Sold Out, Not Priced Out: Nine Months After the RAMpocalypse, You Can't Buy the Server at Any Price

SudoSecurity runs on an OVH VPS. Four vCPU cores, 8GB of RAM, 75GB of NVMe, ten dollars a month.

That plan still costs ten dollars a month.

I just can't buy another one.

That's the thing I didn't see coming when I wrote The RAMpocalypse back in February. I spent that whole article on price — who raised what, by how much, and why AI memory demand was behind it. All of which held up. What I got wrong was assuming price was the shape the problem would keep.

It isn't. Nine months on, the sharp edge isn't the invoice. It's that the thing you want to buy is not for sale, and the moment you try to grow, you get repriced.

Quick update: February's numbers are now wrong, and badly

If you read the February piece, throw out the Hetzner table. I had the April wave at 25–43%.

There were two more waves after that.

The June numbers, in euros per month for Germany and Finland:

A CPX52 went from thirty-six euros to a hundred. That's not an adjustment, that's a different product category.

And look at the Arm line while you're here: CAX11 went €4.49 → €5.99, about +33%. Hold onto that, because it turns out to matter a lot.

The part I actually want to talk about

Here's what happened to me, and I suspect to a lot of you.

My plan didn't go up. It's the same ten dollars it's always been. But when I went looking to add a second one — different region, start building out the thing I actually want SudoSecurity to be — there was nothing there.

Same price. No stock.

For a while I assumed I was doing something wrong, or looking at the wrong hour of the day. I wasn't. This is now a documented operational state, and somebody has written the incident-response guide for it.

Bex's write-up covers Hetzner's side, and the framing is exactly right:

Capacity is not a second hike. It is a third state your autoscaler must handle.

Hetzner's status page has run an incident titled "Limited availability of Cloud plans." Their Cloud API returns "no capacity in this location" as its own distinct error — not a quota error, not a billing error, a separate class meaning the hardware does not exist to sell you.

If you're running Cluster API Provider Hetzner, the failure cascade looks like this:

HCloudMachine   Ready: False      <- limited-availability / no-capacity condition
Machine         Provisioning      <- FailedCreate
Pod             Pending           <- never schedules
cluster-autoscaler: scale-up failed   (NOT "Insufficient cpu")

That last line is the one that'll cost you an afternoon. You go looking for a resource problem, because that's what "pods won't schedule" has meant for your entire career, and the actual answer is that the shop is out of servers. Bex puts the wait at 10–60 minutes until stock returns, for a single pool.

cpx22 sold out in hel1. Then in nbg1.

So if you've built anything that assumes it can buy a machine on demand — autoscaling, blue-green deploys, spinning up a staging environment — you now have a failure mode where the correct code does the correct thing and gets told no.

The dedicated market did something worse

Everything above is cloud and VPS. The dedicated side is where this gets ugly, and it's where my own hunting has gone the most sideways.

A year ago I could get an i9-9900K with 128GB of RAM and 2TB of SSD through Hetzner's Server Auction for somewhere between $50 and $80. Looking today, I'm seeing an E3-1650v3 — a Haswell-era Xeon from 2014, four years older and several generations behind — sitting close to $150.

I want to be straight about the status of those numbers: that's my own shopping, today, not a figure I can point you at a chart for. I tried to pull live auction listings through a public API to check it properly and the request didn't get through from where I was working. So treat it as one operator's observation rather than market data.

But the mechanism behind it is documented, and once you see it the observation stops being surprising.

The auction didn't get a price increase. It got repriced anyway.

Here's the strange part. Hetzner's Server Auction took about +3% in April, and was explicitly excluded from the June restructure that put CPX and CCX up 113–176%.

So the list price barely moved. The market behind it moved instead.

Server Radar, which snapshots the auction every few minutes and keeps three months of history, puts it like this: server DRAM contract prices rose over 40% in late 2025 and kept climbing through 2026, affecting memory-heavy configurations most. And refurbished machines with a lot of RAM have become relatively better value, with sought-after configurations selling faster.

