MarcusLammers/vast-rtx3090-market-6mo

Vast.ai RTX 3090 Spot Market, February-August 2026

Panel data from the vast.ai GPU rental marketplace, restricted to NVIDIA RTX 3090 offers. The public offer listing was polled every 10 minutes between 2026-02-13 and 2026-08-15. Each observation records price, hardware specifications, host reliability, and location. A derived lifecycle table gives the listing duration of every offer. Vast.ai does not publish historical listing data; this dataset was collected independently.

Summary

Observation period	2026-02-13 to 2026-08-15 (UTC), uninterrupted
Scan interval	600 seconds
Successful scans	25,948 (99.97% success rate)
Observations (offer × scan)	1,624,024
Unique offers (asks)	26,425
Unique machines	1,662
Unique hosts	614
Countries	66

Median listed price per month:

Month (2026)	Median price ($/hr)	Median DLPerf per $
Feb	0.122	363
Mar	0.135	331
Apr	0.136	327
May	0.174	249
Jun	0.175	252
Jul	0.149	298
Aug	0.148	300

Files

data/snapshots.parquet (1,624,024 rows)

The raw panel: one row per visible offer per scan.

Column	Type	Description
snapshot_id	string	UUID of the scan
timestamp	datetime (UTC)	Scan time
snapshot_seq	int	Monotonic scan sequence number
scan_interval_sec	int	Configured scan interval (600)
ask_id	int	Vast.ai offer ID
machine_id	int	Physical machine ID (stable across relistings)
host_id	int	Host (owner) ID
price_usd_hour	float	Listed price in USD per hour
verified	bool	Whether the machine is vast.ai-verified
reliability	float	Host reliability score reported by vast.ai (0 to 1)
dlperf	float	Vast.ai deep-learning performance score
dlperf_per_dpth	float	DLPerf per dollar-per-hour
inet_up / inet_down	float	Internet speed (Mbps)
geo_country	string	Country code
geo_region	string	Region within country
cuda_max	float	Max supported CUDA version
gpu_ram	float	GPU memory (MB)
cpu_cores	float	CPU cores allocated
cpu_ram	float	System RAM (MB)
disk_space	float	Disk space (GB)
driver_version	string	NVIDIA driver version

data/ask_lifecycle.parquet (26,425 rows)

Per-offer lifecycle: one row per unique offer, maintained incrementally by the logger.

Column	Type	Description
ask_id	int	Offer ID
first_seen / last_seen	datetime (UTC)	First and most recent scan the offer appeared in
survival_minutes	int	Minutes between first and last sighting (listing duration)
seen_count	int	Number of scans the offer appeared in
miss_streak	int	Consecutive scans missed since last sighting
min_price / max_price / last_price	float	Price range over the offer's lifetime
price_change_count	int	Number of observed price changes
is_active	bool	Whether the offer was still listed at the end of the capture
snapshot_seq	int	Sequence number of the last scan it appeared in
scan_interval_sec	int	Scan interval (600)

data/snapshot_meta.parquet (25,955 rows)

One row per scan attempt: status (ok/error), row_count, fetch_ms, http_status, error details, and logger version. Useful for verifying data completeness.

Methodology

A logger queried the public vast.ai offer listing every 600 seconds with a GPU filter of RTX 3090, appending every visible offer to the snapshot log. The lifecycle table was derived incrementally: an offer's survival_minutes is the time between its first and last appearance across scans.

Known limitations

• Each scan returns at most ~64 offers (the query's result cap), so the dataset covers the visible top of the listing rather than the full RTX 3090 supply at every moment.
• Only offers matching the RTX 3090 filter were captured.
• An offer leaving the listing can mean it was rented or that it was delisted or went offline. The two cases can often be separated using machine_id: a machine that disappears and later relists was likely rented in between.
• survival_minutes measures how long an offer stayed listed, not machine uptime.
• Offers still active at the end of the capture (is_active = true) are right-censored.

Usage

import pandas as pd

snap = pd.read_parquet("data/snapshots.parquet")
life = pd.read_parquet("data/ask_lifecycle.parquet")

# Daily median price index
idx = snap.set_index("timestamp")["price_usd_hour"].resample("D").median()

# Host-level listing survival
hosts = life.merge(
    snap.drop_duplicates("ask_id")[["ask_id", "host_id"]], on="ask_id"
).groupby("host_id")["survival_minutes"].median()

Or with the datasets library:

from datasets import load_dataset

ds = load_dataset("MarcusLammers/vast-rtx3090-market-6mo", "snapshots")

Possible uses

• Spot-market price dynamics: how listing prices respond to supply changes over time.
• Survival analysis: predicting how long an offer stays listed. As a reference point, a baseline model using only host history reaches AUC 0.90 out-of-sample for 24-hour survival.
• Reconstructing rental demand from machine_id disappearance and relisting patterns.
• Studying the persistence of host behavior over multiple months.

License and attribution

CC BY 4.0. Free to use, share, and adapt with attribution.

Collected and published by Marc Lammers (augustus.cloud). Please credit "Marc Lammers, Vast.ai RTX 3090 Spot Market dataset" when using this data.