The Cloud Wars:
Power, Profit & the
$2 Trillion Gamble

Before understanding why Oracle was late to the cloud and Amazon wasn't, you need to understand why a database company and a cloud provider are fundamentally different creatures — even though databases sit at the heart of every cloud.

A database server is software and hardware purpose-built to store, retrieve, and manage structured data. It answers the question: "Where does the data live, and how do I query it?" Oracle built its empire on exactly this — the relational database, the backbone of corporate computing since the 1970s. Companies bought Oracle licenses, ran Oracle on their own servers, and managed it all in-house. Oracle's customer was the enterprise IT department.

A cloud services platform is an entirely different beast. It asks a different question: "What if a company didn't need to own or manage any physical infrastructure at all?" Cloud providers rent compute, storage, networking, databases, and dozens of other services on-demand, paid by the minute. The cloud provider builds and operates the infrastructure at massive scale and passes the economic benefits of that scale — and the flexibility — to customers. The customer is now everyone: startups, governments, Fortune 500s, developers with a credit card.

The Layered Stack: IaaS, PaaS, SaaS

Cloud computing is organized in layers, and understanding them is critical to reading the financials of every company in this report:

The cloud wasn't invented by the company best positioned to invent it. It was invented by a company with a problem no one else had yet faced — and the boldness to monetize the solution.

Market share in cloud infrastructure is measured by spending on IaaS and PaaS services. The Big 3 — AWS, Azure, and GCP — have dominated for a decade, but the competitive dynamics are finally shifting in meaningful ways.

Cloud revenue is the most closely watched number in big tech. Here's how each major player has grown — and what the trajectory actually tells us.

The Individual Stories

We are witnessing the largest peacetime capital investment cycle in human history. Combined hyperscaler CapEx — Amazon, Microsoft, Google, Meta, and Oracle — is projected to reach $443 billion in 2025 and $600 billion in 2026. To understand why, you have to understand what they're actually buying.

What Are They Actually Building?

The CapEx is being deployed across three categories, in rough order of dollar magnitude:

1. Data Center Construction & Expansion (~40-50%): Physical buildings, land, and real estate. Amazon, Microsoft, and Google are building data centers across every major continent. The number of hyperscale data centers has nearly tripled since 2018 — from ~450 to nearly 1,300 globally (Synergy Research Group). Power infrastructure alone — electrical substations, generators, cooling systems — can cost as much as the compute hardware inside.

2. AI/GPU Hardware (~30-40%): Primarily Nvidia GPU clusters (H100, B100, B200 Blackwell), but also custom AI chips: AWS's Trainium/Inferentia, Google's TPUs, Microsoft's Maia AI accelerator. These chips cost $30,000–$40,000 each and consume enormous power. AWS reportedly ordered over 400,000 Nvidia H100 GPUs in 2024 alone.

3. Networking & Storage (~15-20%): High-speed interconnects between data centers, fiber optic cables (sometimes self-built undersea cables), and storage arrays. Google has its own global private fiber network. AWS has submarine cable investments across the Atlantic and Pacific.

Cloud margins tell two stories: the extraordinary profitability of mature cloud infrastructure, and the short-term margin pressure of the current investment cycle. Knowing which story each company is in is critical for investors.

The ROI Question

Historical cloud ROI has been extraordinary. Consider AWS: the service was essentially built by reusing infrastructure Amazon already owned and operated for its retail business. The original capital cost was minimal. By 2016 — ten years after launch — AWS generated $12.2B at 25% margins. Today it runs at $124B annual run rate with ~38% operating margins. The IRR on that original investment is incalculable in the traditional sense — it's simply one of the best capital allocation decisions ever made.

Azure's story is different: Microsoft invested heavily in Azure from ~2010, and the division lost money for years. But the strategic leverage was enormous — Azure kept Microsoft relevant in the enterprise and created a platform for AI deployment that now rivals Google (with a $13B OpenAI stake that has appreciated enormously). Microsoft's management (Nadella-era) has proven extraordinarily effective at capital allocation: the acquisition of LinkedIn, GitHub, Nuance, and the OpenAI investment have all delivered strategic ROI even when financial ROI was initially unclear.

Expected ROI on the Current CapEx Cycle

This is the central analytical challenge. Here's how to think about it by company:

Oracle's cloud story is perhaps the most instructive case study in tech history — both of the dangers of incumbency bias and of the value of a defensible moat when the market finally turns your way.

