Message Queue
Massive buffer holding 200 requests. Place after Firewall to
smooth DDoS spikes, or before Compute to prevent overload. Cheap
insurance against traffic bursts.
Flow: Firewall → Queue → Load Balancer or Load Balancer → Queue
→ Compute
Memory Cache
Caching layer with 35-65% hit rate. Cache hits complete
instantly without hitting DB/Storage. Upgrade tiers for higher
hit rates. Great for reducing DB costs.
Flow: Compute → Cache → DB/Storage
API Gateway
Rate limits incoming traffic. When requests exceed the rate limit,
excess requests are throttled with only -0.2 reputation loss (vs -1.0 for failures).
Upgrade tiers to increase rate limit (20/40/80 RPS).
Flow: Internet → API GW → ALB or WAF → API GW → ALB
NoSQL Database
Faster and cheaper than SQL for READ/WRITE traffic (150ms vs 300ms processing).
Cannot handle SEARCH queries — keep SQL DB for those.
Upgrade tiers for higher capacity (15/30/50).
Flow: Compute → NoSQL or Compute → Cache → NoSQL
Search Engine
$120
Processes SEARCH traffic 3× faster than SQL DB (100ms vs 300ms). Only accepts SEARCH queries — other traffic types will fail. Upgrade tiers for higher capacity (12/25/40).
Flow: Compute → Search Engine or Cache → Search Engine
Read Replica
$100
Processes READ traffic faster than master DB (200ms vs 300ms). Must be connected to a SQL DB or NoSQL DB to function. Only accepts READ queries.
Flow: Compute → Replica → DB or Cache → Replica → DB
Serverless Function
$45
Auto-scales with traffic (capacity 30). Very low upkeep ($2/min) but charges $0.03 per completed request. Cheap for bursty / low-volume traffic; expensive at sustained high RPS. Same connection topology as Compute.
Flow: ALB / Queue → Serverless → Cache / DB / NoSQL / S3
Compute Node
$60
Processes every request type and routes it onward. Upgrade tiers scale it UP (capacity 4/10/18). Toggle AUTO to scale it OUT instead: an Auto-Scaling Group boots an instance whenever utilization stays above 70% and retires one below 30%, between 1 and 5 instances. A new instance needs 3s to warm up and carries no traffic until it does — that cold start is why you pre-scale ahead of a known peak — and every instance is billed from the moment it boots. Real-world equivalents: AWS Auto Scaling Group · Azure VM Scale Set · GCP Managed Instance Group.
Flow: LB / Queue → Compute → Cache / DB / NoSQL / Storage
Monitoring
$75
Collects utilization, queue depth, error rate and latency from every service and raises alerts before overloads become outages. Never receives traffic — place it anywhere. Real-world equivalents: AWS CloudWatch · Azure Monitor · Google Cloud Operations.
Flow: standalone — no connections; watches every service
Dead-Letter Queue
$55
Catches requests that finally failed (retry exhausted or no route) at any connected upstream, parking them instead of dropping them. It then drains slowly on its own, recovering each parked request for a small money cost and refunding a little reputation. If it overflows its cap, further dead requests fail normally plus an extra reputation penalty — an unmanaged DLQ is worse than none. Real-world equivalents: AWS SQS dead-letter queue · Azure Service Bus dead-letter queue.
Flow: Compute / Serverless / LB / API GW → DLQ (drains automatically)
Pub/Sub Topic
$65
Fan-out: one inbound event is delivered to every connected subscriber — one copy each, capped at the number of subscribers. It is the only node that multiplies traffic, so a single event becomes N independent deliveries that each run their own path. Wire it after a Load Balancer or API Gateway and out to processors, storage, or a Notification sink. Real-world equivalents: AWS SNS · Azure Event Grid · Google Cloud Pub/Sub.
Flow: LB / API GW → Pub/Sub → Compute / Serverless / Notification / Storage (×N)
Identity Provider
$55
Sits in-line on the request path and trades latency for security: every request through it pays an extra processing delay, and half the Malicious traffic that reaches it is caught — the session-based attacks a WAF alone misses. What it does not catch slips through and breaches downstream, so it is a second layer, not a replacement for the firewall. Real-world equivalents: AWS Cognito · Microsoft Entra ID · AWS IAM.
Flow: Internet / WAF → Auth → Load Balancer / API Gateway / Compute
Scheduler
$50
A traffic source, not a processor: every 8 seconds it injects its own burst of batch jobs into its downstream, independent of the external request rate. It teaches capacity planning for known peaks — provision for the wave before it lands. It freezes when the game is paused and never leaks jobs. Real-world equivalents: AWS EventBridge Scheduler · Google Cloud Scheduler.
