Hostinger Undercuts Hetzner 70% on the Same RTX PRO 6000: The Own-vs-Rent GPU Math

The same 96 GB Blackwell card currently rents for anywhere between $0.65 and $3.00 per GPU-hour depending on whose logo is on the invoice — a 4.6x spread for identical VRAM. The cheapest row belongs to a web host, the priciest to a CDN-turned-cloud, and the median across 51 tracked providers sits at roughly $2.20. If you run inference on rented GPUs, that dispersion is not trivia: it decides whether your fleet should rent by the hour or own the card outright.

Here are the numbers up front, all from GetDeploying's October 2026 tables for on-demand RTX PRO 6000 rentals:

Hostinger's row is 70% below Hetzner's for the same model name — and Hostinger-only rows list B200s at $4.50, RTX 5090s at $0.64, and A100s at $1.55 that Hetzner simply doesn't stock. Meanwhile Hetzner answered the rental market from the other direction: its new GEX63 dedicated server puts the same 96 GB card in a flat-rate box at €999/month (list price $1,199 in USD) plus a one-time €999 setup fee. So the real question is not which rental row is cheapest. It is at what utilization the flat-rate box beats each rental row — and whether the cheapest row will still be there when you scale.

The utilization table that decides rent vs own​

Assume a 730-hour month and an instance you hold only while you need it — Hostinger bills "for as long as you have the GPU," so spinning down to zero is allowed. Monthly cost is then hourly price times hours held, against the GEX63's flat $1,199:

Two conclusions fall out immediately. First, against Hetzner's own on-demand price the GEX63 pays for itself at about 558 hours a month — roughly 76% utilization, or 18 hours a day average. Against the median row the break-even is nearly identical at 545 hours (~75%), and against Akamai's $3.00 it drops to 400 hours (~55%). Anything serving inference around the clock beats every mainstream rental row on the flat-rate box, even before amortizing the setup fee (spread over a year it adds ~$100/month, pushing break-evens to ~83%, ~81%, and ~59% respectively).

Second, and stated honestly because the table demands it: the $0.65 row wins on price at every utilization level. Run it 24/7 and the month costs $474.50 — 60% below the GEX63. If your workload fits Hostinger's slice and the row is actually bookable when you need it, no flat-rate math beats it. The catch is everything in that sentence after "if."

Why the cheapest row is not the whole answer​

Start with availability. GetDeploying flags the $0.65 Hostinger row as not verified in stock, and names $0.66 from Packet·ai as the cheapest verified-in-stock on-demand row instead. Hostinger's own docs warn that GPU availability varies by region due to demand and tell you to check the dashboard for real-time availability before deploying. Hyperstack's $1.85 row — the next data point up the ladder — is marked sold out. The pattern across the table is that the cheapest rows are the least certain ones, which is exactly the wrong correlation for capacity you plan a fleet around.

Then there is catalog risk. A web host's GPU menu is a marketing surface that can reprice or delist rows a platform built its schedule around. The 90-day median trend is flat, but individual rows move: third-party trackers already list Hostinger's RTX PRO 6000 slice at $0.70 rather than the $0.65 list price. A fleet that arbitrages the cheapest row inherits every such twitch as a migration event, while the flat-rate box reprices on an annual contract cadence, not a dashboard refresh.

Finally, run the payback sanity check on the $0.65 itself. NVIDIA raised the RTX PRO 6000's retail price to $16,000 in August 2026, more than double its ~$7,600 launch price. At $0.65 an hour, a card has to bill 24,615 hours — 2.8 years of continuous, 100%-sold occupancy — just to gross its retail price. No host runs at full occupancy, so Hostinger either procured cards far below today's retail or prices the row as an acquisition funnel. Neither possibility makes the row fake, but both make it fragile: wholesale inventory runs out and funnels reprice. For context, the card's datacenter-class neighbor, the H100, still medians at $3.49 on-demand — the whole RTX PRO 6000 curve, even at its median, is the discount tier of high-VRAM inference.

