The core choice
When deploying AI for your business, one of the first decisions is where the AI runs and where your data goes. The spectrum runs from fully public (cloud API) to fully private (self-hosted on your own infrastructure).
Public AI (cloud-hosted)
Public AI means using cloud services like OpenAI's API, Google Gemini, or Anthropic's Claude directly. Your data is sent to their servers for processing.
- Pros: Easiest to start, no infrastructure to manage, always up-to-date models, lowest upfront cost.
- Cons: Data leaves your control, potential training data usage, higher per-query cost at scale, dependency on vendor availability.
Private AI (self-hosted)
Private AI means running models on your own infrastructure — on-premises servers, your own AWS account, or dedicated cloud instances. Open-source models like Llama 3, Mistral, and Phi can be self-hosted.
- Pros: Full data control, no data leaves your environment, no vendor dependency, no per-query API cost.
- Cons: Higher upfront cost, requires ML ops capability, models may be less capable than frontier options, you manage updates and scaling.
The middle ground: AWS Bedrock. Use frontier models (Claude, Mistral) via API, but your data stays within your AWS account and VPC. Not used for training. Best of both worlds for many use cases.
Comparison table
| Factor | Public API | AWS Bedrock | Self-Hosted |
|---|---|---|---|
| Data location | Vendor servers | Your AWS account | Your infrastructure |
| Model quality | Frontier | Frontier | Good (open-source) |
| Setup effort | Minutes | Hours | Days to weeks |
| Per-query cost | Medium | Medium | Low (after setup) |
| Infrastructure cost | None | AWS services | GPU instances |
| Privacy | Varies | Strong | Maximum |
| Ops burden | None | Low | High |
The hybrid approach
Most of our clients end up with a hybrid:
- AWS Bedrock for production RAG systems and customer-facing AI — strong models with data isolation.
- Public APIs (ChatGPT, Claude) for internal productivity — drafting, brainstorming, coding where data sensitivity is low.
- Self-hosted models for specific use cases requiring maximum privacy or offline access.
Our recommendations
- Sensitive customer data: Bedrock or self-hosted. Don't send it to public APIs.
- Internal knowledge systems: Bedrock with VPC endpoints. Data stays in Sydney.
- General productivity tools: Enterprise plans of ChatGPT or Claude (with training opt-out).
- Regulated industries: Self-hosted or Bedrock. Document your data flows for compliance.
Key takeaways
- Public AI is easier and cheaper to start with. Private AI gives you full data control.
- AWS Bedrock offers a middle ground — frontier models with data staying in your account.
- Self-hosting open-source models (Llama, Mistral) gives maximum control but requires ML ops capability.
- Most businesses use a hybrid: Bedrock/private for sensitive data, public APIs for general tasks.