know.2nth.ai tech google compute
2nth × Google Cloud Compute

From a cron job
to a GPU cluster. In Johannesburg.

The 2nth.ai compute tier on Google Cloud covers the full spectrum — from a scale-to-zero container that costs you nothing when idle, to a persistent VM running a self-hosted ERP, to a GPU instance serving your own AI models. Every workload runs in africa-south1, protected by the same defence-in-depth pattern the hyperscalers use for their own platforms.

Region · Johannesburg Residency · POPIA-aligned Zones · 3 independent Scale · 0 → 1000+ instances
01 · Two ways to compute

Pick the shape that fits the workload.

Most modern applications want serverless containers — deploy once, let the platform handle scaling and billing. Some workloads need full virtual machines — persistent state, a specific kernel, or a GPU. 2nth uses both, picked per workload, not per opinion.

Serverless containers

Cloud Run · default for new work

Ship a container, get a global HTTPS URL. No servers to patch, no scaling to configure, no idle cost.

  • Scales from 0 to 1,000+ instances in seconds
  • Billed per 100ms of actual CPU time
  • Deploys from Docker image or source
  • Integrates with Pub/Sub, Cloud Scheduler, Eventarc
  • Limit: stateless, max 60 min per request

Virtual machines

Compute Engine · when you need full OS control

Full Linux or Windows VMs with persistent disks. You choose the kernel, the libraries, the tuning.

  • Sized from 2 vCPU / 1GB up to 224 vCPU / 896GB
  • Persistent disks + in-memory state
  • GPUs available — T4, L4, A100, H100
  • Spot VMs up to 91% cheaper for batch work
  • 1-year commits save ~25%, 3-year save ~52%
02 · What you can build

Six shapes, one platform.

Real workloads 2nth runs or is ready to run on this tier. Each one picks serverless or VM based on what the workload actually needs — not by rule.

ERP · persistent

Self-hosted ERPNext / Frappe

Run your own ERP, fully in SA, on your own schedule. Wake on demand, sleep between sessions, auto-backup to regional storage. Freedom from subscription traps and offshore data.

Cloud Run · wake-on-request + Cloud SQL · MySQL 8.0
API · serverless

REST / GraphQL backends

Public APIs that scale from zero to a product launch. Pay for requests you actually serve. Fronted by Cloudflare for DDoS, rate limiting, and edge auth.

Cloudflare Worker · edge → Cloud Run · private core
Batch · scheduled

Overnight jobs and reports

ETL pipelines, monthly reports, mail digests, data reconciliation. Triggered on schedule, run to completion, cost nothing for the other 23 hours.

Cloud Scheduler → Cloud Run Jobs · parallel tasks
AI · inference

In-region AI inference

Self-hosted models where data can't leave SA, or Gemini / Claude via Vertex AI for the heavy lifting. Streaming responses, structured output, tool calls.

Cloud Run · proxy + Vertex AI · Gemini + Claude
Webhooks · integrations

Webhook receivers

Endpoints that receive events from Slack, WhatsApp, Stripe, Meta, Gmail, GitHub, Zoho. Parse, classify with AI, fan out via Pub/Sub, write to your systems.

Cloudflare Worker + Cloud Run · verified + signed
Ops · internal

Internal tools & dashboards

Ops dashboards, admin UIs, custom reports. Gated by Cloudflare Access, hosted privately, visible only to named team identities — no public surface at all.

Cloudflare Access → Pages · UI → Cloud Run · API
03 · Cost shape

You pay for what you use. Nothing for idle.

Illustrative monthly numbers for the workload shapes above. Not quotes — indicative cost ranges based on published GCP pricing for africa-south1. Real numbers depend on traffic, data transfer, and committed-use discounts.

Indicative monthly cost

ERPNext on wake-on-request Cloud Run 4 hrs / day business use · 2 vCPU · 2 GB RAM · Cloud SQL db-f1-micro
~$30 – $60 / mo
Public API, moderate traffic 1M requests / month · avg 200ms · 512 MB Cloud Run · Cloudflare edge free tier
~$5 – $12 / mo
Nightly batch job 1 hr / night · 4 vCPU · 8 GB · Cloud Run Jobs
~$3 – $8 / mo
Internal tools cluster 5 services · low traffic · scale-to-zero · Cloudflare Access gated
~$2 – $10 / mo
Persistent VM for legacy workload e2-medium · 24/7 · 20 GB boot disk · 1-yr commit
~$25 / mo
GPU inference on-demand NVIDIA L4 · 10 hrs / month · on spot
~$6 – $15 / mo

The shape that matters: most 2nth workloads spend more time idle than active. Scale-to-zero + spot pricing + 1-year commits together usually cut total cost by 50–70% vs running 24/7 on reserved capacity. For anything that isn't always-on, serverless is effectively free until real users show up.

04 · Why run it through 2nth

You get the platform and the pattern.

Google Cloud gives you the primitives. 2nth gives you the opinionated wiring — the hybrid with Cloudflare, the POPIA posture, the operational patterns, and the skills tree that lets an AI agent help operate what we build together.

01 · Region
Built for South Africa

Every service in Johannesburg, every byte under POPIA. No "data might leave the country" asterisks.

02 · Edge
Cloudflare in front

DDoS, WAF, bot management, rate limiting, and caching handled at the global edge — your compute never sees malicious traffic.

03 · Pattern
Private by default

No public IPs, no open ports. Every internal call is authenticated by identity; every external call is filtered at the edge.

04 · Cost
Pay for work, not idle

Scale-to-zero is the default, spot + commitments are layered where they fit. Most workloads settle well below equivalent hyperscale pricing.

05 · Speed
Deploy in minutes

From a container image to a live HTTPS URL in under 90 seconds. From a git push to production in under 5 minutes.

06 · Agent
Agent-operable

Every skill node lives in an AI-readable knowledge tree. Penny and the other 2nth agents can operate what we build together — from deploy to incident.

Talk to 2nth

If you're a South African business that needs serious cloud infrastructure without losing data residency or paying for idle capacity, we'd like to talk. Compute is just the start.

2nth.ai →