Why Oracle Was Late

Oracle's core business — database licensing — was extraordinarily profitable. In 2010, Oracle generated over $3B per year in software license updates and support revenue at near-100% gross margins. Cannibalizing that with a cloud model (where customers pay monthly and can cancel) would have been economically irrational in the short term. This is the classic innovator's dilemma: the incentive structure of a dominant business makes it rational to defend rather than disrupt.

Larry Ellison compounded this with ideological resistance. He publicly mocked cloud computing in 2008 and 2009, calling it "water vapor" and "the latest fashion." This wasn't ignorance — Ellison is one of the most technically sophisticated CEOs in history. It was denial, possibly strategic, to protect the installed base while building OCI in secret.

OCI's Genuine Differentiation

When OCI Gen 2 launched in 2016, it did something different: it was built for enterprise workloads from the ground up. AWS and Azure had evolved from developer-first products and had accumulated years of architectural debt. OCI offered a flat-rate, consistent pricing model, superior networking architecture (dedicated physical NICs, not virtualized), and — critically — every OCI region is identical in capabilities. This is unusual: AWS regions have different service availability, which frustrates enterprise architects managing global compliance requirements.

The AI Wildcard: Stargate & OpenAI

Oracle is a founding member of Stargate — the $500 billion AI infrastructure consortium with SoftBank, OpenAI, and others. OpenAI trains ChatGPT models on OCI. This gives Oracle legitimacy in AI infrastructure that it never had in general cloud. Larry Ellison's CTO relationships and Oracle's database expertise in handling structured data for AI RAG (Retrieval-Augmented Generation) applications makes OCI a natural fit for AI-era enterprise deployments.

The multicloud database revenue from Amazon, Google, and Azure grew 1,529% YoY in Q1 FY2026 (Oracle fiscal Q1, August 2025). This is small in absolute terms but signals a genuine wave of migration from on-premises Oracle databases to cloud-based Oracle databases — a migration Oracle effectively delayed for a decade.

Not all cloud revenue is created equal. The way companies define, segment, and report their cloud businesses varies enormously — and some of those choices are strategically advantageous. Here's our forensic analysis.

China's cloud market is not an afterthought — it's roughly 10% of global cloud spending ($10B+ per quarter) and growing fast. But it operates under entirely different rules, with different players, different regulations, and an entirely different geopolitical context.

Alibaba Cloud (BABA) — The Wounded Giant

Alibaba Cloud's global market share has declined from ~6% in 2020 to ~4% by 2024 — a casualty of geopolitical tensions, US chip export restrictions (limiting access to Nvidia H20 and other advanced GPUs), and a slowing domestic economy. Yet within China, Alibaba remains dominant with 33–36% market share.

The AI opportunity is real: Alibaba's Qwen AI models are among the most capable open-source models globally. AI product revenue has maintained triple-digit growth for seven consecutive quarters. The company pledged CNY 380 billion ($52.6B) in cloud and AI infrastructure investment over three years (2025–2027) — exceeding its ten-year prior investment total. This is an existential bet. Alibaba Cloud revenue was $4.15B in Q1 2025 (+18% YoY) — modest growth by hyperscaler standards, but accelerating.

Tencent Cloud — The Opaque Player

Tencent's cloud business is uniquely difficult to analyze because the company deliberately buries it inside its "Fintech and Business Services" segment. This is not an accident — Tencent's competitors and investors cannot cleanly benchmark it. What we know: AI-related cloud revenue almost quadrupled in 2024. GPU revenue now represents "teens-percentage" of IaaS revenue. The Hunyuan LLM and its DeepSeek integration have been commercially significant. Tencent's CapEx surged 421% YoY in Q4 2024, signaling a dramatic acceleration in infrastructure investment.

This is the most important analytical question in all of investing right now. Why are the largest companies in the world simultaneously committing to trillion-dollar infrastructure buildouts, and what does it really mean for you?

Theory 1: They Need It for AI — And AI Is Real

The most straightforward explanation: training frontier AI models requires extraordinary compute. GPT-4 reportedly required ~$100M in compute to train. The next generation will cost more. Inference (serving AI models to millions of users simultaneously) requires continuous, massive compute. The demand from enterprises integrating AI into their workflows is genuinely accelerating — cloud spend grew 25% YoY in Q3 2025, the fifth consecutive quarter above 20% growth.

