The whales left IPv4. The crowd took over

For a decade, the IPv4 transfer market had a simple center of gravity. A handful of hyperscalers bought address space faster than anyone else, in blocks no one else could absorb, and the rest of the market priced off their demand. In 2025, that center moved. The biggest buyers stepped back hard, and the gap did not close because demand fell. It closed because demand changed hands.

Part of what is happening is a migration. Workloads, and the customers behind them, are moving off expensive hyperscaler infrastructure toward leaner, cheaper providers. The economics are stark: the Uptime Institute pegs a top-end GPU instance at around $98 an hour on a hyperscaler, compared with roughly a third of that on a specialized neocloud, and that market is projected to surpass $23 billion in 2025, with triple-digit growth. But the migration splits in two the moment it touches IPv4, and the split is the whole story. The cheaper hosting and VPS providers buy address space to scale because their model assigns a public IP to every server they rent. The AI and GPU providers do not, because a training cluster is mostly private fabric with a thin public edge. One-half of the migration drives IPv4 demand. The other half barely touches it.

The giants pulled back

The pullback is not subtle, and it is plain in the RIR transfer record. Amazon, the steadiest large-block buyer in the market, acquired about 17 million addresses in 2024 and 10 million in 2025, a 40 percent cut to its lowest level since 2021, sitting on top of a stable 2022-to-2024 plateau. This is a step down, not noise. Microsoft cut roughly a fifth. Google, always lumpier, cut more than half. Together, hyperscaler buying fell about 40 percent year over year.

The cleaner way to see it is to share. For three straight years, the hyperscalers were roughly two-thirds of all demand-driven transfers. In 2025, they were under a third. In a single year, the market flipped from giant-led to crowd-led.

What makes the retreat strange is that it happened while the same companies were growing fast. AWS revenue grew around 29 percent in 2022 when Amazon was buying 23 million addresses. By late 2025, AWS was re-accelerating toward 24 percent, its fastest in three years, while Amazon’s buying had halved. Azure and Google Cloud told the same story at higher growth rates. For years, more cloud revenue meant more IPv4. Since 2024, the two have moved in opposite directions.

Growing on AI, not on addresses

The reason the lines diverged is what the growth is made of. This cycle is AI, and AI is compute-bound, not address-bound.

Google is the clearest case. Its cloud business ended 2025 at a run rate above $70 billion, which Alphabet attributes directly to AI demand. In the first quarter of 2026, it grew 63 percent, with AI products named the single largest driver, and generative AI usage up nearly 800 percent year over year. None of that required a proportional pile of IPv4. The investment that produced it was silicon. Alphabet has built custom TPUs for more than a decade and is now shipping its seventh generation, with an eighth aimed squarely at the agentic era. A bet on AI infrastructure made over ten years is paying off now, and it is paying off in compute, not in addresses.

The tell is in what is scarce. Google has said its cloud growth is capacity-constrained, with its backlog doubling to over $460 billion. The constraint on AI is GPUs, TPUs, and power. It is never IPv4. A modern training cluster is mostly an internal machine, tens of thousands of accelerators talking to each other over a private, non-routable fabric with only a thin public edge. Growth that runs on silicon and electricity rather than on addresses is growth that decouples from IPv4, and the same logic determines who replaces the hyperscalers as buyers.

So who took over, and why they buy

If the giants left, who filled a market that did not shrink? A record number of buyers. Close to 3,000 distinct organizations acquired address space through demand-driven transfers in 2025, the most in the market’s history, and the typical deal was small, a /22. The market did not get smaller. It got wider.

The biggest share by volume is access networks. Telecoms and ISPs are expanding broadband and fixed-wireless, relieving carrier-grade NAT, building out rural coverage under programs like the US BEAD rollout, and laying connectivity infrastructure. The largest single buyers in 2025 were a submarine-cable operator and the consolidated address pools of carriers like AT&T, Verizon, and Vodafone.

The other driver is hosting, and the structural story here is worth slowing down on, because it explains why two kinds of fast-growing compute providers behave in opposite ways.

A traditional hosting or VPS provider sells servers. Every virtual machine, website, or game server it rents needs its own public, routable address, the way every unit in an apartment block needs its own street number. Grow the customer base and the address count grows with it, close to one for one. That is why providers like Hostinger and Hetzner keep buying blocks: IPv4 is a direct input that scales with the business. It is the same model that drove the cloud’s address hunger a decade ago, the model that had Amazon buying nearly 27 million addresses in a single year, only now at a leaner price point.

