VoidMob
Dashboard
Mobile proxies by use case

Proxies for AI Training Data

Dataset crawls across sites, countries and languagesMobile proxies for AI training data: crawl public pages across many sites and languages from real carrier IPs inside each country, rotated per request and billed by the gigabyte the crawl transfers.

Open dashboardSee the recommended setup

From $3.99 per GB, or a dedicated device from $49 a month.

Live pool for AI training data
and 180+ more countriesAll locations
Why it fits

Why a pay-per-GB pool fits AI training data

  • A dataset crawl is many short, independent fetches across many domains, and per-request rotation is the pool's default, so there are no sessions to manage

  • A corpus is text: with media off a page costs a small part of its full weight, and pay-per-GB billing passes that saving straight through

  • A proxy list per country or language edition, each with its own login and target, lets every record carry the edition it was collected from

Setup

Recommended configuration for AI training data

The settings you would set in the dashboard, with the reasoning for this workload.

Session settingsAI Training Data
PlanPay-per-GB pool
RotationPer request, default
Sticky sessionSticky list setting
ProtocolsHTTP(S), SOCKS5
Rotation
  • Per request for page fetches across the domain list
  • Hold one exit only through an archive or search listing that keeps its page position in a cookie
Sticky session
  • Sticky keeps one carrier exit for as long as that device stays online
  • A timer list swaps it every few minutes on a long archive walk
Concurrency
  • Run many domains side by side and few requests into each one: keep every domain at the rate its owner publishes, or at a slow default, and back off on 429 and Retry-After
Protocol
  • HTTP or SOCKS5 on the list's host, port and login, set once as the proxy URL in Scrapy, a requests session or any HTTP client the crawler runs on

AI agents

An agent can run this AI Training Data setup

Once you create an API key and fund the balance, an AI agent can price the AI Training Data setup above, buy it when you say yes, pull the connection details, rotate the IP and watch usage through VoidMob's open-source MCP server. Works with Claude Code, Codex and other local MCP clients.

purchase_proxyrotate_proxy_ipget_proxy_status
$ npx -y @voidmob/mcp
Explore the MCP server
Pitfalls

What goes wrong with proxies for AI training data

Symptom, cause, fix. Each one traces back to a setting on this page.

  • The GB bill climbs on every re-crawl while the dataset barely grows

    Each pass downloads every page in full again, most of which have not changed since the last one

    Store each page's ETag and Last-Modified values and send them back as conditional requests: an unchanged page answers 304 with no body. Where a sitemap carries lastmod dates, re-queue only the URLs it marks as changed.

  • The corpus holds the same article several times under different URLs

    Tracking parameters, print views, mobile paths and syndicated copies all serve one text, and the crawler queued each URL as a new page

    Strip tracking parameters before queueing, key each record on the URL in the page's rel=canonical link, and hash the extracted text so any copy that slips through is dropped before training.

  • Records collected through one country's list come back in another language

    The site chose the language from the Accept-Language header or a stored cookie rather than from the exit's location, and the crawler sent its default

    Send an Accept-Language that matches each list's country, start every edition without cookies, and tag each record with the exit country and the language the page itself declares.

    Check with Location Test

  • A domain starts answering 403 or 429 partway through the crawl

    The crawler sent a generic library User-Agent to a site whose policy asks for a named one, or fetched faster than that site's stated pace

    Send the descriptive User-Agent the site's policy asks for, with a contact address, and slow that domain to the rate its documentation or Retry-After header gives before queueing more of its pages.

Why it works

Why mobile IPs work for AI training data

Training datasets are built from public pages on many sites and in many languages, collected in bulk and refreshed as sources change. VoidMob sends each fetch through a real 4G/5G carrier exit in the country you target, rotates it per request across that country's pool, and bills by the gigabyte, so a text-only crawl pays for the text it brings back.

A training dataset is a crawl across many domains and languages, repeated as the sources change. Each site sets its own pace and its own language editions. Exits in the rotating pool are real 4G/5G devices in the country you pick, so a German crawl collects the pages people in Germany get.

Give each country or language edition its own proxy list, rotate per request, and keep media off: the corpus is text, and billing is per GB. Conditional requests and sitemap dates hold a re-crawl to the pages that changed. Web scraping covers general crawler setup; the calculator below turns pages per day into GB.

Collection here means public pages, fetched in bulk ahead of training. Each site's robots.txt and terms set the rules for crawling it. Agents that browse and act while a task runs are a different job, covered on the AI agents page.

Pricing

What it costs

Pay-per-GB shared pools, or dedicated devices billed per day, week or month.

Pay-per-GB

HTTP(S) and SOCKS5 with UDP. Rotate per request or hold a session.

