OVHcloud and Hetzner Both Raised Prices in 2026: What the Budget EU Squeeze Means for Your Self-Hosted Fleet

Your favorite cheap EU server just got less cheap — twice — and the other one just told you it will keep getting more expensive every year through 2028.

In early 2026, Hetzner and OVHcloud, the two names that anchor nearly every "just self-host it on a budget EU box" cost comparison, both repriced. Hetzner did it in two sharp moves: roughly 30–37% across cloud lines on April 1, then another 107–204% on its popular CCX and CPX families on June 15 for new orders and rescales. OVHcloud did it once but with a long tail: 9–11% per year for any Public Cloud, Private Cloud, or Bare Metal deployment between 2026 and 2028, plus 2–6% on pre-2025 hardware, all effective April 1, 2026. Same root cause, two very different pricing shapes — and together they turn a frozen baseline that half the self-hosting internet hard-codes into a moving target.

Here is the recomputed math you actually need. For a representative 20-node fleet of CCX13-class machines (2 dedicated vCPU, 8 GB RAM — the median production node, not the cheapest CX you benchmark on a weekend), the 12-month run rate looks like this before you change anything else:

A Hetzner fleet you do not rescale still looks cheap — until you need a 21st node or resize a 20th one. An OVHcloud fleet looks predictable — but predictable at a slope that compounds. Neither is the flat line your spreadsheet still has.

This post does three things: names the shared DRAM shock driving both moves, recomputes a real Cluster API fleet at the new numbers across a 12/24/36-month horizon, and prices what diversifying a node pool across providers actually costs versus hoping one vendor's price holds.

The Same Shock Hit Both Providers​

It is tempting to read two price hikes as two business decisions. It was one commodity shock, passed through at two different speeds.

TrendForce called Q1 2026 a seller-driven DRAM super-cycle. Contract prices for server DDR5 rose 90–95% quarter over quarter in Q1 and another 58–63% in Q2. Spot prices for some server-grade modules jumped more than 400%. HP disclosed that DRAM went from 15–18% of a PC build cost to 35% in a single quarter. UBS tracked automotive DRAM up 180% in three months. The culprit is not mysterious: AI servers and high-bandwidth memory (HBM) now consume roughly 40% of world DRAM output, with some estimates pushing that toward 70% for high-end production, and the capacity was physically reallocated — fabs that used to run commodity DDR5 now run HBM stacks.

OVHcloud was explicit about it. Founder Octave Klaba published a pricing memo on March 5, 2026 and a longer landing page that frames the 9–11% path as deliberate absorption: OVHcloud projects RAM pricing up 250–300% by the end of 2026 versus September 2025, plus rising NVMe and disk costs, but chose not to pass the full increase through. For services deployed between 2026 and 2028 the average lands at 9–11%, and for equipment deployed before 2025 it is only 2–6% depending on age. The company also says it stockpiled memory and storage to delay the hit — the same playbook that delayed rather than prevented the increase.

Hetzner told the same story louder. Its April 1 adjustment cited increased infrastructure and hardware costs across Germany, Finland, the US, and Singapore — cloud servers up 30–37%, US cloud up 38–40%, dedicated AX42 from €47.30 to €57.30, object storage base from €4.99 to €6.49. The June 15 standardization round was more surgical: the dedicated-vCPU CCX and shared-vCPU CPX AMD lines, the families most production teams actually run, absorbed the biggest DRAM exposure. Hetzner's own tables showed June alone adding 107–204% on top of already-elevated April prices for those families, while the cost-optimized CX and ARM-based CAX lines rose a more modest 30–38%. A CCX23 (4 dedicated vCPU, 16 GB) that cost $39.99 before April was listed at $102.99 after June. Hetzner grandfathered existing cloud instances until a rescale or new order, which is why a legacy fleet still prints the old number — and why every growth or migration event reprices at the new one.

Two important dynamics sit behind both:

• This is not cyclical. IDC and EE Times commentary around the same reports describes the reallocation as structural, not a 12-month blip. Foundries invested in HBM capacity that does not swing back to commodity DDR5 quickly. Forecast windows that predicted relief in late 2026 have slipped to late 2027, and even then at higher absolutes.
• Stockpiling buys quarters, not immunity. OVHcloud's memo and Hetzner's setup-fee cuts both advertise mitigation, not avoidance. When the buffer thins, the next increase lands on whatever hardware is queued behind it.

In short: both vendors are translating the same memory super-cycle into your invoice. OVHcloud smoothed it into a published escalator. Hetzner left it as two ad-hoc steps with no forward cap stated. Your fleet plan should treat them as correlated variables, not independent bets.

