You probably shouldn't host your own LLM. And definitely shouldn't train one.

From time to time, I get questions from insurance leaders: Should we opt for proprietary models or self-host an open-source LLM? And should we train our own domain-specific model, full LLM or smaller SLM, on our claims, underwriting and policy data? 

For most insurers, my simple answer to both is no. 

There are exceptions. Global tier-1 groups like Allianz or Zurich, and some regional industry associations, may have the engineering depth and the capital to do either credibly. But that is a small minority. In my view, more than 95% of insurers don't have the in-house capability, the talent pool or the budget to self-deploy or self-train. They also don't need to. Pursuing it is rarely a competitive advantage and often a long, expensive detour. 

A nuance worth keeping. Don't self-host is not the same as don't use open-source. Open-weight models like DeepSeek, Qwen or Mistral have closed most of the gap to proprietary frontier models, at a fraction of the token cost. The right way for most insurers to access them is through commercially managed hosting, such as DeepSeek on AWS Bedrock. You get the cost-performance of open source without owning the GPUs. Direct self-hosting on your own infrastructure only pays off at very high, very predictable usage volumes. 

Why direct self-hosting doesn't make sense for most insurers. 

• Total cost of ownership. GPU procurement or leasing, MLOps, security hardening, regional redundancy, 24/7 SRE, and compliance are all expensive. It's a similar logic to why traditional compute has moved from own data centers to public cloud. Unit economics only work if your inference volumes are very large, sustained and stable.  

• Pace of model improvement. The frontier in model performance moves every few months. Whatever you launch today is twelve months behind by next year. With a managed gateway, you swap models as configuration. With a self-hosted set-up, you swap them as a custom project. 

Why in-house training is usually worse. 

• Premature convergence and unstable generalisation. Train an LLM or SLM on narrow insurance data and it locks in fast on that distribution. Performance outside that distribution degrades. New product lines, new geographies, new regulations — all of it becomes a retraining problem. 

• Reasoning regression. Custom-trained domain models routinely under-perform a frontier model with good RAG and prompting on the same task. This could mean six months and a seven-figure budget to land below where you could have started. 

Fine-tuning can be a different conversation. There may be real cases for SLM fine-tuning where latency, privacy or cost demand it.  

In any case, we are flexible to support insurers in whichever way they want to proceed. Our AI solutions are built to be agnostic of any underlying model and deployment option.

-Bill Song, Peak3 Co-Founder and CEO