The original promise of AI inference rested on a clean, portable exchange of inputs and outputs—a semantic record you could take anywhere. But as Earendil Engineering highlights in a recent analysis, that abstraction is rotting. We are moving from transparent transcripts to provider-sealed 'states,' where the illusion of data ownership masks a hard technical tether. While early LLM adopters could archive and replay conversations across different backends, modern APIs have pivoted to returning opaque reasoning tokens, encrypted blobs, and internal subagent instructions that stay firmly behind the provider’s curtain. What was once a stateless transaction has become a proprietary black box.

The industry is quietly erecting barriers through the aggressive use of provider-bound IDs and non-portable artifacts. Features marketed as performance optimizations—like prompt caching, compacted context, and restricted vector-store references—ensure that a session cannot be resumed by a competitor. Even integrated web searches frequently hide source material from the client while allowing the model to see everything. According to Earendil, this architecture creates a structural dependency on specific GPU instances and server-side storage. For an engineering lead, the dream of seamlessly switching from OpenAI to Anthropic or a local Llama instance is dead if the session’s logic is locked in a vendor-dereferenced ID.

This isn't just about technical debt; it’s a strategic play for deep vendor lock-in. Earendil proposes a litmus test for true session ownership: portability requires a self-contained archive that functions without a provider’s proprietary key. For platforms building complex agentic workflows, surrendering these hidden states means surrendering the very logic of their product. The industry sold these features as 'enhanced security' and 'latency reduction.' In reality, they delivered a leash. If you can't dereference your own session history without paying a specific vendor's toll, you don't own your AI—you're just renting a seat in their walled garden.

AI in BusinessLarge Language ModelsCloud ComputingDigital Transformation