Fleet Math Recomputed: What a Real Cluster API Fleet Pays Now​

Take a concrete fleet. Twenty nodes of CCX13-class capacity (2 dedicated vCPU, 8 GB RAM, 80 GB SSD) is a believable small platform: enough for a control plane, monitoring, a build pool, and a dozen tenant workloads with headroom. Pre-2026, a CCX13 ran €11.99 per month — so the raw compute line was €239.80 per month before tax, load balancers, storage, and snapshots.

Here is that fleet repriced at each moment, holding everything else constant. The point is not the absolute euro total — your real bill has egress, volumes, and snapshots — but the delta and what triggers it.

Hetzner: still cheap if you do not touch it​

• Pre-April 2026: 20 × €11.99 = €239.80 per month, €2,877.60 per year.
• After April 1 (every order): CCX13 moved to about €15.99. Fleet = €319.80 per month, €3,837.60 per year, +33.4%. Every node, old or new, repriced.
• After June 15 (new and rescale only): The CCX line repriced again steeply. Depending on region and whether you catch a "-Ltd" limited tier, a CCX13-class replacement now lists between roughly €32 and €36.50 per month. A fleet that has to add or rescale nodes therefore jumps to €640–€730 per month, €7,680–€8,760 per year — 2.7 to 3.0× the pre-2026 baseline. Nodes you never touch keep their April price until you resize them, which is exactly why this hike hides in the average until growth forces it visible.

Add a realistic 21st node in Q3 2026 and the story gets sharper: that one node costs 2.5× what the first twenty average. Average fleet price becomes a function of churn, not just count.

OVHcloud: predictable slope, compounding pain​

OVHcloud's path is simpler to model because it is published as a schedule. For the same 20-node capacity on VPS 2026 or Public Cloud instances:

• If deployed before 2025: year-1 increase 2–6% depending on hardware age. Fleet = €245–€254 per month in 2026, barely noticeable.
• If deployed in 2026–2028: year-1 increase 9–11%. Assume a comparable €12 per node starting point → fleet = €262–€266 per month in 2026 (+9–11%), €286–€295 in 2027, €312–€328 in 2028 at compounding, for a three-year total of roughly €10,300–€10,700 versus €8,640 at a flat baseline — a €1,660–€2,060 escalator you should have in the budget now, not discovered in the renewal email.
• VPS specifics matter. OVHcloud lifted entry VPS plans single-digit to ~30% depending on tier, with the new VPS 2026 range starting at $6.46 per month — above Hetzner's CX22 at €3.79 but below Hetzner's post-June CCX pricing. The gap between the two EU budget providers narrowed materially.

What the three-year comparison actually says​

Put the two paths side by side for the same capacity growth of 10% per year (22 nodes in 2027, 24 in 2028):

*2026 blended: most nodes at April price, growth nodes at June price.

Two takeaways survive any quibble about the exact instance match:

1. Not touching Hetzner is the cheapest way to stay on Hetzner — which is a strange property for infrastructure. It rewards stasis and taxes growth.
2. OVHcloud's published escalator is easier to budget even though it is not the cheapest year-one number. A CFO would rather model 9–11% compounding than explain why a Q3 scale event tripled the unit cost on a slide.

Either way, the era of pasting a single Hetzner row into a cost-vs-Render table and calling the comparison done is over. The comparison needs a time axis now.

Diversification: What a Second Provider Actually Costs​

If the "cheap EU box" premise just became a moving target, the obvious hedge is a second provider. The less obvious part is what that hedge costs to stand up and run.

What you get from adding OVHcloud alongside Hetzner​

OVHcloud compensates for a higher entry price with properties Hetzner does not try to match:

• Breadth. More than 30 global locations across Europe, North America, and Asia-Pacific versus Hetzner's concentrated footprint (Nuremberg, Falkenstein, Helsinki, plus Ashburn and Hillsboro in the US and Singapore). For workload placement and latency diversity, that is not a marginal difference.
• Vertical integration. OVHcloud manufactures its own servers, owns its data centers, and runs its own fiber network. Hetzner assembles and operates at high quality but at smaller scale and fewer sites.
• Operational guarantees that matter to regulated buyers. SecNumCloud 3.2 qualification on the Bare Metal Pod (360+ technical, organizational, and legal controls vetted by ANSSI), SOC 2 Type II and HIPAA posture in US regions, always-on anti-DDoS at terabit scale with no add-on fee. Hetzner offers solid DDoS and ISO 27001 but not the same French sovereign-cloud certification.
• Billing shape. No egress fees on most OVHcloud plans — structurally different from per-GB egress that metered PaaS vendors keep expanding — and now a published 2026–2028 escalator instead of ad-hoc steps.

