Why your next cloud bill could be a trap

A few months ago, I worked with a global manufacturer that considered itself conservative on AI. They focused on stabilizing their ERP migration to the cloud, modernizing a few key customer-facing apps, and tightening security. The CIO’s position on generative AI was clear: “We’ll get there, but not this year. We’re not ready.”

On paper, they were officially “not doing AI.” In reality, they were already deeply involved. Their primary cloud provider had quietly integrated AI-native features into the services they were currently using. The search service they adopted for a new customer portal came with semantic and vector modes turned on by default. Their observability platform was now AI-assisted, altering logs and telemetry processing. Even their database service had a new “AI integration” checkbox in the console, which developers began enabling because it appeared useful and was inexpensive to try.

Six months later, their infrastructure bill had risen sharply, and their architecture had become so integrated with that provider’s vector engine and AI tools that shifting away had become dramatically harder. Key data stores were now optimized around that provider’s vector engine. Workflows were wired into proprietary AI agents and automation tools. The CIO’s team woke up to a hard truth: They had unintentionally become an AI-focused organization, more locked in than ever.

Whether you asked for it or not

For years, we have talked about cloud-first strategies, with the big hyperscalers competing on compute, storage, databases, and global reach. Generative AI changed the game. The center of gravity is shifting from generic infrastructure to AI-native platforms: GPUs, proprietary foundation models, vector databases, agent frameworks, copilots, and AI-integrated everything.

You can see the shift in how providers talk about themselves. Earnings calls now highlight GPU and AI accelerator spending as the new core investment. Homepages and conferences lead with AI platforms, copilots, and agentic AI, while traditional IaaS and PaaS take a back seat. Databases, developer tools, workflow engines, and integration services are all being refactored or wrapped with AI capabilities that are enabled by default or just a click away.

At first glance, this appears to be progress. You see more intelligent search, auto-generated code, anomaly detection, predictive insights, and AI assistants integrated into every console. However, behind the scenes, each of these conveniences typically relies on proprietary APIs, opinionated data formats, and a growing assumption that your workloads and data will stay within that cloud.

A bigger problem than you realize

Lock-in is not new. We have always had to balance managed services with portability. The difference now is the depth and systemic nature of AI-native lock-in. When you couple your workloads to a provider’s proprietary database, you can often extract the data and re-platform with effort. When you couple your entire data platform, embeddings, fine-tuned models, agent workflows, and security posture to a single AI stack, the cost and time to exit increase by an order of magnitude.

Training and inference pipelines are expensive to rebuild. Vector indexes and embeddings may be tied to a provider’s specific implementation. Agent frameworks are increasingly integrated with that cloud’s eventing, identity, and security systems. Once you start relying on a provider’s proprietary model behavior and tool ecosystem, you are no longer just “using compute.” You are buying into their approach to AI.

What worries me most is that many enterprises are drifting into this locked-in position rather than choosing it. Teams turn on AI-native features because they come bundled with existing services. Line-of-business units experiment with AI assistants hooked into core data without a broader architectural or financial strategy. Over a few release cycles, the outlook moves from just experimenting to “We can’t move off this platform without a multi-year, multi-million-dollar transformation.”

What ‘AI-ready’ really means

Providers market their platforms as “AI-ready,” implying flexibility and modernization. In practice, “AI-ready” often means “AI–deeply embedded” into your data, tools, and runtime environment. Your logs are now processed through their AI analytics. Your application telemetry routes through their AI-based observability. Your customer data is indexed for their vector search.

This is convenient in the short term. In the long term, it shifts power. The more AI-native services you consume from a single hyperscaler, the more they shape your architecture and your economics. You become less likely to adopt open source models, alternative GPU clouds, or sovereign and private clouds that might be a better fit for specific workloads. You are more likely to accept rate changes, technical limits, and road maps that may not align with your interests, simply because unwinding that dependency is too painful.

The rise of alt clouds is a signal

While hyperscalers race to become vertically integrated AI platforms, we are also seeing the emergence of alternative clouds. These include GPU-first providers, specialized AI infrastructure platforms, sovereign and industry-specific clouds, and environments run by managed service providers. These alt clouds are not always trying to be “AI everything.” In many cases, they prioritize providing raw GPU capacity, clearer economics, or environments where compliance, data residency, or control are the main value propositions.

For companies not prepared to fully commit to AI-native services from a single hyperscaler or in search of a backup option, these alternatives matter. They can host models under your control, support open ecosystems, or serve as a landing zone for workloads you might eventually relocate from a hyperscaler. However, maintaining this flexibility requires avoiding the strong influence of deeply integrated, proprietary AI stacks from the start.

Three moves to stay in control

First, be deliberate about where and how you adopt AI-native services. Don’t let free trials or default settings define your architectural strategy. For each major AI-integrated service a provider pushes—a vector database, agent framework, copilot, or AI search—ask explicitly: What will it cost us to switch later? What data formats, APIs, and operational dependencies does this introduce, and how difficult will it be to replicate them with another provider, an alt cloud, or a self-managed stack?

Second, design your AI and data strategy from the start with portability in mind, even if you don’t plan to move soon. Use open formats for embeddings whenever possible, store raw data in portable structures, and separate application logic from proprietary AI orchestration. When evaluating AI services, consider alternatives such as open source models, GPU-first alt clouds, or private and sovereign clouds that avoid a single provider’s AI ecosystem. It’s entirely reasonable to move some workloads away from providers that are heavily focused on AI if their AI services do not align with your current or upcoming needs.

Third, prioritize AI costs and dependency as key governance issues alongside security and compliance. Incorporate observability into AI deployment to track which teams enable AI-native features, understand how these affect costs, and identify long-term platform risks. Before choosing a cloud provider that is rapidly shifting toward AI-native solutions, step back and ask if their AI services truly match the problems you need to solve over the next three to five years. If AI is on your radar but not yet essential, consider a more neutral infrastructure approach and selective AI implementation rather than adopting every new AI-native feature your provider offers.

The bottom line is simple: AI-native cloud is coming, and in many ways, it’s already here. The question is not whether you will use AI in the cloud, but how much control you will retain over its cost, architecture, and strategic direction. Enterprises that pose tough questions now, focus on portability, and maintain real options across hyperscalers, alt clouds, and private environments will turn AI into a strategic advantage instead of a costly pitfall.