Data residency determines where a model runs and where sensitive data is stored or processed. Sending regulated data to a third-party API may violate compliance requirements; adaptation alone does not solve that problem. section 1.3 section 2.4
Treat residency as a deployment decision, separate from the choice of adaptation technique. A team can self-host whichever rung it selects, from LoRA to full supervised fine-tuning. section 1.7
Two deployment paths can keep data within a required boundary:
- Use a provider’s tenant in the customer’s cloud region, with data-residency and no-training commitments. Azure OpenAI, Amazon Bedrock, and Google Vertex AI are named options. section 1.3
- Self-host on organizational infrastructure, accepting responsibility for GPU capacity, serving, latency, and throughput. section 2.4
Hospital deployment
When patient data cannot leave the hospital, a team can start with a self-hosted open model, train a LoRA adapter on a few hundred curated clinical examples, and consider full fine-tuning if the adapter plateaus. section 1.7
Table 2.3 Buy versus build: criteria for choosing between a managed API and a self-hosted model.
|
Criterion
|
Buy: managed API
|
Build: self-hosted
|
|
Setup time
|
Minutes
|
Days to weeks to provision GPUs and
configure the serving stack
|
|
Upfront cost
|
Low, pay per token
|
High, physical GPU hardware or cloud
GPU VMs
|
|
Cost at scale
|
Grows linearly with usage
|
One-time infrastructure cost amortized
across requests
|
|
Customization
|
Prompt engineering plus limited
fine-tuning APIs
|
Fine-tuning, weight access, custom
serving stack
|
|
Data privacy
|
Sent to vendor servers
|
Stays on your infrastructure
|
|
Model choice
|
Vendor decides versions and
deprecations
|
Any open-weight model you choose
|
|
Latency control
|
Vendor SLA on shared infrastructure
|
You own the hardware and networking
path
|
|
Maintenance burden
|
Zero infrastructure burden; vendor
management overhead (deprecations, pricing, terms) is real.
|
Ongoing operational commitment for
hardware, serving stack, and scaling
|
Compare managed APIs and self-hosting across privacy, setup time, cost, customization, and operational burden. section 2.4