Under this theory, hyperscalers are simply meeting real demand. The $443B in 2025 CapEx is justified because the market will grow to $1.76 trillion by 2029 (per multiple analyst projections) and whoever owns the infrastructure owns the revenue.

Theory 2: This Is an Infrastructure Land Grab

The more strategic view: hyperscalers aren't just meeting demand. They're creating demand by building infrastructure that makes certain futures possible and others impossible. By deploying 1,300 data centers across 200+ countries, by signing 20-year power purchase agreements with utilities, by cornering the supply of advanced GPUs — they are making it structurally difficult for competitors to enter. This is moat-building at civilizational scale.

Theory 3: They're Replacing SaaS With AI Agents — And Vertically Integrating Everything

Consider what Microsoft is doing: they invested $13B in OpenAI, built Copilot into every Microsoft 365 product, and are now selling AI agents that can autonomously operate enterprise software. The endgame isn't just to provide compute for AI — it's to deliver AI as the software layer itself, making traditional SaaS vendors (Salesforce, ServiceNow, Workday) increasingly irrelevant. Azure becomes the substrate for a Microsoft-controlled AI economy.

Amazon is pursuing something similar with Alexa+, Amazon Bedrock (which hosts Claude, GPT, Llama), and AWS AI services. Google has Gemini embedded in Workspace and GCP. Each hyperscaler is building a vertically integrated stack: silicon → data center → cloud → AI model → AI application → end user. If they succeed, they don't just win infrastructure — they win the application layer too.

Theory 4: The Rental Economy — Owning Your Compute vs. Renting It

There is a fourth, more unsettling possibility. Right now, the cloud computing model requires that you rent compute from hyperscalers. Your data lives on their servers. Your AI model runs on their GPUs. Your business logic executes in their data centers. The alternative — running powerful AI locally on personal hardware — has been limited by the fact that training and inference for large models requires enormous compute.

But models are getting dramatically more efficient (DeepSeek R1 showed 95% cost reduction vs. OpenAI for comparable tasks). Edge AI hardware is improving. Apple Silicon M-series chips can run meaningful local LLMs. The potential threat to the cloud rental model is local AI — where your personal device handles inference privately and securely, without sending data to a hyperscaler's server.

Theory 5: It's the Dotcom Bubble, Part 2

The most bearish view: the AI infrastructure spending is a capital cycle that will end badly. The fiber optic boom of 1998–2001 destroyed $1 trillion in investor value despite the internet ultimately succeeding as a technology. The same dynamic could apply here: AI succeeds as a technology, but the infrastructure buildout overshoots demand by a factor of 3–5x, GPU prices collapse, data center utilization falls to 40%, and hyperscaler CapEx falls sharply in 2027–2028 — creating a severe earnings recession even as AI adoption continues.

Goldman Sachs and tech strategist Jac Arbour have both flagged this risk explicitly: "By 2026, investors will need to see tangible earnings that justify those investments." AI-related services are expected to deliver only ~$25B in revenue in 2025 against $443B in infrastructure investment — a ratio that implies a 20+ year payback at current trajectories. Something has to give: either AI monetization accelerates dramatically, or CapEx will fall.

The Questions Every Investor Should Be Asking

1. Can hyperscalers monetize AI fast enough to justify CapEx at 94% of operating cash flows? Watch quarterly AI revenue disclosures (AWS AI revenue, Azure AI services growth) as the key leading indicator.

2. Will Microsoft's Azure bundling strategy eventually face analyst or regulatory scrutiny that forces a breakout disclosure? If it does, Azure's "pure" IaaS share numbers could be a significant negative surprise.

3. Is the DeepSeek efficiency revolution a sign that AI compute costs will fall faster than expected — making current infrastructure investments partially redundant — or will lower inference costs simply drive 10x more demand?

4. When does the local AI threat become real? When Apple M-series and Qualcomm Snapdragon can run GPT-4-equivalent models on-device, does the cloud rental model face genuine disruption?

5. Is Oracle's $455B RPO real? Watch how much converts to revenue in fiscal 2026 (expected ~$45B recognized) as the key validation test of management credibility.