. The same day, RTX 3090s started near $0.07/hr, RTX 4090s near $0.14/hr, A100 SXM4 80GB near $0.60/hr and B200s near $4.63/hr. Treat all of those as a snapshot, not a price list.

Every listing carries far more metadata than a cloud console typically shows. The two numbers that matter most are DLPerf and reliability.

DLPerf is Vast's own composite benchmark, meant to predict iterations per second on standard deep learning workloads rather than advertise theoretical TFLOPS. Vast's documentation gives a V100 about 21 points and a 1080 Ti about 10, and states that the score is tuned for CNN training, transformers and standard computer vision, and is less meaningful for unusual compute patterns. In sampled listings on 24 August 2026, the median DLPerf figures ran roughly 44 for an RTX 3090, 97 for a 4090, 344 for an H100 SXM and 680 for a B200 — a far more honest ranking for training work than VRAM alone.

Reliability is a separate per-machine score reflecting how consistently that host has stayed online. Alongside it, each offer exposes disk bandwidth, PCIe generation and lane count, internet up and down speeds, CPU model and core count, geolocation, whether the machine has a static IP, and the host's storage and bandwidth prices. Filtering on those fields — rather than sorting by price and taking the top row — is the single biggest determinant of whether Vast works well for you.

On-demand instances run at a fixed host-set price and cannot be interrupted. Interruptible instances work on a direct bid: you name a price, and a higher bidder can stop your instance. Vast's docs put the saving at 50 to 80 percent, and unlike Google Cloud's preemptible instances there is no 24-hour ceiling, so an unchallenged bid can run indefinitely. When you lose a bid the instance stops and running processes die, which makes checkpointing non-negotiable for anything longer than a coffee break.

Reserved instances are the third option: prepay for one, three or six months at on-demand priority for up to 50 percent off, convertible from an instance you are already running. That tier only makes sense once you know your steady-state usage, which most Vast renters do not.

This is where new users get caught. Vast charges independently for active rental, storage and bandwidth, and all three vary by machine.

Active rental accrues for every second the instance is connected. Storage accrues for every second the instance exists — including while it is stopped — which means a paused pod with a large disk quietly burns money. Sampled listings on 24 August 2026 clustered around $0.20/GB/month for storage, but ranged from a few cents to over a dollar. Bandwidth is billed per byte in both directions, commonly around $2.60/TB but running an order of magnitude higher on some hosts. If your pipeline streams a large dataset in and writes checkpoints out, price the bandwidth line before the GPU line.

Billing itself is straightforward: prepaid credit topped up by card via Stripe, or crypto through BitPay and Crypto.com, with optional auto-billing against a threshold. Hit zero and instances stop rather than being destroyed — but storage keeps charging, so destroy what you have finished with.

Vast runs two security tiers. Verified hosts have been manually tested for reliability and performance and are pitched for general-purpose AI and HPC work. Secure Cloud is a filter that restricts search to vetted datacentre partners: professionally managed facilities with a minimum of five flagship-class GPU servers, signed data processing agreements, and due diligence on ownership and source of funds. Those partners may hold ISO 27001, SOC 2 Type 2, PCI DSS or HITRUST certifications themselves, though Vast notes certification strengthens an application without being strictly required.

At the company level, Vast has completed SOC 2 Type 2, provides a SOC 3 report on request, and offers Business Associate Agreements for HIPAA-covered workloads on Secure Cloud. That is a meaningfully stronger compliance position than the platform's reputation suggests, but it applies to a filtered subset of supply — not to the cheapest offer in the list. Workloads run in unprivileged Docker containers with namespace, network and filesystem isolation, and Vast's own security guidance still recommends encrypting sensitive data before upload and keeping credentials off instances entirely.

Runpod is the closest comparison and the more curated of the two. Its Secure Cloud pods carry published list prices — $3.29/hr for an H100 SXM on 24 August 2026 — with no ingress or egress charges at all, plus a serverless tier Vast's marketplace does not directly match. Vast will usually win on the GPU line and can lose on bandwidth. Vast also offers hardware Runpod does not stock, which matters if you specifically need, say, a consumer 5090 or an older card that is fast enough for your model.

Against AWS, Azure or Google Cloud the comparison is not really about price. It is about whether you need quota systems, IAM, VPC peering, support contracts and a single audited vendor. If procurement requires those, no marketplace discount will substitute. If it does not, the hyperscaler premium on per-GPU compute is difficult to justify for training and batch inference.

Vast fits people who are paying attention. If you will filter on verification and reliability, checkpoint aggressively, destroy idle instances, and sanity-check storage and bandwidth rates before renting, it is the cheapest serious option available and the marketplace transparency is a genuine advantage over opaque cloud pricing. If you want to click one button and get a machine that behaves identically every time, the extra few dollars an hour elsewhere buys exactly that.