How much does it cost to run a multiplayer game backend?

Game backends have an unusual cost shape: the compute is genuinely expensive because sessions need dedicated capacity with low latency, but the thing that catches studios out is bandwidth. Patch day is a bigger infrastructure event than peak concurrency. Here is what a mid-sized multiplayer title actually costs.

The workload we are pricing

Assume a session-based multiplayer game in us-east-1 with 200,000 monthly active players, 20,000 daily, and a peak of 3,000 concurrent. Each session server instance (c5.2xlarge) hosts about 100 concurrent players, so peak needs 30 instances and the daily average is around 15, with headroom. Players average one hour per session-day. There is a matchmaking and account API, a leaderboard and inventory store, a relational database for accounts and purchases, and a monthly patch of roughly 2 GB that most of the player base downloads.

The monthly breakdown

Patch day is the biggest bill of the month

Forty terabytes of client downloads at CloudFront tiered rates is 3,328 dollars plus request charges. The arithmetic: 200,000 players times a 2 GB monthly patch is 400 TB if everyone downloads everything, which is why delta patching exists. Assuming a 10 percent effective delta and partial adoption gets to 40 TB.

The optimization here is entirely about patch size. A build pipeline that produces 200 MB binary deltas instead of 2 GB full packages reduces egress by an order of magnitude. Studios that ship weekly rather than monthly and do not delta-patch can easily spend more on CDN than on every other line combined. Tuning cache behaviour matters too: patch files are immutable and should carry a one-year TTL so edge locations serve nearly all requests, keeping origin fetches and object storage request fees near zero.

Spot instances for session servers

Spot at roughly 70 percent off works for game servers because sessions are short. An instance that receives a two-minute interruption notice can stop accepting new matches, let current matches finish, and drain cleanly. The pattern is to keep a reserved or on-demand baseline covering the floor of concurrency so interruptions never take the service below capacity, then burst on Spot. That blended approach saves about 2,700 dollars a month against pure on-demand here, and the trade-offs are covered inpurchase options compared.

Gameplay egress does not go through the CDN

Real-time UDP game traffic cannot be cached, so it leaves the network at standard data transfer rates. At 20,000 daily players averaging one hour at 30 kbps each way, that is 600,000 player-hours a month times about 13.5 MB, or 8 TB, costing 737 dollars after the 100 GB free allowance. This line scales directly with engagement, unlike patches which scale with release cadence.

Reducing it means reducing tick rate or packet size, which is a gameplay decision, not an infrastructure one. Dropping from 30 to 20 updates per second cuts this line by a third but changes how the game feels. Most studios accept the cost. What is worth checking is that session servers and players are in the same region, since cross-region traffic between game servers and matchmaking adds avoidable charges.

Unit economics

At 8,325 dollars for 200,000 monthly players, infrastructure costs 0.042 dollars per monthly player and 0.139 dollars per daily player-month. Against a free-to-play ARPU of, say, 1.20 dollars per monthly player, infrastructure is 3.5 percent of revenue. The dangerous scenario is a viral spike: a 10x player surge multiplies session compute, egress, and patch delivery simultaneously, taking the bill to roughly 70,000 dollars in a month. Price the scaling behaviour from Terraform against theresource catalog so the 10x number exists before the spike does.

FAQ

What to do next

Model the 10x spike before it happens. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.