IPv6 proxies fit AI data work in two places: bulk collection from the sources that publish IPv6 addresses — Hugging Face, PyPI, npm, Docker Hub, Wikipedia, Common Crawl — and the outbound web fetches of agents built on the model APIs, which are all dual-stack except DeepSeek's. They do not fit GitHub, Stack Overflow, arXiv, Kaggle or Reddit, which are IPv4-only, and they never fit a login.
Every provider page that ranks for "proxies for AI training data" says proxies are essential and none of them says which sources an IPv6 proxy can actually reach, so we measured it: AAAA records on the hosts AI pipelines pull from and on the API hosts agents call, looked up from two public resolvers on 20 September 2026, on top of our 102-domain census of 14 September. The result decides the routing rule this article ends with — IPv6 exit where there is an AAAA record, residential where there is not — and the arithmetic that makes it worth the trouble: $0.20/GB down to $0.083/GB on the IPv6 network against $1.00/GB down to $0.80 on residential.
Which AI data sources accept IPv6
An IPv6 exit reaches a host only if the host publishes an AAAA record. The sources, with the number of AAAA records each answered:
| Source | What pipelines pull from it | AAAA records | Over IPv6 |
|---|---|---|---|
| huggingface.co | Datasets, model weights, the Hub API | 8 (16 in the census) | Yes |
| pypi.org | Packages and their metadata | 4 (8 in the census) | Yes |
| www.npmjs.com | Package registry pages | 2 (4 in the census) | Yes |
| www.docker.com | Images and documentation | 2 (4 in the census) | Yes |
| en.wikipedia.org | Articles, the MediaWiki API | 1 (2 in the census) | Yes |
| commoncrawl.org | Crawl indexes and manifests | 1 | Yes |
| medium.com | Articles (403 to plain clients over IPv4 too) | 2 (4 in the census) | Yes, behaviour-gated |
| github.com | Code, issues, READMEs | 0 | No — IPv4-only |
| stackoverflow.com | Q&A | 0 | No — IPv4-only |
| arxiv.org | Papers | 0 | No — IPv4-only |
| www.kaggle.com | Datasets, notebooks | 0 | No — IPv4-only |
| www.reddit.com | Threads (apex has AAAA, www does not) | 0 on www | No — the site lands on www |
The pattern is not random. The sources that live behind a modern CDN or a large cloud front door are dual-stack; the ones that run their own edge often are not. GitHub is the important absence: a pipeline that clones repositories or reads issues at volume has to leave from an IPv4 exit, whatever its proxy provider says about "unlimited IPv6". Wikipedia is the important presence: a licensed, structured corpus whose API answers over IPv6 and whose guidelines ask for nothing more than a descriptive User-Agent with a contact and a modest rate — and whose bulk copies are published as dumps, so that the live site is for fresh pages and targeted samples, not for downloading an entire edition. The Wikipedia over IPv6 page has the etiquette in full.
The model APIs and the assistants: all dual-stack but one
The second place IPv6 matters is the agent side. Agents built on the model APIs fetch web pages, search and read documentation as part of a task, hundreds of requests per session, and their egress is ordinary HTTP traffic that can be routed like any other. First, the APIs themselves:
| API host | AAAA records | Note |
|---|---|---|
| api.openai.com | 2 | platform.openai.com 2, chatgpt.com 2; the corporate openai.com apex has none |
| api.anthropic.com | 1 | claude.ai 1, console.anthropic.com 1 |
| generativelanguage.googleapis.com | 8 | gemini.google.com 8, www.googleapis.com 8 |
| api.x.ai | 2 | grok.com 2; x.com itself is IPv4-only |
| api.perplexity.ai | 2 | www.perplexity.ai 2 |
| api.mistral.ai | 2 | |
| openrouter.ai · api.groq.com · api.together.xyz | 2 each | Aggregators and inference providers, all dual-stack |
| api.deepseek.com | 0 | IPv4-only, as is www.deepseek.com |
Routing API calls through a proxy does not change any provider's rate limits — those are per key, per project or per tier, never per address — so an IPv6 exit on the API path buys exactly two things: the cheapest egress for a service that calls the API from a fixed place, and a choice of country for the outbound address. What it does not buy is a way around a login: the assistants' web apps (ChatGPT, Claude, Gemini, Grok, Perplexity, Copilot) are account sessions and belong on a sticky residential exit, a point each assistant's own page on the use-case hub makes before it says anything else.