  • 1 GB$3.99 ($3.99/GB)
  • 10 GB$35 ($3.50/GB)
  • 25 GB$81.25 ($3.25/GB)
  • 50 GB$149.50 ($2.99/GB)
  • 100 GB$250 ($2.50/GB)
  • Above 100 GBCustom volume plans

Dedicated devices

HTTP(S), SOCKS5 and OpenVPN on every device. VLESS/Xray on premium tiers.

  • United States flagUnited Statesfrom $129/mo
  • United Kingdom flagUnited Kingdomfrom $97/mo
  • Germany flagGermanyfrom $119/mo
  • France flagFrancefrom $79/mo
  • Brasil flagBrasilfrom $129/mo
  • Indonesia flagIndonesiafrom $72/mo
  • Italy flagItalyfrom $127/mo
  • Spain flagSpainfrom $119/mo
  • Canada flagCanadafrom $159/mo
  • Poland flagPolandfrom $59/mo
  • Thailand flagThailandfrom $59/mo
  • Australia flagAustraliafrom $119/mo
  • Singapore flagSingaporefrom $119/mo
  • Ukraine flagUkrainefrom $49/mo
  • Romania flagRomaniafrom $49/mo
  • Austria flagAustriafrom $79/mo
  • Ireland flagIrelandfrom $59/mo
  • Netherlands flagNetherlandsfrom $139/mo
  • Portugal flagPortugalfrom $89/mo
  • Lithuania flagLithuaniafrom $59/mo
  • New Zealand flagNew Zealandfrom $129/mo
  • Georgia flagGeorgiafrom $109/mo
  • Moldova flagMoldovafrom $55/mo
23 locations, monthly fromTiers and features

Prices updated

Data estimate

How much data does AI training data use?

Pre-filled with a typical setup for this use case. Adjust the inputs to match yours.

Rough estimate, not a billing forecast. Real usage varies by target, media mix and caching.

Use case
Per page

About 0.2 MB per page without media, over 30 days.

Rough estimate
~60GB/mo
Fits the 100 GB tier at $250/moSized with 20% headroom over the estimate.
See proxy pricing

Prices updated

Run AI Training Data on real mobile IPs

Pools from $3.99/GB, or a dedicated device for 1 to 30 days.

Get AI Training Data ProxiesInstant activation, no KYC, crypto accepted.
FAQ

Frequently Asked Questions

Common questions about mobile proxies for AI training data.

Which proxies fit AI training data collection?
For a crawl over many public sites in several countries, a pay-per-GB pool of real mobile carrier IPs, rotated per request. A proxy list per country collects that country's edition of each page, and billing by traffic suits a text-only crawl, which transfers a small part of what a full page load would.
How is collecting training data different from proxies for AI agents?
Training data collection is a batch job: a crawler fetches public pages in bulk before training starts and can rotate freely between them. An agent uses a proxy at run time, browsing or signing up while it works on a task, and often holds one exit for the whole task. The [AI agents page](/proxies/for/ai-agents) covers that second case.
Should my crawler identify itself with its own user agent?
Where a site asks for it, yes. Wikimedia's User-Agent policy asks scripts for a descriptive name with contact details and may block generic ones without notice. Under RFC 9309 a crawler also finds its robots.txt group by its own name, which should appear in the User-Agent it sends. A mobile exit changes the network a request leaves from, not which of those rules apply to the crawler.
How much data does a training-data crawl use?
Pages per day times what each fetch transfers. Raw HTML is a small share of a full page, so a crawler that skips images, video and scripts uses far less, and on a re-crawl an unchanged page answered with 304 sends no body at all. Sources that publish a dump, as Wikimedia does for its wikis, cost one download instead of a crawl.
Can I collect the same pages in several languages?
Yes. Create a proxy list per country, each with its own login, and send an Accept-Language that matches it, since a site may pick the language from that header as well as from the visitor's location. Tag every record with the exit country and the page's declared language, so the language mix of the dataset can be checked before training.
Are VoidMob proxies for AI training data real mobile IPs?
Yes. Traffic exits from physical 4G/5G devices on consumer carrier plans behind carrier-grade NAT, so the sites you work with see an ordinary mobile subscriber IP with that carrier's ASN and DNS, not a datacenter or residential range.
Do I need a dedicated device or the pay-per-GB pool for AI training data?
Dedicated is one real device on one carrier that only you use, with an IP you rotate on demand or hold. Pay-per-GB shares a country pool, rotates per request by default, and is billed by traffic. The fit card on this page says which one matches AI training data and why; most projects can start on the pool and move to dedicated once a workflow settles.
Can I check what a site sees from my current IP before I start?
Yes. The free IP checker, platform detection and location consistency tools on this site show IP type, ASN, geolocation consistency and browser signals, no account needed.
Keep exploring

More for AI training data and beyond

Other use cases

By platform

Locations for AI training data

Guides

Free tools

More