Now put that together with how the auction actually works. It's a Dutch auction — the longer a box sits unsold, the cheaper it gets. That's the whole appeal. You wait, and you catch something good on the way down.

Which means a configuration that sells fast never reaches the bottom of its slide. The 128GB boxes aren't necessarily priced at €150. They're gone at €150, before anyone gets to see €80.

That's not a price increase. That's a queue, and you're not at the front of it.

For reference, recent auction pricing around September had an AX41-NVMe — Ryzen 5 3600, 64GB DDR4, 2×512GB NVMe — at about €37.30, and an AX42 with 64GB of DDR5 ECC at about €46. Sixty-four gigabytes is the comfortable configuration now. One-twenty-eight is the one you fight for.

And the floating is real even at the low end: the same source warns that "today's €46 box can be €52 or €41 next week." Auction stock is also Germany and Finland only, which matters a great deal if your plan involved being anywhere else.

The uncomfortable thought I keep having

I have no evidence for this and I want to flag it as speculation rather than reporting. But when server DRAM contract prices climb past 40% and keep going, and the secondhand market for DDR4 ECC gets hot, at some point the 128GB of RAM inside an eight-year-old box is worth more as 128GB of RAM than as part of an eight-year-old box.

I don't know that anyone's stripping machines. I'm saying the arithmetic that would make it rational now exists, and it didn't two years ago.

If you're enjoying what you've read so far, we recommend Oracle Free Tier ARM Cutoff: What to Do Before Aug 18 as your next read.

Why the cheap tier evaporated

This is the bit that clicked for me when I read OVHcloud's own announcement, and it's not a conspiracy. It's an entirely reasonable policy that had a predictable second-order effect.

OVHcloud let customers renew at existing rates for up to two years before the April 2026 increase landed. Lock in now, pay old prices until 2028.

That's a generous thing to do for existing customers. I'd have done the same. And every single person who was paying attention did exactly what you'd expect: locked in, for the maximum term, on every box they had.

Which means the cheap inventory is now spoken for until 2028 by people who already had it. The price on the website is real. The availability behind it is not, because the stock is committed.

That's why my ten-dollar plan is simultaneously still ten dollars and impossible to buy. Both things are true and they're the same fact.

And you can't grow into it either

Hetzner's June change applies to new orders and rescales. Existing servers keep their pricing.

Read that again if you run anything that grows. Rescales.

You are grandfathered exactly as long as you never need a bigger box. The moment your traffic doubles and you resize, you're not on your old price any more — you're on today's, which for a CPX is up to 176% higher. The grandfather clause is a cage with a very polite door.

I don't think that's malicious. If the hardware under a rescale costs what it costs now, they can't sell it at 2024 prices. But the practical effect for someone trying to build something is that success triggers a repricing event, and you need to model that before it happens rather than during.

What's actually driving it, nine months on

The February thesis was right and I'm not going to re-litigate it. But the data got sharper, and OVHcloud now says it in their own words rather than leaving us to infer:

Since autumn 2025, the global memory market has been going through a major disruption.

Memory makers, they say, "redirected a significant portion of their production capacity to meet the massive demand for GPUs." And then the line that matters:

This reallocation took place without a corresponding reduction in the historical demand for RAM and storage.

Nobody stopped needing normal server RAM. The supply just left.

OVHcloud's own planning assumption is that RAM and storage costs "could increase by 15% to 300% compared to 2025 prices," and they expect no stabilisation until late 2026, with prices possibly not returning to earlier levels before 2028. They also note they order components up to 12 months ahead, without knowing the final price at purchase, which is a horrible way to run a business.

The market data from TrendForce backs the trajectory:

That RDIMM row is the whole problem in one line. Bit supply growing 15–20% a year, which TrendForce calls "significantly below" expected server CPU shipment growth. The CPUs are coming. The memory to put next to them isn't.

Some servers have already started shipping with 32GB and 64GB modules where they used to use 96GB and 128GB. The industry is quietly downgrading.

The inequality nobody designed but everybody feels

Here's the structural bit, and it's the part that actually annoys me.