Routing an agent's fetches by AAAA
An agent that browses does not care which network its fetch leaves from; its operator's bill does. In our census 45 of 102 major sites accept IPv6 on the host their homepage lands on and 57 do not, and the split is close to even across sectors. So the rule that pays is a lookup, not a policy: before a fetch, resolve the host's AAAA record — cached, it costs nothing — and send the request to the IPv6 exit when there is one and to the residential exit when there is not. Roughly half the agent's traffic then moves at $0.20/GB or less instead of $1.00, and no fetch fails for want of a route. Three practical notes:
- Resolve at the exit, not at the agent. Use SOCKS5 with remote resolution (
socks5h://in Python,--socks5-hostnamein curl) or plain HTTP proxying, so that the exit — the machine with IPv6 connectivity — does the lookup and the connection. An agent running on an IPv4-only box does not need IPv6 itself. - Pace by /64 on the destination. A fresh address per request does not reset a site's counters if all the addresses share a subnet; our exits are spread across /48s and /64s so rotation changes the prefix. Keep a low rate per destination and honour a challenge page by backing off.
- Keep the login off the cheap path. If the agent has to sign in somewhere, that session is a sticky residential one in the account's country; everything anonymous can rotate.
The same logic is packaged in the QuanticData MCP server, whose search and scrape tools bill per successful result and choose the network for you, and in the web data API for AI for pipelines that want Markdown or JSON rather than raw HTML. The proxy route is for teams that run their own fetch layer and want the bytes at the lowest price.
The cost of a corpus on IPv6
Training and retrieval corpora are measured in gigabytes, which is why the per-GB price is the whole cost story on the address side. The IPv6 ladder — $20 for 100 GB, $70 for 500 GB, $110 for 1 TB, $250 for 3 TB — against residential at $1.00/GB on the $5 starter and $0.80/GB at 1,000 GB:
| Corpus size | IPv6 | Residential (at volume) | Difference |
|---|---|---|---|
| 100 GB | $20 | $100 | 5× |
| 500 GB | $70 | $400 to $500 | 6–7× |
| 1 TB | $110 | $800 to $1,000 | 7–9× |
| 3 TB | $250 | $2,400 to $3,000 | 10–12× |
Two honest limits. First, the difference applies only to the dual-stack share of the corpus; the IPv4-only sources still cost residential or datacenter rates, and a mixed job lands somewhere between the columns. Second, cheap bytes are not permission: the acquisition side of a training corpus is now the part US courts look at hardest — the 2025 rulings treated training on lawfully obtained material as fair use and the keeping of a pirated library as not, whatever it was later used for — so a pipeline that fetches public pages it is entitled to fetch, keeps per-record provenance (source URL, timestamp, response status, what robots.txt said) and takes bulk copies from published dumps where they exist is on the right side of that line. Our overview of US scraping law covers the cases; the AI crawler user-agent list covers how to identify your client honestly.
What an IPv6 proxy does not do for AI data
- It does not reach IPv4-only sources. GitHub, Stack Overflow, arXiv, Kaggle, Reddit, X, Discord, TikTok, eBay: no AAAA record, no route. Those go through residential or datacenter exits on the same account.
- It does not make a datacenter address look like a household. Our IPv6 addresses are datacenter-issued on /48 and /64 subnets. A site that rejects hosting ranges by ownership rather than by behaviour is not helped by a cheaper hosting address.
- It does not change API quotas or a platform's rules. Rate limits are per key; robots.txt and terms are per site; a proxy changes the address the traffic leaves from and nothing else.
- It does not carry a login well. Accounts want one stable address for hours. Rotating cheap addresses under a session is the fastest way to a verification loop.
Twenty-five platform-by-platform guides — the seven assistants, the search engines, Wikipedia, Amazon and the rest — are on the IPv6 proxies by use case hub, each with the DNS answer for its host and the jobs that fit.