Tier-0 · Local GPU (Local Zero-Cost)
Own hardware: privacy, offline operation, zero external-provider limits.
Ollama -- Qwen2.5-Coder 14B (code) · DeepSeek-R1 14B (reasoning, second opinion).
AIOS Cloud Architecture on Google Cloud
Services and Their Roles
- Vertex AI (Planned): model inference for coding and vision workers.
- Cloud Run: stateless worker execution, scale-to-zero.
- Cloud Storage: artifacts -- textures, Unity packages, task outputs.
Cost Control
- Tier cascade: local inference -> free API -> paid cloud.
- Daily limits per provider.
- An overarching global daily ceiling.
- An audit trail for every paid call BEFORE execution.
FinOps -- three layers, one rule
Free first, then local, paid cloud only with explicit, one-time approval -- never a silent fallback.
Tier-1 · Free Cloud (Free API)
Fast, free APIs are the first choice for any task attempted in the cloud.
Cloudflare Workers AI · Google AI Studio (Gemini Flash) -- already active in production.
Tier-2 · Paid Cloud (guarded)
Enabled in code, but gated on every call: a hard $0.50/call limit, a closed list of allow-listed purposes, and a full audit trail before execution. Activating broader usage is always a single, explicit operator decision -- never a silent fallback once cheaper tiers are exhausted.
Azure OpenAI is the Tier-2 provider with a coded, active call path today. AWS Bedrock and Google Vertex AI are named in the routing configuration and tracked by the same budget guard (Shadow Ledger), but the API call implementation for these two does not exist in the code yet -- activating them requires writing that integration, not just supplying credentials.
Google Cloud is part of this architecture at two levels: Google AI Studio (Gemini Flash) as the free, first-choice layer for most tasks -- already active in production -- and Google Vertex AI as a named, budget-tracked Tier-2 target whose API integration is not yet coded; wiring it up, and the same explicit operator decision required for the other guarded tiers, are both still ahead.