TrendForce notes that several US cloud providers hold multi-year long-term agreements that cap what suppliers can charge them — so future increases "fall mainly on buyers without LTAs."

Omdia's Vlad Galabov, quoted in The Register, puts the same thing plainly: smaller providers without large purchasing power, Hetzner and OVHcloud included, have to pass costs through. Hyperscalers like AWS, Microsoft and Google are insulated by long-term direct vendor contracts.

So the companies whose AI buildout created the shortage are the companies least exposed to it.

The gap between "cheap European VPS" and "hyperscaler cloud" has been the single most useful fact in independent hosting for a decade. It's what made it possible to run a real service on ten dollars a month. That gap is closing — and it's closing from the wrong end. Not because AWS got cheap. Because the cheap option got expensive while AWS held its contracts.

Hetzner's own statement was short and, I thought, honest: "There have been drastic price increases in various areas in the IT sector recently. The costs to operate our infrastructure and to buy new hardware have both increased dramatically."

netcup's forum post was blunter: "The memory market is being bought out by large AI corporations."

And if you wanted multiple regions, read this first

This is the one that actually killed my plan, and I only found it while researching this piece.

Hetzner's included outbound traffic is not the same everywhere. It is not even close.

The US allowance is twenty times smaller than the EU's. Singapore's is forty times smaller, at more than seven times the overage rate.

Work through what that does to a small node. On a roughly €6/month server, per Bex's numbers: 25TB in the EU costs you €10.99 all-in. The same 25TB in the US is €29.99. And 5TB — five — in Singapore is €39.29, which as they put it "exceeds the server's own price more than fivefold."

Inbound and private-network traffic is free everywhere, and most ordinary app servers never come close to their allowance. The exposure is on bandwidth-heavy things: container registries, media origins, backup and artifact stores.

But look at what that pricing is telling you to do. Stay in Europe. The economics of "put a node near your readers in Asia" are not subtly worse, they're an order of magnitude worse, and that's before you find out whether the SKU you want is in stock there.

I wanted multiple regions so readers aren't all hitting one box in one datacentre. The bandwidth maths says the cheap way to do that is a CDN in front of one origin, not nodes in three continents — which is probably the right answer anyway, but I'd have liked to arrive at it by choice rather than by having the alternative priced out of reach.

What this doesn't cover

I'm describing my own experience with OVH and the publicly reported situation at Hetzner, netcup and Scaleway. I haven't audited every provider, and I can't tell you what's in stock in your region tonight — availability moves, and anything I write here about specific SKUs will be wrong within weeks.

I also can't quantify how much of a VPS price is memory. Nobody publishes that. I can tell you memory costs rose dramatically and prices followed, but the exact pass-through ratio — and therefore whether any given provider used this as cover for margin expansion — is not something I can prove, and I'm not going to insinuate it without numbers.

And one genuine complication to my own argument: Contabo appears not to have raised prices at all. They relaunched their lineup in May 2025, and reporting from this year still shows VPS plans in the sub-$8 range. The widely-cited Contabo "+15% increase" you'll find online is from October 2022 and was about European energy prices, not AI memory demand. I nearly cited it myself before checking the date.

I don't know how Contabo is absorbing this when nobody else is. Older hardware, different margin structure, a willingness to eat it for market share — pick your theory. But a thesis that says "everyone had to raise prices" has to account for someone who didn't.

What I'd actually do

Go Arm if your workload allows it. This is the single biggest lever and almost nobody talks about it. Hetzner's CAX line went up 31–33%. The x86 CPX line went up 143–175%. That is a five-fold difference in how hard you got hit for running the same workload on a different instruction set. Bex also notes CAX had better availability through the first half of 2026. If your stack is containerised and your dependencies have Arm builds, you're leaving a lot of money on the table.

Build multi-arch images now, before you need the fallback.

Treat capacity as a state your tooling handles. Separate node pools per location and server type, never one pool. Use the priority expander rather than random, with an Arm pool as fallback. Shorten your retries to 30–60 seconds and let it fail over, instead of hammering one sold-out SKU for an hour. If you're on Hetzner specifically, datacenter is gone — use location.