Flow: Scheduler → Queue / Load Balancer / Compute / Serverless
Notification
$40
A terminal 'send': it completes an event and, unlike Storage or the database, its success earns bonus reputation (user goodwill), so it is valuable even when the money reward is small. Its overload failures are silent — no alarm, only a quiet drip of dissatisfaction rather than a scored, sonified drop. Real-world equivalents: AWS SES/SNS · SendGrid-class email/SMS senders.
Flow: LB / Pub/Sub / Scheduler → Notification (terminal 'send')
Container Cluster
$120
The economic anti-twin of Serverless: dense fixed capacity (12, three times a Tier-1 Compute) at a flat per-cluster fee with no per-request charge. It shares Compute's processing brain, so the different choice is purely economic — it wins steady high-throughput, where the flat price beats Serverless's per-request bleed and its density beats paying upkeep for many Compute instances. Toggle AUTO to scale it out, but its node pool warms up far slower than an ASG Compute VM, so pre-scale before a known peak. Real-world equivalents: Amazon EKS · Azure AKS · Google GKE.
Flow: LB / Queue / API GW → Container Cluster → Cache / DB / NoSQL / Storage
Stream
$90
High-throughput ingestion split across independent partitions, each processed in strict order. It is the only node that models head-of-line blocking: if a partition's head cannot forward because its consumer is saturated, the whole partition waits behind it — it never skips ahead — while the other partitions keep flowing. That is the ordering-vs-throughput lesson: order is preserved per partition, but one slow consumer stalls its partition's tail. Real-world equivalents: Amazon Kinesis · Azure Event Hubs · Apache Kafka.
Flow: LB / API GW → Stream (partitioned) → Compute / Serverless / Storage / Warehouse
GeoDNS
$50
The front-most entry point: it sits in front of the internet-facing tier and splits traffic across two or more independent regional stacks, each its own WAF → ALB → … pipeline. Where a Load Balancer distributes workers within one stack, GeoDNS distributes whole stacks at the very front — the basis for multi-region and regional failover. Wire the Internet to it, then wire it out to the front door (WAF / ALB / API Gateway) of each region. Real-world equivalents: AWS Route 53 · Azure Traffic Manager · Google Cloud DNS.
Flow: Internet → GeoDNS → WAF / ALB / API GW of each region (×N)
Data Warehouse
$90
An analytics store with the inverse profile of a database: it completes WRITE and UPLOAD traffic — slowly and cheaply at high volume — but fails any realtime READ or SEARCH that reaches it, because a warehouse is OLAP, not OLTP. You load it, you do not serve user reads from it. Feed it an analytics copy from a Pub/Sub fan-out, a scheduled batch load from the Scheduler, or a Stream. That read-rejection plus cheap-high-capacity is the OLTP-vs-OLAP lesson. Real-world equivalents: Amazon Redshift · Azure Synapse · Google BigQuery.
Flow: Pub/Sub / Scheduler / Stream → Data Warehouse (analytics terminal)
GPU Cluster
$300
The only node that batches: INFERENCE requests accumulate (up to the batch size, or 1.5s from the first arrival) and run as ONE job, so a full batch amortizes the fixed cost and a lonely request pays nearly full price — utilization IS the economics. Tiers are model size: bigger batches AND a lower bad-answer risk (10%/4%/1%, printed on the upgrade card), paid for with a longer model load (12/20/30s) during which the node serves nothing — upgrade in a lull, never mid-surge. It draws 6 kW from the power grid. Real-world equivalents: AWS P5/Inferentia · Azure ND-series · GCP A3/TPU.
Flow: Compute / Serverless / Cluster / Inference GW → GPU (terminal; batches INFERENCE only)
Inference Gateway
$70
The honest pitch: a direct-wired GPU has only its tiny bounded intake — during a model load or an overload, requests die fast. The gateway holds up to 20 requests with deadline honesty: anything older than 6 seconds is failed as an SLO breach (a stale answer is worthless), everything else is dispatched to the least-loaded GPU with room in its batch. Its value is measured in reputation saved during warmup and burst windows, not in revenue. Real-world equivalents: vLLM router · NVIDIA Triton · Amazon Bedrock.
Flow: LB / API GW / Compute tier → Inference GW → GPU (deadline queue, INFERENCE only)
Substation
$150
Not a traffic node — it is never wired to anything. The base grid supplies 8 kW, every GPU draws 6, and a GPU cannot be placed past the cap: each substation adds +6 kW, so a three-GPU fleet needs two of them. Watts are a standing marginal cost of AI capacity, not a one-time unlock — and a substation whose loss would strand powered GPUs refuses to be demolished until the GPUs go first. Real-world equivalents: datacenter power provisioning · rack power budgets · utility interconnect queues.
Flow: standalone — no connections; +6 kW grid capacity per substation