LLM Lock-In is the New Cloud Lock-In

LLM performance and cost curves are moving fast. The best foundation model today is likely not the best foundation model twelve months from now — and almost certainly not three years out. No one knows which model or which provider wins next. That is a good reason to be cautious about any decision that narrows your future options. 

The logic here maps almost one-to-one from what we already learned about cloud: resilience, performance, cost, and commercial leverage.

• Resilience. You need to be able to switch. Infrastructure outages happen. Geopolitical events happen. Recent disruptions due to drone strikes at cloud data centres in the Middle East were a reminder. For most insurers, it is also not feasible to just switch to another cloud region due to local storage and processing requirements for sensitive data. If your entire agentic claims or customer service stack depends on a single LLM provider (particularly if hosted on just one cloud provider), you have a concentration risk that belongs on your operational risk register. 

• Performance and cost. Model quality is moving in two directions at once: frontier capability at the top and dramatic cost-performance improvements in the mid-tier. You want to be able to shift your underlying LLMs when higher performing models emerge (mainly for complex use cases such as in underwriting) or when smaller, faster and cheaper models become available (for most use cases). We have just made that shift moving from some of the most powerful but expensive closed models for coding to some of the latest open source models released with token costs at only around one quarter of the price for almost equal performance. 

• Commercial leverage. Optionality is negotiating power. Procurement teams that have run a multi-cloud RFP know this. While pricing and discounting has been less liberal for AI tokens so far, I believe the same discipline needs to apply to foundation-model sourcing. 

There is also a trend that shapes certain regions: sovereign AI and sovereign cloud. In Europe in particular, regulators and large enterprises are increasingly explicit about data and model sovereignty. Insurers operating across jurisdictions will need the flexibility to route certain workloads to in-region, in-sovereignty compute and models to realise scale and reusability. 

This is why Peak3 applies the same design philosophy to our AI stack as we do to our core system offering. We build cloud-agnostic and LLM-agnostic. 

In practice, our customers choose their preferred deployment model (Peak3-managed or client-managed), their cloud provider, and their region. For each agentic use case, they select which foundation model to call — as configuration, not redevelopment. I need to make the honest caveat that some prompt optimisation across model families is unavoidable. 

Besides “AI-first” and "AI-native" phrases commonly used these days, let's also think about “AI-resilient”. Resilience and optionality don't come from picking the right providers today. They come from designing so that the wrong or suboptimal providers tomorrow don't take you down with them. 

If your current AI roadmap hard-codes a cloud or a model, it is probably worth pausing for an architecture review. I am happy to facilitate a session with my architecture team to share our philosophy and approach. Just message me or comment here. 

-Bill Song, Peak3 Co-Founder and CEO