GS2-Steady

Running costs no in-house build can match.

For titles above a certain scale, backend cost of around $0.016 per user per month.
A new pricing plan aiming for around $0.008 with a one-year commitment.

  • ~$0.016

    Monthly cost per user
    for titles above a certain scale

  • ~$0.008

    Monthly cost per user
    with a 1-year commitment

  • 16.6ms or less

    Average response time
    within the same data center

  • 1.02M RPS

    Requests per second
    the largest fleet absorbs

Why we built it

Building got cheaper.
Running it did not.

As coding agents advance, the cost of building a game backend has dropped sharply.

But the servers and databases you pay for every month after launch cost about the same as they always have — and the longer you operate, the more that adds up.

So we redesigned the GS2 backend architecture and built GS2-BufferedSession to bring down the cost of running itself.

Price level

Per user, per month:
around $0.008.

Example: a 1,800,000 MAU title that fits in std1.large

Standard price

$23,500/mo ÷ 1,800,000

Per user $0.0131 /mo

1-year commitment

$11,800/mo ÷ 1,800,000

Per user $0.0066 /mo

The fleet you need depends on your game design and API usage. To estimate with your own numbers, use the calculator.

Architecture

We rebuilt the backend,
starting from the architecture.

GS2-BufferedSession sits between the GS2 API and the database, keeps player data in memory, and writes updates in batches.

It delivers its best cost performance in games that update the same player data many times a minute.

STEP 01

The database is where game server costs go

Every player action calls the API, and every API call reaches the database. Database access costs more than the servers themselves.

STEP 02

Every API call is a round trip to the database

Each request becomes a database access, and the number of accesses is what you pay for.

STEP 03

Absorb in memory, write in batches

GS2-BufferedSession keeps player data in memory and writes updates to the database in batches. However many times the same data changes, it takes a single write.

STEP 04

Critical updates are written on the spot

When a critical update arrives — spending paid currency, for instance — the pending updates for that player are flushed first, and then it goes straight to the database. Even mid-window, important data never waits.

STEP 05

No waiting on the database, so faster than a frame

Fewer accesses mean less waiting. With the database round trip gone, the API responds in under one frame (16.6ms) on average.

Monthly game server cost

API servers
Database
Network
Other

The biggest share goes to the database

* Illustrative. Actual shares vary by setup and game.

Requests 0

DB accesses (without buffering) 0

DB accesses (BufferedSession) 0

Reduction (in this animation) -

When a critical update arrives, pending updates are flushed first, and then it is written to the database immediately.

  • Regular update
  • Critical update, e.g. paid currency (written immediately)
  • * Batches are written every 15 minutes; shortened to seconds here.

One game frame(60fps)

16.6ms

Without buffering

API
DB
DB
DB

BufferedSession

API
Memory
  • < 16.6ms Average response
  • < 100ms p99

* Measured within the same data center, excluding internet latency. The bar without buffering is illustrative.

The buffering window

Fixed at 15 minutes. Repeated writes to the same data inside the window never touch the database — they are handled in memory. An unexpected failure can lose whatever the window still held.

Critical writes go straight through

Because GS2 exposes fine-grained game context as APIs, it knows how important each write is. A write that matters — spending paid currency, for instance — flushes the buffer and reaches the database immediately, even mid-window.

Scale without stopping

Whether you scale up or scale in, a retiring instance flushes its buffer before it leaves. Nothing in the buffer is lost, and there is no downtime.

Faster than a frame

With the database round trip gone from ordinary API calls, responses average under one frame (16.6ms) and stay under 100ms even at p99 (round-trip time within the same data center).

An illustrative AWS cost comparison

Make the assumptions clear.
Then compare costs.

An illustrative AWS Tokyo configuration for 60,000 CCU and about 4,000 peak external RPS. With an average fanout of four, the GS2 side processes about 16,000 internal RPS, compared here using std1.large.

