A database vendor removes its free tier to control cost. How should a platform team model the true cost of serving many tiny tenants, and which architectural choices change the economics?

Why tiny tenants are disproportionately expensive

The intuition is that a database with almost no data and almost no traffic costs almost nothing. That is only true if the cost is proportional to usage, and for a managed database it largely is not.

The costs that do not scale down:

• Baseline infrastructure per tenant. A dedicated instance, however small, has a floor: memory reservation, storage minimum, network attachment, monitoring agents. An idle database is not a free database.
• Control-plane cost per tenant. Provisioning, health checking, metrics collection, backup scheduling, certificate management, upgrade orchestration. This is per-tenant, not per-request, and at large tenant counts it dominates.
• Backups and snapshots, which have a floor per tenant and are retained regardless of activity.
• Upgrades and migrations, where the engineering effort scales with the number of instances, not their size — a fleet-wide version upgrade costs the same per tenant whether the tenant is enormous or empty.
• Support burden, which correlates poorly with revenue. Free and small tenants generate support volume at a rate that is not lower than paying tenants, and often higher.
• Abuse and fraud handling, which is concentrated almost entirely in free tiers.
• The long tail of abandoned resources, retained indefinitely because nobody deleted them.

PlanetScale's 2024 removal of its free tier — with the stated reasoning that supporting it consumed engineering and infrastructure resources disproportionate to its value — is a concrete instance of this arithmetic concluding against the tier.

How to model it properly

Compute fully-loaded cost per tenant per month, including:

infrastructure floor + amortised control-plane cost + backup floor + (support tickets × cost per ticket) + (engineering time on fleet operations ÷ tenant count) + abuse handling + the cost of resources retained after abandonment.

Then segment by tenant size and plot cost against revenue. The characteristic finding is a large population of tenants whose fully-loaded cost exceeds their revenue by a wide margin, subsidised by a small number of large ones — which is a viable model only if the small tenants convert at a rate that justifies it. The conversion rate is the number that decides whether a free tier is marketing spend or a loss.

The architectural choices that change the economics

• Shared clusters with logical isolation instead of an instance per tenant. This collapses the per-tenant infrastructure floor, which is usually the largest term, at the cost of weaker isolation and a noisy- neighbour problem requiring quotas.
• Scale to zero. Idle tenants consume no compute, with storage separated so that data persists cheaply. This is what makes small tenants economically viable, and it is why serverless database architectures exist. The cost is cold-start latency on the first request after idleness.
• Storage and compute separation generally, so the two scale independently and an idle tenant pays only for bytes.
• Aggressive automatic cleanup of abandoned resources, with clear notice. The retained-forever tail is pure loss and is usually significant.
• Hard quotas on free tiers, enforced rather than advisory, which bounds the worst case per tenant.
• Self-service support only below a threshold, since support cost per tenant is frequently the term that makes the arithmetic fail.
• Automated fleet operations, so per-instance engineering effort approaches zero — the difference between a fleet of 100,000 tiny tenants being viable or not is almost entirely automation maturity.

The strategic framing

A free tier is a marketing expense, and should be evaluated as one: cost per acquired paying customer, compared against other channels. Framed that way the decision is analysable. Framed as "we cannot afford to run free accounts," it is an infrastructure complaint that obscures the actual question — which is whether the tier produces customers at a competitive cost.