How AI Infrastructure Is Changing IPv4 Demand
Training fleets, inference endpoints and multi-region deployments all need public IPv4 — and most third-party services still don't speak IPv6. Here is how the AI buildout is reshaping the IPv4 market.
Why AI workloads still need IPv4
Every stage of the modern AI stack touches the public internet — and almost none of it can live on IPv6 alone. Training pipelines run massive crawler fleets that pull data from websites, feeds and APIs across the internet, and most of those targets are only reachable over IPv4. Inference services expose endpoints that third-party clients and on-premise systems call — and those callers overwhelmingly speak IPv4. Multi-region deployments need stable, routable public addresses in every region where they operate. The gap is in the ecosystem, not the protocol. IPv6 has been deployed for two decades, yet the long tail of the internet — legacy services, enterprise networks, ISPs in some regions, countless internal systems — still answers only to IPv4. An AI company cannot decide on its own to be IPv6-only, because the third-party services it depends on did not make that decision. All ~4.3 billion IPv4 addresses are already allocated; the free pools at the regional registries ran dry between 2011 and 2020. Every crawler, every inference endpoint and every region adds to demand against a supply that is structurally finished.
The $44-per-IP line item on the cloud bill
Cloud providers charge around **$44 per address per year** for a public IPv4. A small deployment gets by with a handful of addresses; an AI company running training, inference and data pipelines across several regions counts them in the hundreds or thousands. At that scale the number stops being a footnote: a thousand addresses at $44 each is $44,000 a year, before bandwidth and compute. It is the kind of line item finance teams notice — and why AI infrastructure teams increasingly look outside their cloud bill for address space.
Crawlers live and die by block reputation
A crawler fleet can send millions of requests a day. From the receiving side, that traffic is indistinguishable from abuse unless the source addresses carry a clean history. Blocks with spam listings, blocklist residue or a hijack record get throttled, challenged or simply refused — which quietly destroys the value of the data collection effort. Clean blocks therefore command a premium, and for AI teams they are not a nice-to-have but operational infrastructure. Email reputation works the same way: a dirty block cannot deliver mail, and a block abused by a previous tenant is hard to rehabilitate. Verifying a block's listing history before acquiring it is standard due diligence for serious buyers.
Where the market is: RIPE leads
The **RIPE region — Europe and Türkiye — carries the highest transfer volume in the world**. That concentration matters to AI companies for two reasons. First, supply is deeper: more blocks change hands, which means more clean, documentable space to choose from. Second, prices are more competitive in a liquid market than in thin ones. For teams that want to see the whole market rather than a broker's shortlist, knowing which blocks exist, which are dormant and which are clean is the real advantage — information that never appears on a price sheet.
Lease or buy: AI projects favour leasing
AI projects have uncertain lifespans: a training run has a finish line, a model gets replaced, a product pivot can dissolve an infrastructure need in a quarter. Committing capital to a balance-sheet asset for a need that may shrink is a poor fit. That is why leasing is the natural instrument for much of the AI buildout: at roughly **$0.40 per IP per month**, a /24 of 256 addresses runs about $100 a month — a modest operating expense that scales down when the project does. Buying the same /24 costs around **$27 per IP** (about $6,900), which only makes sense when the address space is a long-term asset. The rule is simple: short or uncertain horizon, lease; address space core to the business for years, buy. AI teams usually find themselves on the leasing side of that line.
The takeaway
The AI buildout is not creating a new supply of IPv4 — there is none. It is adding a new, structurally growing source of demand against a fixed base of ~4.3 billion addresses. The result is a market where clean space, liquidity and timing matter more than ever. Whether you lease or buy, the winning move is the same: know which blocks exist, which are dormant and which are clean before they appear on a broker's list. --- *See the dormant side of the address space on the IPDORM dark space map — [ipdorm.com](https://ipdorm.com).*