Infrastructure costs for this example

GS2-Steady Standard price

$23,500

GS2-Steady 1 year prepaid, monthly equivalent

$11,800

AWS example 1-year All Upfront, monthly equivalent + excluded costs

$26,300 On-demand, 30 days $35,000

The faded AWS bar shows 30-day on-demand costs. The one-year commitment is the annual cost divided by 12.

This is an estimate of selected infrastructure costsunder the stated assumptions. It does not demonstrate matching performance, durability or operations. Excluded AWS costs must be added.

AWS example breakdown About 4,000 peak external RPS, 30-day on-demand

Item Configuration / assumption On-demand, 30 days 1-year prepaid, monthly equivalent
API servers Fargate x86 · 16 vCPU / 32 GiB × 18 (6 × 3 AZ) $12,800 $10,100
Database instances Aurora I/O-Optimized db.r8g.2xlarge × 12 $15,000 $9,650
Database storage Aurora I/O-Optimized · 1 TiB $276 $276
Cache Redis cache.r7g.xlarge × 6 $2,260 $1,470
Log ingestion CloudWatch Logs Standard · 2 KiB / external request $4,510 $4,570
Log storage CloudWatch Logs · 30 days · assumed compression ratio 1.0 $196 $196
Log queries Logs Insights · 1 full scan / month $45 $46
Subtotal (excluded costs extra) $35,000 $26,300
  • Scenario: 60,000 CCU and about 4,000 peak external RPS (about 16,000 internal GS2 RPS on average), with an illustrative AWS Tokyo setup.
  • Pricing: 30 days on demand, or an annual upfront commitment divided by 12. AWS conversion: JPY 160/USD, before tax. Performance and operating conditions are not proven equivalent.
  • Additional costs: networking, extra backups, monitoring, support, development and operations. WAF and temporary Blue/Green deployment capacity are also excluded.
View detailed assumptions, calculations and sources
  • Workload assumptions: 60,000 CCU, 4,000 peak external RPS, and roughly 1:1 reads to writes. Average GS2 fanout of four gives about 16,000 internal RPS; not every request expands fourfold.
  • API sizing references Sumzap’s published example: 225 RPS at 40% CPU on a 16 vCPU / 32 GiB Fargate task. We assume 18 tasks and an x86 runtime. This is a sizing reference, not evidence of identical workload performance.
  • Six tasks in each of three AZs. Linear extrapolation from that example gives about 39.5% CPU normally and 59.3% across the remaining 12 tasks after losing one AZ. This does not maintain the normal 40% CPU utilization after a failure; workload testing is required.
  • Four database shards, each with one writer and two readers: 12 instances total. Six Redis nodes and 1 TiB of stored database data across all shards are assumptions, not validated capacity for a particular game.
  • Average load is 30% of peak. Logs total 2 KiB across all services per external request, retained for 30 searchable days in CloudWatch Logs Standard, with one full monthly Logs Insights scan.
  • AWS Tokyo public rates. On-demand uses 720 hours; one-year All Upfront commitments use annual cost divided by 12 (730 monthly hours for ongoing charges). Fargate Compute Savings Plans, Aurora RI normalized for I/O-Optimized, and Redis RI are calculated individually, without blanket or volume discounts. Logs use 720 hours for on-demand and 730 hours for the annual monthly equivalent. AWS yen conversion uses JPY 160/USD; GS2 dollar billing uses its separate JPY 120/USD product rate, so the conversion bases differ.
  • Excluded: ALB, NAT, data transfer and other networking; backup overage; metrics and dashboards; support; taxes; development and operations staff. WAF and temporary additional Blue/Green deployment costs are also outside this comparison.
  • Sumzap: Fargate sizing reference · AWS Fargate · Compute Savings Plans · Aurora · ElastiCache · CloudWatch (2026-11-09)

Different performance, resilience and operations

The AWS example uses one region and multiple AZs; its topology and operations are not identical to GS2. GS2 ordinary updates may lose unpersisted data within the 15-minute buffer window, while important updates are persisted immediately. No equivalence in performance, durability or operations is claimed.Load testing and failure planning must follow your requirements.

