The Cloud Wars: AWS, Azure, GCP, and the Infrastructure Battle

Zusammenfassung

In 2006, Amazon launched a service that let anyone rent computing resources by the hour. The idea seemed modest — a side business for a bookseller. Nineteen years later, Amazon Web Services generated $107 billion in revenue annually and accounted for most of Amazon’s profit. Microsoft Azure and Google Cloud Platform had grown into multi-hundred-billion-dollar businesses. The three largest technology companies in the world derived enormous fractions of their enterprise value not from consumer products but from the infrastructure that every other company’s software ran on. The cloud transformed the economics of software development, made global-scale deployment available to two-person startups, and shifted competitive advantage from hardware ownership to operational scale. It also concentrated the infrastructure of the internet into three companies — a concentration with implications for privacy, sovereignty, and the architecture of digital power that are still being negotiated.

The Pre-Cloud Landscape: Buying Servers

Before cloud computing, deploying software at scale meant buying servers. A startup expecting significant traffic purchased or leased physical hardware, placed it in a data center, configured networking, installed operating systems and middleware, and hoped the hardware could handle the load. This approach had several problems:

Capacity planning was guesswork. Buying for peak load meant paying for idle capacity during normal operations. Buying for normal load meant failing during peaks. The 2004 Democratic National Convention’s website failed because the candidate John Kerry’s team had not anticipated the traffic spike after his nomination; the Slashdot Effect — a popular post linking to a small site would crash it within minutes — was a routine hazard.

Lead time was weeks to months. Ordering servers, receiving them, racking and cabling them, and configuring them took time that a rapidly growing startup did not have. Capital requirements were large: a few thousand servers cost millions of dollars in hardware alone.

Engineering attention was consumed by infrastructure. A team that should be writing product code spent weeks managing databases, load balancers, and storage systems.

Amazon had solved these problems internally. By 2003, Amazon’s engineering team had rebuilt its infrastructure on standardized, virtualized services — computing, storage, database, messaging — accessible through internal APIs. Every Amazon team could request capacity by calling an API rather than filing a hardware request. The infrastructure had become programmable.

Andy Jassy, who ran Amazon’s technology infrastructure, recognized that this internal capability had commercial value. The same virtualized infrastructure services that Amazon’s own engineering teams used could be offered externally.

Amazon Web Services: The Accidental Infrastructure Giant

Amazon launched its first cloud services in 2006: S3 (Simple Storage Service, March 2006) for object storage and EC2 (Elastic Compute Cloud, August 2006) for virtual machine instances. The pricing model was revolutionary: pay by the hour, stop when finished, no minimum commitment.

The initial customers were startups. Dropbox launched in 2007 on S3. Airbnb built on AWS from its 2008 founding. Instagram ran on AWS through its hypergrowth to 100 million users. These companies could have launched on owned infrastructure — but AWS allowed them to do so without upfront capital, scale dynamically as they grew, and focus engineering attention on their products rather than their servers.

AWS’s S3 and EC2 pricing established the unit economics of cloud computing: $0.15 per GB-month of storage, $0.10 per compute-hour for a small instance. The prices fell continuously — every year, AWS reduced prices, reinvesting scale economics into lower costs. By 2023, S3 storage cost approximately $0.023 per GB-month — a 6× reduction over seventeen years. AWS had grown to 200+ services covering every layer of computing infrastructure, from virtual machines (EC2) to managed databases (RDS, DynamoDB) to machine learning (SageMaker) to satellite communications (AWS Ground Station).

AWS’s competitive advantage was not merely first-mover. Amazon invested the cash flows from its retail and AWS businesses in building data centers at a pace competitors could not match without equivalent capital commitment. By 2024, AWS operated 33 geographic regions, each consisting of multiple physically separate data centers (Availability Zones), providing redundancy guarantees that customers could not replicate themselves.

AWS and Amazon’s Profits

AWS generates extraordinary margins relative to Amazon’s retail business. In 2023, AWS represented approximately 17% of Amazon’s revenue but the vast majority of its operating income. Amazon’s retail operations operate on thin margins; AWS’s infrastructure services, once the data center is built, generate software-level margins on hardware-like costs. Jeff Bezos described AWS as a once-in-a-generation business opportunity; in practice it allowed Amazon to cross-subsidize its retail expansion and Prime growth for years.

Microsoft Azure: The Enterprise Counter-Attack

Microsoft was late to cloud computing, which was surprising given its infrastructure assets and enterprise relationships. Steve Ballmer dismissed cloud computing through most of the 2000s; Microsoft’s cloud bet was Windows Azure, launched in February 2010, built on Windows Server virtualization and positioned as a natural extension of the Windows ecosystem.

Azure’s initial positioning — “Windows in the cloud” — was both its advantage and its limitation. Enterprise customers running Windows Server workloads could migrate to Azure with minimal changes. The limitation was that the cloud market was not exclusively Windows; Linux workloads, open-source databases, and cloud-native development favored platforms agnostic to Microsoft’s stack.

