The cloud is a metaphor; the providers are very real. A handful of companies operate the planet-scale data centers that most of the world’s software now runs on. And choosing among them is one of the stickiest decisions in modern architecture.
📌 TL;DR: CSPs operate hyperscale infrastructure and rent it as services: compute, storage, databases, ML platforms. The big three (AWS, Microsoft Azure, Google Cloud) cover similar ground with different strengths. The strategic question isn't features; it's dependence: how portable is your stack if terms change?
What Are Cloud Service Providers?
Cloud service providers are companies that operate large-scale computing infrastructure and rent it as on-demand services. They run the physical layer of cloud computing (data centers, networks, hardware) and package it as everything from raw virtual machines to fully managed databases and AI platforms.
How do the major providers differ?
Amazon Web Services pioneered the market and offers the broadest service catalog. Microsoft Azure leads where organizations already live in Microsoft’s ecosystem. Google Cloud trades on data and machine-learning tooling. Beyond the big three, regional and specialist providers compete on locality, compliance, and price.
For most data teams, the honest differences are narrower than the marketing: all three run warehouses, pipelines, and ML capably. The deciding factors are usually existing skills, existing contracts, and where the rest of the company already is.
The lock-in question
- Managed convenience has a cost: the more provider-specific services you adopt, the harder leaving becomes
- Egress is the moat: moving data OUT of a cloud costs real money at scale; data gravity is a business model
- Portability is a spectrum: some teams pay the CSP-agnostic tax for independence; others accept lock-in as the price of speed
- The answer is per-workload: commodity components stay portable cheaply; deeply managed services rarely do
Choosing a CSP: the evaluation that actually predicts satisfaction
Feature matrices converge; these differentiators don’t. Skills gravity: the provider your team already knows ships faster for a year than the marginally better one nobody knows, because switching costs include every runbook and reflex. Ecosystem fit: deep Microsoft estates integrate naturally with Azure; analytics-first shops lean Google; the broadest service catalog and job market favor AWS. Region and compliance coverage: where are the data centers, which certifications matter to YOUR auditors, and does data residency work for your markets? Pricing shape at YOUR workload: list prices mislead; model your actual compute, storage, and egress against each provider’s structure. Differences of 2x on the same workload are routine.
And negotiate: at any real scale, cloud pricing is a conversation, not a rate card. Committed-spend discounts, credits, and migration support are all on the table for the asking.
The dependency management playbook
- Tier your services by portability: commodity compute and object storage move easily; proprietary databases and ML platforms don’t; know which tier each workload sits in BEFORE adopting
- Keep data in open formats: the data outlives every platform decision; open table formats and standard exports keep the exit door oiled
- Contain the proprietary blast radius: wrap provider-specific services behind your own interfaces where cheap; CSP-agnostic patterns where the independence justifies its tax
- Price the exit annually: a once-a-year estimate of ‘what would leaving cost?’ converts lock-in from vague dread into a managed number
Beyond the big three
The hyperscalers dominate but don’t exhaust the market. Regional and sovereign providers win where data-residency law or public-sector rules demand local control. Specialist clouds compete on focused workloads: GPU fleets for ML, bare-metal performance, or VPS-style simplicity at flat prices. And the private-cloud vendors serve estates that must stay on-premise while still wanting cloud ergonomics. The pattern for buyers: hyperscaler as the default center of gravity, specialists where a specific workload’s economics or rules say otherwise.
Real-World Examples
The abstract trade-offs get clearer with faces on them. So here are four setups I see over and over.
The all-in startup. A ten-person team goes fully managed on one hyperscaler. Lock-in? Sure. But they ship in weeks, and speed is worth more than portability they’d never use.
The Microsoft estate. An enterprise running on Office, Teams, and Active Directory lands on Azure. Not because the feature matrix won. Because the identity layer and the contracts were already there.
The split ML shop. A data team trains models on a specialist GPU cloud, then serves them from a hyperscaler. Each workload sits where its economics make sense.
The regulated firm. A bank keeps customer data with a regional provider for residency law, and runs everything else on a hyperscaler. Compliance drew that map, not preference.
Reading a CSP relationship like a contract
Past the technology, a CSP relationship is a long-term commercial arrangement with clauses worth understanding before they matter. Service level agreements: what uptime is promised per service, what the remedy actually is (usually credits, rarely meaningful), and which of YOUR obligations void it. Support tiers: response times and escalation paths are bought separately, and the gap between free-tier and enterprise support becomes visible at the worst moments. Price-change mechanics: how much notice, on which services, and what committed-use agreements protect against.
And the compliance inheritance: the provider’s certifications cover their layer; your auditors will ask how you verified them (audit-report access is part of the contract) and what compensating controls cover your layer. Treat the annual contract review with the seriousness of a vendor renewal, because at cloud spend levels, it is one of the largest vendor relationships the company has, and the shared-responsibility line it encodes is the one your incident postmortems will quote.
One relationship habit that pays
Assign the CSP relationship an internal owner: someone tracking spend against commitments, service deprecation notices, new-region openings, and the quarterly conversation with the account team. Providers ship changes constantly; the owned relationship converts that stream from ambient noise into managed input, and captures the credits and programs that unowned accounts simply never hear about.
Common Mistakes
Cloud decisions go wrong in the same handful of ways. And most of them are avoidable with one honest afternoon.
- Choosing by feature matrix. The checklists converge. Your team’s skills, your contracts, and your workload’s pricing shape don’t.
- Discovering egress fees during the migration quote. Data goes in cheap and comes out expensive. Price the exit before you enter.
- Going multi-cloud everywhere on day one. Running every workload on two providers doubles your operational surface. It rarely halves your risk.
- Leaving spend unowned. Without a named owner, commitments drift, credits expire, and the bill becomes a quarterly surprise.
- Treating list prices as final. At real scale, pricing is a negotiation. Teams that don’t ask simply pay more.
- Forgetting the shared-responsibility line. The provider secures the infrastructure. Your misconfigured storage bucket is still yours.
None of these are exotic. But every one of them shows up in postmortems, usually with a large number attached.
Frequently Asked Questions
What is a cloud service provider?
A company that operates large-scale computing infrastructure and rents it as on-demand services, the platform most modern software runs on. AWS, Microsoft Azure, and Google Cloud are the dominant three.
What do CSPs actually provide?
Everything from raw virtual machines and storage to managed databases, data platforms, and machine-learning services. The catalog spans the whole IaaS-PaaS-SaaS spectrum.
How do you choose a cloud provider?
Existing skills, existing ecosystem, regional and compliance needs, and the workloads you actually run, tested against real pricing. Feature checklists differentiate less than the marketing suggests.
How do you compare cloud providers on cost?
Model YOUR workload (compute hours, storage, and especially egress) against each provider’s pricing structure, then negotiate; list prices are the start of the conversation. Identical workloads routinely differ 2x across providers before discounts.
What is vendor lock-in in cloud computing?
The accumulated cost of leaving: proprietary services to rebuild, data egress to pay, and skills to retrain. Manage it by tiering workloads by portability, keeping data in open formats, and pricing the exit annually, not by avoiding managed services entirely.
What should you review annually in a cloud contract?
Committed-spend terms against actual usage, SLA remedies, support-tier fit, price-change notices, and audit-report access for compliance. At real scale it’s among the company’s largest vendor relationships, so review it like one.