Over the past decade and a half, cloud computing has become a foundational technology. It started as a way to rent servers but has evolved into a complex ecosystem that supports everything from basic infrastructure shifts to transformative AI initiatives. Having advised enterprises on thousands of cloud projects over the years, I have seen that most projects fall into a handful of categories. I can say with certainty that success depends less on hype and more on understanding each project’s nature, risks, costs, and lessons.

Cloud migrations

Enterprises continue to migrate existing workloads from data centers to public, private, or hybrid environments. This can involve rehosting (lift and shift), replatforming with minor changes, or full refactoring into cloud-native architectures. The goal is usually cost reduction, scalability, or the end of hardware refresh cycles. The risks here are well documented. Many projects underestimate dependencies, leading to performance surprises or integration failures. Data egress fees and unexpected operational costs can wipe out projected savings.

Cost profiles vary widely. Initial migrations often run 20% to 50% over budget due to discovery gaps and testing. Ongoing expenses can decline through rightsizing and reserved instances, but poor management often leads to 25% to 35% waste from idle resources. These lessons underscore the importance of modeling the total cost of ownership up front, including people, training, and change management.

What we’ve learned: Pure lift-and-shift rarely delivers the promised ROI. Organizations that succeed treat migration as an opportunity for modernization rather than a simple move. Phased approaches with strong governance and finops practices minimize overruns, which have historically plagued most efforts.

Cloud-native applications

Teams build microservices, serverless functions, or containerized apps on platforms such as Kubernetes, AWS Lambda, or Azure Functions. This approach leverages elasticity, devops pipelines, and managed services to accelerate time to market.

Risks focus on architectural complexity and skills gaps. Overengineering with too many microservices creates operational nightmares, while underengineering leads to unscalable monoliths. Distributed systems need constant security vigilance. New apps often begin well but gain technical debt when teams prioritize features over observability and resilience. Entry costs are usage-based, which sounds attractive, but they often spike at scale due to poor design.

What we’ve learned: Based on my years of observation, successful teams embed cost awareness in CI/CD, use spot instances strategically, and design for observability from day one. Cloud-native development accelerates innovation when paired with disciplined architecture.

Business analytics projects

Enterprises are moving data lakes, data warehouses, and ETL processes to services such as Snowflake, BigQuery, or Redshift. Real-time analytics, dashboards, and predictive modeling become possible at scale. The primary risks are data gravity and quality issues. Moving petabytes is expensive and complex, while poor governance leads to compliance headaches or “garbage in, garbage out” results. Integrating with legacy systems often delays the realization of value.

What we’ve learned: Fifteen years later, we know that centralized data strategies outperform fragmented ones but only when paired with strong data mesh or data fabric approaches that respect domain ownership. Cost profiles include storage, compute for queries, and egress. Optimization through partitioning and materialized views pays off, but many organizations waste money on unused data. Lessons emphasize starting small with high-value use cases and building governance early rather than bolting it on later.

Artificial intelligence projects

Artificial intelligence and machine learning projects represent the current frontier of cloud. This includes training models, deploying inference endpoints, and integrating ML into applications. Managed services lower barriers, but custom needs often require GPU clusters or specialized hardware. Risks are significant: model drift, explainability issues, high compute demands, and ethical concerns. Many projects stall after the proof of concept because production deployment exposes scalability or cost issues. Managed AI offerings from providers help, but enterprises still struggle to integrate them into core business processes.

Costs run high, especially for training. Inference can be optimized, but it often dominates bills. What we have learned is that AI succeeds when treated as part of a broader cloud-native architecture, not as a standalone science project. Hybrid approaches and cost controls are essential.

Generative AI projects focus on large language models, image generation, code assistants, and custom agents using services like Bedrock, OpenAI integrations, or fine-tuned open source models. Enterprises are experimenting with retrieval-augmented generation for grounded responses and agentic workflows. Risks include hallucinations, data privacy leaks, intellectual property issues, and runaway token costs. Many early adopters built impressive demos only to face governance and compliance walls in production.

What we’ve learned: After observing the wave, the lessons are clear. Start with narrow, high-value use cases and layer in strong prompting, evaluation, and human oversight frameworks. Cost profiles are usage-driven and can escalate quickly with volume. Optimization through caching, smaller models, and hybrid on-prem inference helps. Generative AI delivers ROI fastest when embedded in existing workflows rather than used as standalone tools.

Other project types

Modernization of legacy mainframe or monolithic applications falls between migration and new development. Internet of Things (IoT) initiatives use the cloud for device management and edge analytics. Disaster recovery and backup projects leverage the cloud to improve resilience. Edge computing projects move processing closer to users or devices. Compliance-focused sovereign cloud deployments address data residency requirements. Finally, sustainability initiatives focus on reducing carbon footprints by implementing efficient architectures.

What we’ve learned: Each approach carries tailored risks and cost dynamics. Modernization often uncovers hidden dependencies. IoT requires reliable connectivity. Edge computing introduces latency considerations. Lessons across all types highlight the value of multicloud strategies for negotiation leverage and risk diversification, though they increase complexity.

Common themes

Most projects do not fail because of technology itself but from inadequate planning, cultural resistance, or neglect of operational realities. Cost overruns are often caused by the absence of strict finops discipline. Security and compliance issues remain ongoing and require integrated design considerations. Skills shortages hinder progress, which makes managed services appealing despite concerns about vendor lock-in.

Successful cloud stories share common traits: strong executive sponsorship, iterative delivery, cross-functional teams, and continuous optimization. Enterprises that treat the cloud as a business transformation rather than an IT project perform best. They measure outcomes using business metrics, such as revenue impact, customer satisfaction, and speed to market—not just uptime or instance counts.

The cloud landscape continues to evolve as capacity markets, neoclouds, and AI-driven operations offer new options. Yet cloud fundamentals endure. Choose the right project type for your cloud maturity and goals. Understand risks thoroughly. Model costs realistically. Apply lessons from the thousands of cloud deployments that came before.

My advice sounds simple, but it will determine which cloud projects and enterprises will thrive in the next decade of cloud computing. Those who chase hype without discipline will only become another cautionary tale.