The transformation came under Satya Nadella, who became CEO in 2014. Nadella reoriented Microsoft toward “mobile-first, cloud-first” and made Azure the center of the company’s strategy. Critically, he abandoned Microsoft’s historical hostility to Linux: Azure became the cloud with the best Linux support among the major providers, and Microsoft joined the Linux Foundation, contributed to the Linux kernel, and acquired GitHub (the center of the open-source developer community) in 2018.

Azure’s competitive advantage was its enterprise sales force: Microsoft had decades-long relationships with IT departments at large corporations, government agencies, and regulated industries. An enterprise that already paid Microsoft for Office 365, Azure Active Directory, and SQL Server licenses could purchase Azure services through existing procurement relationships without a new vendor qualification process. The Azure + Microsoft 365 bundle became the dominant enterprise productivity and infrastructure package, with Azure growing from essentially zero in 2010 into one of the world’s two largest cloud platforms, the core of a Microsoft Cloud business that reported roughly $137 billion in revenue in fiscal 2024.

Azure Government — a physically separate cloud infrastructure meeting US government security requirements — captured significant federal contracts, including a disputed $10 billion JEDI (Joint Enterprise Defense Infrastructure) contract initially awarded to Microsoft over Amazon in 2019 (later cancelled after years of litigation from Amazon).

Google Cloud Platform: The Technical Underdog

Google had invented many of the foundational technologies of cloud computing: MapReduce (2004), BigTable (2006), Chubby (distributed lock service), and the Borg container orchestration system (which became Kubernetes in its open-source form in 2014) — all described in academic papers that had bootstrapped the cloud computing industry. Amazon and Microsoft had built their cloud services partly by reading Google’s research papers.

Paradoxically, Google was behind AWS and Azure in commercial cloud revenues for years. The reasons were organizational: Google’s culture was engineering-first and enterprise sales was historically deprioritized. Selling to enterprise customers requires long relationship-building, procurement compliance, and support guarantees that Google’s engineering-focused organization resisted building.

Google Cloud Platform (GCP) offered technical capabilities that AWS and Azure matched only later: BigQuery (serverless data warehouse, 2010), Kubernetes Engine (2014), and the best-in-class machine learning infrastructure (TPUs, Vertex AI). But technical superiority did not automatically convert to enterprise revenue. Customers who had built on AWS saw no reason to migrate; customers choosing for the first time defaulted to AWS or Microsoft.

Under Thomas Kurian (CEO of Google Cloud from 2019), GCP built an enterprise sales organization and began winning significant contracts. Google Cloud revenue grew from $8.9 billion in 2019 to $33 billion in 2023 — substantial but still approximately one-third of AWS’s revenue. Google’s investments in Anthropic — beginning with roughly $300 million in 2022 and growing to multi-billion-dollar commitments — and Microsoft’s parallel multi-billion-dollar investment in OpenAI signaled that AI infrastructure was becoming the next competitive battleground.

The Multi-Cloud Reality and Market Structure

By the early 2020s, the enterprise market had largely converged on multi-cloud strategies: organizations using services from multiple cloud providers to avoid vendor lock-in, satisfy data sovereignty requirements, or select best-in-class services for specific workloads.

The cloud market was simultaneously highly concentrated and highly competitive:

Concentrated: AWS, Azure, and GCP together held approximately 65% of global cloud infrastructure market share (AWS ~33%, Azure ~22%, GCP ~11%). The capital requirements for competing at scale — each company invested $40–60 billion annually in data center infrastructure — made new entry by full-service cloud providers essentially impossible.

Competitive: Within the leading three, competition was intense on pricing, service breadth, geographic coverage, and AI capabilities. AWS’s pricing pressure had driven continuous price reductions industry-wide.

The concentration had geopolitical implications. When the European Commission assessed digital sovereignty, the dependency of European governments and businesses on US-headquartered cloud providers was a central concern. GAIA-X — a European cloud initiative — attempted to establish a framework for European cloud sovereignty, with limited success in competing against the technical and economic advantages of the established three. China’s equivalent cloud providers (Alibaba Cloud, Tencent Cloud, Huawei Cloud) dominated the Chinese market behind trade barriers but made limited inroads in markets where US providers operated.

Dead End: The On-Premises Counter-Trend

The cloud’s economics were not uniformly superior to on-premises infrastructure for every workload. For predictable, large-scale compute — the cases where a company needed specific hardware running at high utilization 24 hours a day — buying servers was cheaper than renting them. David Heinemeier Hansson (DHH) of Basecamp published analyses showing that their particular workloads cost significantly more on AWS than on owned hardware; Basecamp repatriated significant portions of their cloud workloads to owned servers.

The cloud repatriation trend was real but selective. The economics of ownership beat the economics of rental only for large-scale, predictable workloads where utilization was high and organizational capability to manage infrastructure existed. For variable-load applications, global distribution requirements, or organizations without infrastructure teams, cloud remained economically superior. The cloud wars had not ended; they had settled into a stable competitive dynamic in which AWS, Azure, and GCP competed for an expanding total market while specific workloads moved in and out based on economics, compliance requirements, and technical fit.