Model the rescale. Write down what your bill becomes at current prices when you outgrow your current box, because that's the real number, not what you pay today. If you're grandfathered, you're grandfathered until the day you succeed.

Consider whether it needs to be in a datacentre at all. The mini PC maths has flipped. A ~$630 box against a €28/month VPS pays for itself in about two and a half years. Do check that against your own power price, because the naive division is misleading — $630 ÷ $28 is 22 months, but the box draws power, so it's really $630 ÷ ($28 − running cost). At roughly $6/month for electricity that's $630 ÷ $22 ≈ 29 months. Electricity runs about $2–8/month at $0.20/kWh, and a UPS adds maybe $60 once. For CI runners, staging, monitoring, dashboards and backups — things with no SLA and no public traffic — the cloud premium stopped making sense a while ago. Keep a cheap VPS as a public edge and tunnel back.

That's not the right answer for everything. Revenue-generating public services, anything needing an SLA, bandwidth-heavy workloads, DDoS exposure, genuine multi-region — those still want real hosting. But the honest split is smaller than most of us run.

And check the free tiers. I've written up Oracle Cloud's free tier before, and in a market like this one it deserves a second look, Arm instances included.

The second cliff, which is the one that worries me more

Everything so far is about SudoSecurity, and a publication is a light tenant. Mostly it serves text.

The thing that actually keeps me up is the other stuff I'm building.

I run Tower of Elements, a browser-based idle game, with a card mode and a tactics mode alongside it. Right now the hosting is nothing. It's a rounding error. That's the whole joy of a side project — the cost of having it is basically zero, so it can sit there being half-finished for a year without anybody having to justify it.

But idle games are stateful. Every player has a world ticking away whether they're looking at it or not, and the backend holds sessions, game state and open connections. That's not CPU-bound work. It's memory-bound work, and memory is the exact thing that got expensive.

So there's a threshold out there where it stops fitting on whatever it's on now and needs a real box. And crossing that threshold used to be a shrug — fifty, eighty dollars, fine, that's a tolerable hobby expense for something people actually play.

At current prices I'm not sure what that number is any more. The 128GB machine I'd have reached for is the configuration that vanished. And it arrives as a new recurring cost, at the moment the project gets popular, before the project makes any money — which is precisely the worst possible ordering.

That's the part people miss when they talk about this as a price story. It isn't really that hosting got more expensive. It's that the cost of succeeding a little bit went up, and it went up at the stage where you have the least slack.

I think a lot of small projects are going to quietly not cross that line. Not because they failed. Because the next rung of the ladder moved, and the person holding them up looked at the number and decided to cap the player count instead.

The bit that actually bothers me

I can afford ten dollars. That's not the problem. If OVH doubled my bill tomorrow I'd grumble and pay it, the same way I'd grumble and pay a doubled electricity bill.

What I can't do is buy the second box.

SudoSecurity is supposed to grow. Multiple regions, so readers aren't all hitting one box in one datacentre. Redundancy, so a single provider incident isn't an outage. The ordinary, unglamorous infrastructure that turns a project into a thing. All of that assumed I could buy small servers, cheaply, whenever I wanted — and that assumption was so reliable for so long that none of us ever wrote it down as an assumption.

It isn't true any more. The entry-level tier that let a person with ten dollars and an idea put something real on the internet is the exact tier that got cannibalised, because it's the tier with the thinnest margin and the least pricing power behind it.

The trillion-dollar AI buildout isn't taking anything from me directly. It's just that I'm bidding against it for DRAM, and I'm bidding with ten dollars.

I'll keep running on the box I've got. It's grandfathered, it works, and I'm not resizing it unless something forces my hand. But "don't grow, you'll lose your price" is a hell of a thing to have as an infrastructure strategy, and I don't think I'm the only person who's quietly ended up with it.

If you've found capacity somewhere, or you've moved something to Arm and it went fine — I'd like to hear about it.