As of this morning the four-euro plan is gone from the configurator. It will not come back, and nothing will replace it at that price. Existing instances keep running at their current price for twelve months, after which we will help those customers move up a tier at a discount.
This is the decision we have argued about internally for longer than any other, so the numbers that settled it are below in full.
The numbers
Twelve months to the end of July, across the whole fleet.
| Measure | The four-euro plan |
|---|---|
| Share of monthly revenue | 4.2% |
| Share of running instances | 37% |
| Share of support tickets | About 90% |
| Share of abuse reports | Every single one |
| Median instance lifetime | 41 days |
| Gross margin after support cost | Negative, about minus eleven percent |
| Instances per engineer-hour of support | Roughly one hour of attention per instance per quarter |
Read the first two rows together. Thirty-seven percent of everything we operated produced four percent of what we were paid. Then read the third and fourth rows, which is where the argument actually ended.
What the tickets were about
Not hardware. The nodes running cheap plans were the same nodes running everything else, and they failed at the same rate as any others.
The tickets were about the consequences of selling infrastructure for less than the price of a sandwich: instances suspended for outbound scanning, mail deliverability on addresses inherited from a previous tenant, people who had bought a server without knowing what one was, and a steady trickle of accounts created to be disposable. Every one of those is a legitimate ticket. Answering them well takes as long as answering a ticket from an EPYC customer, and there were nine times more of them.
What the abuse reports were about
Every abuse report we had received to that date concerned an instance on the cheap tier. Not most. All of them.
That is not a coincidence and it is not a moral judgement about people with small budgets. A four-euro server is disposable by construction. If your plan involves an IP address that will be burned within a fortnight, you buy the cheapest thing available, and no amount of terms-of-service language changes that arithmetic.
What happened next
Revenue fell about nine percent in the first quarter after the change and had recovered by the second. Abuse reports fell by roughly ninety percent within two months. The support queue emptied to the point where median first response went from a little over forty minutes to eleven, where it has stayed.
The unexpected result was on the sales side. Removing the entry-level plan made the rest of the catalogue legible. A customer looking at a page whose cheapest option is twenty-nine euro understands what is being sold; the same page with a four-euro plan at the top reads as a discount host that also has expensive plans, and people price the whole company accordingly.
The part that still bothers us
We lost customers who were perfectly good and simply poor. Students, hobbyists, people running a personal mail server or a small project on a budget that was real to them. Several wrote to say so, politely, and they were right that nothing about their behaviour caused any of the problems above.
We do not have an answer for them. A cheap tier for the well-behaved is not a product you can build, because the abuse arrives through the same price point rather than through a personality test. What we could do is not pretend otherwise, so: if you were one of those customers, the reason you lost your server was other people, and we are aware that is an unsatisfying explanation.
What this commits us to
Everything we sell from now on is high-end hardware at a price that pays for the support behind it. No entry tier, no promotional pricing that quietly reintroduces one, and no shared-core plan with a fraction of a CPU sold as a whole one.
If we ever reverse this, the reversal will be a post in this journal with numbers as specific as the ones above.