The Hidden Cost of Cheap AI: A New Era of Vendor Lock-in
AI-generated image · Bay Street Wire
As compute costs drop, industry leaders warn that integrated AI platforms are creating 'walled gardens' that make switching providers nearly impossible.
While the cost of running AI is decreasing through faster models and cheaper GPU infrastructure, these savings may be offset by strategic vendor lock-in, as BetaKit first reported. Avery Pennarun, CEO and co-founder of Tailscale, warns in a BetaKit piece that AI is mirroring the restrictive business models of the 1970s mainframe era.
Pennarun argues that providers are bundling models, storage, networking, and management into integrated platforms. While these tools save initial engineering time, they create deep dependencies. If a company wishes to move to a cheaper provider later, they may find that their application relies on proprietary services that do not exist elsewhere, forcing them to rebuild significant portions of their infrastructure.
BetaKit cites the recent bankruptcy of Character.ai as a cautionary example of these economics. Despite efforts to reduce costs—including adopting open-weight models and doubling inference throughput—the company struggled with sustainability. According to bankruptcy filings, Character.ai spent approximately $10 million USD ($14 million CAD) per month on compute in 2025.
Pennarun compares this trend to the evolution of AWS, which expanded from basic S3 storage and EC2 compute to specialized services like DynamoDB and Lambda. While these tools offer efficiency, they tie customers to specific APIs. Pennarun suggests that true "sovereign AI" requires more than just keeping data within national borders; it requires the ability to change providers without being forced into a single vendor's entire technology stack.