Is it the same box, though?​

Not quite, and the bundle matters more than the GPU row suggests. The GEX63 is a dedicated machine: 20-core Intel Core Ultra 7 265, 128 GB of DDR5, two 1.92 TB NVMe drives, and a full IPv4 address in Hetzner's own European data centers. Trackers list Hostinger's slice around 8 vCPUs, 16 GB of RAM, and 200 GB of disk. For pure token inference — model resident in VRAM, requests in and out over an API — the host bundle is mostly irrelevant, and the slice is genuinely sufficient. The moment the workload needs a large local model library, CPU-side pre- and post-processing, datasets on fast disk, or RAM for offloaded context, you are no longer pricing the GPU row; you are pricing the host around it, and the comparison shifts toward the dedicated box.

One genuine specification caveat: Hetzner's card is the Blackwell Max-Q bin, while Hostinger lists a "Server" variant. Both are 96 GB RTX PRO 6000s and every price tracker treats them as the same model, but bins can differ in sustained clocks. For memory-bound inference serving — which is most serving — VRAM capacity dominates and clocks are second-order, so this should not decide the purchase. It is worth one line in your evaluation notes, not a veto.

Egress deserves a quantified dismissal rather than a hand-wave. Hostinger charges no egress fees on GPU instances; Hetzner's cloud side meters egress past generous allowances at around €1/TB. A billion output tokens is roughly 4 GB of text, so even a busy inference endpoint moves single-digit gigabytes a month and the meter reads in cents. Egress only enters the math for image or video generation at real volume — and there Hostinger's zero-meter is a genuine second discount on top of the hourly rate.

The rest of the menu matters too​

The single-card spread understates how different the two vendors' GPU businesses are. Hostinger-only rows list B200s at $4.50, RTX 5090s at $0.64, A100s at $1.55, plus RTX 4090s and L40S — Hetzner stocks none of them, answering only with the entry-level RTX 4000 SFF Ada at $0.42 that Hostinger lacks. A fleet that standardizes on one vendor's cheapest row inherits that vendor's catalog gaps the day it needs a second card type: the B200 you want for a big-context serving tier or the 5090 you want for cheap batch work may mean a second provider, a second billing relationship, and a second operations surface.

There is also a third way between on-demand rental and the flat-rate box: committed reservations. HyperAI lists an in-stock one-month reserved RTX PRO 6000 row at $0.48 an hour — $350.40 for a full month, below even Hostinger's 24/7 on-demand cost and less than a third of the GEX63. The trade is the obvious one: reserved hours behave like a mini flat rate, paid whether the card works at 3am or not, and the commitment sits with a niche provider rather than hardware in your own rack or account. For a known-steady, single-month burst — a training run with a fixed end date — it is the sharpest price on the table. For capacity you will still need in month thirteen, it just restarts the same rent-vs-own question with a different counterparty.

Verdict: price the workload shape, not the row​

The decision rule falls out of the utilization table, not the headline spread. Spiky, batchable, or dev/test GPU work — nightly fine-tunes, eval harnesses, burst rendering — belongs on the cheapest bookable rental row, because the flat-rate box charges you for idle VRAM at 3am while the metered instance spins to zero. Always-on inference with steady traffic belongs on the flat-rate box the moment utilization clears ~75% against mainstream rental rows, which is most of what "production serving" means.

And the dispersion itself is the argument for owning the card. When identical VRAM rents from $0.65 to $3.00 across providers — with the cheapest row unverified in stock, the next one sold out, and the card retailing at $16,000 — the rental market is telling you it has no consensus on what the capacity is worth. Building a fleet schedule on the bottom row of a table that disagrees with itself by 4.6x is not optimization; it is a bet that the row you built on will still be listed, stocked, and priced when you need your tenth card. The GEX63's pitch was never that €999 beats $0.65 an hour on a spreadsheet. It is that the spreadsheet row can't be delisted.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.