Vast.ai bid pricing strategy: rates that actually win jobs

Updated April 2026 · Live Vast.ai rates as of 16 minutes ago

Vast.ai is a bid-priced marketplace — you don't pick a fixed rate, you set a per-hour ask and renters' jobs match against the cheapest box that fits their requirements. Most new hosts mis-set this two ways: too high and they sit idle for weeks; too low and they leave dollars on the table for months. This guide walks through how Vast actually clears jobs, how to set your first bid, and what to do when the dashboard shows your card going unused.

How Vast's pricing actually clears

Renters search Vast with a query — "GPU type, VRAM, dlperf score, region, on-demand or interruptible" — and see boxes sorted by ascending $/hour. Your card surfaces in their list if it matches the filter. The cheapest box that fits the job's spec wins. Three implications:

• You compete inside a peer cohort, not against the whole site. An RTX 4090 with 24 GB VRAM and dlperf 39 is competing against other 4090s with similar dlperf and similar disk speed. You don't need to undercut H100s.
• Bid floor matters more than ceiling. If the cohort floor is $0.30/h and you're at $0.45/h, you'll see almost no jobs. If the floor is $0.30/h and you're at $0.32/h, you'll still match most queries because not every renter sorts strictly by price.
• "On-demand" and "Interruptible" are separate auctions. Same physical card, two price tracks. Interruptible clears at 30–60% of on-demand. Most boxes serve both.

Setting your first bid

The simplest method, in this order:

1. On the Vast cloud console, search for renters' view of your exact GPU model. Note the cheapest 5–10 listings' $/h.
2. Set your on-demand ask at the 25th percentile of that band. You'll be visible to most queries without burning yield by living at the absolute floor.
3. Enable interruptible at 50% of your on-demand rate. This catches spillover demand at low marginal cost — a job that preempts you only costs the time it ran, and Vast pays you for that time.
4. Wait 48–72 hours. Your reliability score is provisional in the first week, so utilization will start lower than steady state.

As of right now, RTX 4090 cards on Vast are clearing around $0.136/h. A new 4090 host using the rule above would set on-demand at roughly $0.142/h and interruptible at $0.075/h.

Reliability score — the multiplier you can't bypass

Vast publishes a reliability score (0–100%) per box. Renters who care about long-running training jobs filter to ≥95%. Boxes below 90% get only spot / experimental traffic. The score moves down whenever your card fails a heartbeat, the SSH connection drops, or a job exits with an instance-side error. Three things lift it:

• Stable network. 100 Mbps symmetric is the practical minimum. Drops > 30 seconds register as offline.
• Cooling headroom. Cards that thermal-throttle into a job failure tank the score fast. Aim for <78°C under load.
• Long enough uptimes. Reboots, kernel updates, and Docker daemon crashes all read as instability. Schedule maintenance into a single weekly window, not ad hoc.

Adjusting after launch

After your first 7 days, look at the dashboard's "Recent earnings" graph. Three patterns:

• Dashboard shows <30% utilization. Drop on-demand by 10% and interruptible by 15%. If still under 30% after another 72 hours, the cohort floor has moved against you — re-survey the listings page.
• Dashboard shows 80%+ utilization with constant queue. Raise on-demand by 5%. If utilization stays high, raise another 5%. Stop when utilization dips below 70%.
• Interruptible at 100% but on-demand at 20%. Renters are price-sensitive on this card right now. Lower on-demand by 8–10% and interruptible by 5%.

Datacenter cards play differently

H100 / H200 / B200 hosts are mostly competing on dlperf and reliability, not raw $/h. Right now H100 SXM on Vast is clearing around $1.468/h. The cohort spread is tighter (5–10%) than on consumer cards (often 30–50%). Bid the median, not the floor — datacenter renters favor stability over saving $0.10/h.

Common mistakes

• Setting one rate and never revisiting. Vast's clearing prices move 10–30% week to week. Stale rates leave 20% of yield on the table.
• Disabling interruptible to "force higher rates." Cuts you out of the largest queue without raising on-demand prices for anyone — renters just find the next box.
• Optimizing for top-of-listing. Floor pricing wins jobs but at margin near zero net of power. The 25th percentile is the actual sweet spot.
• Ignoring dlperf. An undervolted card running 5–10% slower at lower watts is more profitable per kWh, but only if the dlperf score doesn't fall out of the renter's filter band. Test before committing the undervolt to 24/7.