The tab that looks too good
You’ve got two listings open. One is a secure-tier box on a name provider, priced at whatever the current H100 SXM floor shows today. The other is a marketplace listing for the same card, quoted well below that. Your job needs to be done within a few hours.
The marketplace price gap is real and it’s large. Community and marketplace listings on cards like the RTX 3090, A100 PCIe, and MI300X run 70%+ below secure-tier rates for the identical card as of this writing. That’s not a rounding error, that’s a different cost structure entirely, and it’s why marketplaces exist. But the discount buys you a machine someone else configured, on a network you haven’t tested, with a disk that might still hold the last tenant’s cache.
Before you click rent, the question isn’t “is this cheaper.” It’s “what does cheaper cost me in setup hours.”
Check ingress before you check price
Rental clocks start the moment the instance boots, not the moment your job starts running. If your first step is pulling model weights, the download is billed time.
Take an INT4 quant of a mid-size MoE, say something in the range of DeepSeek V4 Flash’s footprint (284B total params at roughly 0.55GB per billion gives a rough 150GB file). On a host with a real 10Gbps link, that’s a few minutes. On a marketplace box where the advertised bandwidth is aspirational and the actual sustained rate is a fraction of that, the same download can eat two to four hours before your GPU does a single forward pass. At an hourly rate, that’s not a rounding error either, it can erase the entire discount you rented the box for.
The fix is cheap. Before committing to a multi-hour job, pull a small test file (a few GB) and time it. If the sustained rate implies your model weights will take longer to land than your actual compute job will take to run, walk away from that listing regardless of price. This is the single most common way people lose the marketplace discount without realizing it, and it never shows up on the price comparison page, because that page tracks rental rate, not delivered bandwidth.
Data hygiene isn’t optional on someone else’s box
Marketplace and community-tier machines are frequently re-provisioned between tenants with minimal cleanup. That cuts both ways: your predecessor’s leftover files might still be on disk when you boot, and your own files might survive for whoever rents the box after you.
So don’t assume the disk is clean, and don’t assume it’ll be wiped after you leave. Encrypt anything sensitive before it touches the volume, pull secrets from environment variables rather than files that persist in a snapshot, and if your workload touches proprietary data (a fine-tuning set, customer prompts, anything under an NDA) treat the local disk as public until proven otherwise. On a secure-tier instance from a name provider, isolation and wipe policy are part of what you’re paying for. On a marketplace box, they’re not guaranteed, they’re a courtesy.
When secure tier earns its premium
Skip the marketplace discount when any of these apply: you’re fine-tuning on data you can’t afford to leak, you’re running a multi-day job where a mid-run preemption costs you a checkpoint restart you can’t easily absorb, or you’re serving production traffic where a host going dark mid-request is a customer-facing outage, not an inconvenience.
Take the marketplace discount when the job is short, the data is public or synthetic, and a restart costs you minutes at most. Batch inference against an open weights model, benchmarking a quant, running eval suites, spinning up a throwaway vLLM or SGLang server to sanity-check a model page’s numbers against your own hardware, these are exactly the jobs marketplace tiers were built for.
Price spreads compound this decision. Some models show provider-to-provider spreads over 20x for identical weights, visible on pages like Llama 3.1 405B and Qwen3 235B A22B. Combine a marketplace GPU discount with a low-cost API provider quote on the same model and you can end up comparing an insecure, unverified rental against a fully managed, SLA-backed API on price alone, which is comparing two different products. Pull tier and provider from the same table before you decide which axis you’re optimizing.
Our tiers guide, linked from /guides, breaks down what secure, community, and marketplace actually promise contractually. Read that before you assume “marketplace” means “worse hardware.” Usually it means identical hardware, weaker guarantees.
The rule
Before renting any marketplace listing, run a small test download and compute the implied time to pull your model’s weights at that host’s real sustained rate. If that time exceeds roughly 10-15% of your expected job runtime, the discount is an illusion once rental hours are counted, and you should rent secure tier off /gpus instead. What this can’t tell you is whether that specific host’s bandwidth holds steady under load an hour into your job. Marketplace listings vary rig to rig and hour to hour. The only way to know is to have burned an hour on that exact host before.