What you pay to actually run two providers​

This is where Cluster API semantics matter.

• Hetzner has CAPH. The Cluster API Provider for Hetzner is mature, maintained, and already the substrate for fleets like Bex.co's. Declarative Machine and MachineDeployment objects, automated node provisioning, rolling upgrades — it works.
• OVHcloud does not have an equivalent mature CAPI provider. The OVHcloud API is capable and well-documented for Public Cloud and Bare Metal, but a production-grade cluster-api-provider-ovhcloud that matches CAPH's ergonomics would need to be built or adopted from a community effort still finding its footing. That is not a weekend project. It is weeks of provider development, testing against OVHcloud's rate limits and availability-zone model, and ongoing maintenance as OVHcloud's own 2026 portfolio changes flow through.
• The pragmatic alternatives each carry a different tax: UpCloud (Finnish, €3 per month Starter plans in May 2026, data centers in London, Frankfurt, Amsterdam, Warsaw, Madrid, Helsinki) benchmarks storage competitively and has a clean API, but also lacks a turnkey CAPI provider — same build cost, different hardware supply chain. Equinix Metal would have been a natural second source for bare-metal diversity, but its shutdown pushes that capacity toward OVHcloud Bare Metal itself, which collapses the diversification you were trying to buy.

So price the hedge honestly:

None of that is a reason not to diversify. It is a reason not to pretend diversification is free or that quoting an OVHcloud list price next to a Hetzner list price is the analysis. The real hedge ratio is a fraction of your fleet — say 30% on the second provider — enough to keep provisioning velocity when one provider throttles or reprices, without doubling your operational surface on day one.

For a platform team that already runs on Cluster API, the cheapest near-term diversification is often not a second cloud at all: it is a bare-metal slice next to your cloud slice on the same provider, with different failure and pricing exposure. The second-cloud provider is the next step once the bare-metal/cloud split is already automated.

Operating a Fleet When the Floor Keeps Moving​

You cannot control DRAM futures. You can control how exposed your fleet is to them and how honestly you budget.

Pin the instance families that moved least. Hetzner repriced CCX and CPX hardest because dedicated and AMD-shared nodes carry the most memory and NVMe per vCPU. CX (Intel shared) and CAX (ARM shared) rose 30–38% — painful, but not 150%. If a workload runs well on ARM or tolerates shared cores, it saw the smaller hike. Audit every deployment for "CCX by habit" and right-size the ones that do not need dedicated cores.

Treat bare metal and cloud as a mix, not a religion. A dedicated AX or Bare Metal server amortizes the same DRAM shock differently than a cloud slice — the price is higher up front but the per-GB-RAM cost and the noisy-neighbor profile are different. For steady-state tenant apps, a bare-metal pool under Cluster API can be the cheaper, more predictable host. For burst or spiky workloads, cloud elasticity still wins. The right answer in 2026 is probably both, with autoscaling that prefers the cheaper pool.

Budget an escalator, not a point estimate. OVHcloud gave you the template: model 9–11% annually for new capacity through 2028 and 2–6% on legacy hardware. Even if you stay on Hetzner, adopt the same discipline — put an explicit escalator in the forecast and name the trigger (rescale, new MachineDeployment, storage-class change) that fires it. A forecast that says "€320 per month plus 10% annual for growth nodes" will survive contact with reality. One that says "€240 forever" will not.

Negotiate on commitment, not on spot. Both vendors reward commitment and punished spot-like assumptions in 2026. Hetzner grandfathered existing instances and cut setup fees to soften the monthly increase. OVHcloud banded the increase by deployment age — older commitments moved least. If you can commit to a bare-metal shape for 12 months, commit. The cheapest node in 2027 will be the one you already reserved.

Recompute the PaaS comparison at the new numbers. Even after Hetzner's hikes, a flat-rate Hetzner or OVHcloud box remains 3–5× cheaper than equivalent always-on capacity on metered PaaS platforms that also repriced in 2026 — Render's April workspace and egress changes, Vercel's fourth repricing since 2024, Fly.io's continued billing-line expansion. The self-hosting cost advantage did not disappear; it compressed and became conditional on growth and instance choice. Update the table that your migration pitch actually uses, not the one from a 2024 blog post.

The broader lesson is not that budget EU providers failed. It is that they stopped being a fixed baseline you can anchor a multi-year pitch to. They are now what every other infrastructure input is: a market price with a slope. A platform that owns the machines still wins on control, egress, and the ability to choose where the slope hits — but only if it treats the provider as a variable in the model.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Own the fleet, keep the PaaS ergonomics, and model the floor as it moves. Star the repo on GitHub or deploy your first app today.