Launch day roulette: how a startup can size its server budget with a probability

Every startup shipping a new feature faces the same quiet anxiety on launch day: what if it actually takes off? A founder posting to Hacker News or Product Hunt has no way of knowing, in advance, whether the launch will draw a trickle of curious users or a flood that melts the servers. The instinct is to either over-provision "just in case" and eat the cost, or wing it and hope. This is a case study in replacing that guess with a number.

The problem: two ways to be wrong

Picture a small AI tooling startup about to ship a new feature. The team can pre-book extra burst capacity for launch day — more compute headroom, a bigger database connection pool, a support engineer on standby — for a flat cost of €800. If they skip it and the launch stays quiet, they have lost nothing. But if they skip it and the launch goes viral, the site slows or crashes during the exact hours when the most new users are arriving. Some of those visitors never come back, some vent publicly, and the founders estimate the real cost of a launch-day outage — lost signups, refunds on annual plans sold at a launch discount, a bruised reputation — at around €6,000.

So the decision has the same shape as a retailer deciding how much stock to buy: one path has a small, certain cost, and the other has a larger cost that only lands in one specific scenario. Ordering Black Friday stock is the same trade-off with boxes instead of servers. Neither problem is solved by intuition — it is solved by attaching a probability to "will this go viral" and doing the arithmetic.

Putting a number on "will it go viral"

This is exactly the kind of question a prediction market is built to price. A prediction market works by asking a pool of participants to bet on a yes/no outcome, and the price that emerges is a crowd's best estimate of the probability. A startup does not need its own market for this — plenty of founders already watch odds on Product Hunt "featured" outcomes, competitor launch timing, or broader AI-adoption milestones, and use those prices as an input alongside their own judgment.

Say the founders land on a working estimate of 15% for "more than 5,000 signups in the first 24 hours" — informed by how similar launches in their space have performed, plus whatever market-based signal they can find on launch-day traction. That 15% is the number the rest of the decision hangs on.

Running the expected value

With a probability in hand, the two options can be compared directly.

Reserving the extra capacity costs €800, full stop, regardless of what happens. Skipping it costs €0 if the launch is quiet (85% of the time, by the estimate above) and €6,000 if it goes viral (15% of the time). The expected cost of skipping it is:

0.15 × €6,000 + 0.85 × €0 = €900

Reserving the capacity (€800) is cheaper in expectation than gambling on quiet traffic (€900) — and it also removes the tail risk entirely, which matters more to a small team than the average outcome does. If the estimated probability of virality were only 5% instead of 15%, the expected cost of skipping falls to €300, and skipping becomes the better bet. The point is not that reserving capacity is always right; it is that the right call flips depending on the probability, so getting that probability right is the whole game.

The same logic applies beyond launch day

This kind of calculation generalizes well past server capacity. Startups constantly make resourcing bets tied to uncertain external events: whether a major lab ships a competing feature before their own launch, whether a funding round closes before a runway deadline, whether an integration partner's API changes break a core workflow. Each of these is a binary-ish outcome with a probability attached, and each has a "reserve now" option that trades a small certain cost for protection against a larger uncertain one. Teams that write the probability down — rather than leaving it as an unspoken gut feeling — end up making the same kind of calculation, whether or not they call it that.

Practise the calculation before it costs you anything

The hard part of this method is rarely the arithmetic — it is getting comfortable putting a probability on things that feel unpredictable. Trading with virtual money on Outcomer is a low-stakes way to build that muscle: pick an upcoming event, form a view on how likely it is, and see how the market prices it against your own estimate. No signups crash and no real money is at risk — just a faster way to get better at the one skill this whole exercise depends on.