Small titles want pay-as-you-go

GS2-Steady starts at $2,700/mo. For titles below a certain scale, GS2's pay-as-you-go plan may be cheaper. Once you grow into it, you can switch across without moving any data.

Pricing

Three fleet families

Balanced games want the std family.
Games with many concurrent users want ccu, which carries more memory for buffering.
Games that call the API very frequently want rps, with extra throughput.

Fleet Max CCU Max RPS Standard 1-year commit
std1.nano 3,000 1,000 $2,700/mo $1,300/mo
std1.micro 9,000 2,000 $4,000/mo $2,000/mo
std1.small 12,000 4,000 $6,300/mo $3,200/mo
std1.medium 28,000 8,000 $12,200/mo $6,100/mo
std1.large 60,000 16,000 $23,500/mo $11,800/mo
std1.xlarge 124,000 32,000 $46,700/mo $23,300/mo
std1.2xlarge 248,000 64,000 $90,000/mo $45,000/mo
std1.4xlarge 496,000 128,000 $176,700/mo $88,300/mo
std1.8xlarge 992,000 256,000 $343,300/mo $171,700/mo
std1.16xlarge 1,984,000 512,000 $671,700/mo $335,800/mo
std1.32xlarge 3,968,000 1,024,000 $1,313,300/mo $656,700/mo

All prices are monthly and exclude tax. "1-year commit" is the monthly equivalent when you prepay a full year of usage. Prices are preview estimates and may be revised when the service becomes generally available.

About max CCU and max RPS

  • Max CCU The hard limit on concurrent connections. Once a fleet reaches it, no further logins are accepted.
  • Max RPS A level that keeps the fleet below 60% of its capacity, leaving headroom for failover. Going above it does not stop the service right away, but a failover or node replacement may then cause an outage. With GS2-Script the fleet may fall short of this level, so run your actual workload to judge.

What's included

  • Storage and data transfer
  • Cold backup restorable to any second within the last 30 days
  • Log retention for the past 30 days
  • Log and metrics dashboards powered by GS2-Insight

Moving between plans

You can move between GS2-Steady and the pay-as-you-go plan in either direction. Both plans share the same database, so no data migration is required. A move does involve downtime, so we expect it around launch or at a turning point in live operations. Moving between plans requires a project that uses database schema v2.

Calculator

Find the fleet you need

Four numbers is all it takes. We size the fleet for the launch month and for everything from day 30 onward.

Typically 7–14 minutes

Launch month

—

—

How this was calculated

Downloads through day 7

—

Your input

Day 1 downloads

—

× 0.5

Peak-hour actives

—

× 0.5

Login RPS

—

÷ 3600 × fetches

In-session RPS

—

Peak CCU × requests per minute ÷ 60 × 4

Peak CCU

—

× session minutes ÷ 60

RPS needed
—
CCU needed
—

Day 30 onward

—

—

How this was calculated

Downloads through day 30

—

Day 7 total × 1.5

Day 30 active users

—

× 0.1

Peak-hour actives

—

× 0.5

Login RPS

—

÷ 3600 × fetches

In-session RPS

—

Peak CCU × requests per minute ÷ 60 × 4

Peak CCU

—

× (session minutes × 0.8) ÷ 60

RPS needed
—
CCU needed
—

Estimated from average GS2 statistics. Per-user prices use downloads through day 30 as the launch-month MAU, and day 30 active users converted at a 50% DAU/MAU ratio from day 30 onward.

Building a backend is easier than ever.

So GS2 gives you a foundation that keeps your game running for years, for less.

That is the value the new GS2 aims to bring as a BaaS.

Preview

GS2-Steady.
Be among the first to use it.

We are accepting preview registrations ahead of general availability. Register and we will let you know when it opens, and talk through a configuration sized for your title.