A GPU neocloud sells compute, and a GPU cluster is built the opposite way. Tens of thousands of accelerators are wired to one another over a private, high-speed fabric that never touches the public internet, more like machines on a single factory floor than tenants in a building. The cluster needs only a thin public edge, a handful of addresses to manage it, pull in training data, and serve the finished model. A ten-thousand-GPU cluster might need a few dozen public addresses, whereas a hosting fleet of the same size needs tens of thousands. So the providers building the AI boom, CoreWeave, Lambda, Crusoe, Nebius, barely register as buyers, and what little they need they lease or take from their data-center partners rather than acquire.

That is the structural change underneath the market. The compute growing fastest, AI, is the least address-hungry per dollar ever built. The compute that still consumes IPv4 the old way, hosting and VPS, is the steadier, slower-growing, price-sensitive layer. When workloads migrate from expensive hyperscalers to cheaper providers, the hosting half of that migration shows up in the transfer market, and the GPU half does not.

What is missing from the buying tells the same story from the other side. The proxy networks and AI data-collection businesses that are among the heaviest users of routable IPv4 hardly appear as buyers either, because their need is for clean, distributed, well-reputed addresses spread across many networks, which is a leasing need, not an ownership one. So the market splits cleanly: access networks and traditional hosting acquire IPv4, while AI compute, proxy, and data-collection demand is met by leasing.

Why prices crashed, and why they are turning

When the biggest buyers leave a market, the instinct is to assume prices were being propped up by their demand and should now hold elsewhere. The opposite happened, and it happened most at the top of the market.

Large blocks crashed. By IPv4.Global’s published figures, a /16 that peaked above $60 in 2021 and traded in the low $30s early in 2025, fell under $13 by the end of the year, a drop of roughly 60 percent, and large blocks fell so far they slipped below small and mid-sized ones. The reason is straightforward once you see who was bidding. The hyperscalers were the natural buyers of /16s, the only ones who routinely absorbed blocks that size. They were the top bid. When they paused, the bid disappeared, and the largest blocks had the furthest to fall. Their constant buying had been holding large-block prices up, not down.

That also explains the fragility. A market that leans on two or three buyers is exposed to exactly one event: those buyers pausing. It happened, and the price followed.

The recovery is the more durable part, and it is built differently. By early 2026, the slide had stopped, large-block pricing had begun to firm, and available large blocks were tightening. A market with thousands of active buyers competing for a shrinking supply gives sellers pricing power that two episodic whales never did. No single buyer pausing can crater it now. The floor is forming on a broader and steadier base than the one that broke.

Own or lease? The question just got sharper

Leasing moved too, but gently. On the IPXO platform, the average lease price sits around $0.35 per address today, below its four-year average of about $0.43 and down from roughly $0.60 in 2022. That easing tracked the purchase market lower, but it played out as a slow drift across four years, not a 60 percent crash compressed into one. There is no great migration from owning to leasing here. Lease rates are steady, and they have been.

What the crash and recovery change is: the math, and the math decides, not the use case. When prices were falling, owning looked like catching a falling knife. As large blocks firm and tighten, holding an asset that is appreciating again comes back into play. But leasing works either way, for a large steady need or a small spiky one. What moves the answer is the lease rate set against the purchase price.

At about $0.35 per month, leasing an address costs roughly $4 per year. Against a large block bought near today’s lows that pays for itself in three to four years, and against pricier small blocks closer to six or seven. So leasing is the cheaper path for anything shorter or less certain than that horizon, and for anyone who needs clean, distributed, well-reputed addresses across many networks rather than a block on their own books. Owning wins when the need is long and certain, the capital is there, and prices are stable or rising enough to beat the rent. The point is that this is a price decision, not a profile one. Today’s low lease rates make leasing attractive across a wide range of needs, and as purchase prices recover and pull rates up, the payback shortens and owning starts to compete again. The volatile year did not answer the own-versus-lease question. It made clear that the answer is a number, and the number keeps moving.

The whales did not kill IPv4 demand when they left. They revealed its real shape. Demand is broader than it has ever been, spread across thousands of buyers instead of a handful, split between the networks and hosts that own address space and the AI, proxy, and edge workloads that lease it. The market that priced off two or three buyers is gone. What replaced it is larger, more distributed, and harder to knock over.