A customer can request computing capacity through an API in seconds because a cloud provider has already spent years assembling the infrastructure behind it. Buildings, processors, networks, storage and operating software become a service that someone else can consume on demand. AI expands that opportunity, but also makes the provider’s wager larger: capital must be committed before the full pattern of future usage is known.
The investment question
The cloud investment question is whether AI strengthens platform economics or requires so much additional capital that returns become harder to sustain. Revenue growth is only one part of the answer. Providers must finance capacity, keep it productive, manage hardware transitions and capture enough customer value after competition passes efficiency gains through to prices.
Our view is that the strongest cloud platforms combine infrastructure with data, development tools, security and distribution. Those complements can make computing more useful and deepen customer relationships. Yet AI workloads may also strengthen specialist providers or give large customers more negotiating leverage. The outcome depends on the workload, service layer and capital required to serve it.
How the cloud business works
Infrastructure services expose resources such as virtual machines, storage and networking. Managed platform services take on more operational work, including running databases or application environments. Model services add access to AI capabilities through APIs and related tooling. Customers can combine these layers, and the provider’s responsibility and margin opportunity change with the service chosen.
The distinction is especially important for security. Microsoft’s shared-responsibility guidance explains how responsibilities vary across service models. The provider operates underlying infrastructure, while customers retain obligations for data, identities and the configurations they control. Buying cloud services does not eliminate the need for sound application and data governance.
Physical infrastructure remains regional. A service available in one region or instance type may not be available everywhere with the same capacity. Customers also need suitable latency, data location and resilience arrangements. A global brand is therefore not a guarantee that any workload can be placed anywhere immediately.
Market structure and competitive advantage
| Provider model | Illustrative participants | Main competitive asset |
|---|---|---|
| Broad public cloud | AWS, Microsoft Azure, Google Cloud | Service breadth, customer distribution and operating scale |
| Enterprise and infrastructure cloud | Oracle Cloud | Database relationships, infrastructure and application integration |
| Specialist AI cloud | CoreWeave and other operators | Focused capacity and workload operations |
| Customer-owned infrastructure | Enterprises and large technology users | Control and economics for suitable workloads |
AWS combines a broad infrastructure and services portfolio with custom silicon. Microsoft connects Azure with enterprise software and developer relationships. Google links infrastructure with its data, AI and custom-compute capabilities. Oracle combines database and application relationships with cloud infrastructure. These descriptions identify strategic positions, not a claim that one provider is best for every customer.
Specialists can compete where capacity access or focused operational support is valuable. Large platforms can respond through their service breadth, financial resources and existing commitments with customers. Market share should therefore be examined at the relevant workload layer rather than inferred from a provider’s overall cloud ranking.
Economics: growth, depreciation and cash
Cloud revenue depends on consumed resources, contracted commitments, discounts and service mix. Some customers buy flexibility; others exchange commitment for better pricing. A contract can improve visibility without producing immediate consumption. Remaining performance obligations and similar backlog measures differ by company and should not be compared without understanding their definitions.
The provider funds infrastructure ahead of use. Depreciation spreads asset cost over an accounting life, while capital expenditure records investment cash flows. Changes in estimated useful lives can affect reported margins without changing the underlying cash already spent. Investors should compare operating profit with capital requirements and the economic competitiveness of the installed fleet.
Illustratively, a service generating 100 of revenue and 30 of operating profit can still consume cash if growth investment materially exceeds depreciation and operating cash generation. That does not automatically make the investment unattractive: new capacity may create future returns. It does mean current margins cannot establish that the growth is self-funding or that incremental returns are adequate.
Service mix matters too. A managed database or software service can have different economics from renting an accelerator. Customers may move up the service stack because it reduces their engineering burden. The provider must then deliver reliability, security and productivity that justify the additional charge rather than assuming every higher-level service earns premium margins automatically.
AI and hyperscalers: competing across the stack
Hyperscalers buy merchant accelerators and develop custom chips, while also offering model platforms and using AI in their own applications. They can optimise across these layers and accept lower returns in one area if it improves the broader business. This makes component-level competition difficult to interpret without understanding the platform’s overall strategy.
Custom silicon can lower costs for suitable workloads and strengthen negotiating leverage. Its success depends on software support, customer adoption and enough volume to amortise development. Merchant platforms remain valuable where flexibility and broad compatibility matter. A provider’s choice can vary between internal workloads and the services it offers external customers.
Data is another source of advantage. Applications often run near their databases, storage and governance systems because moving data adds cost and complexity. That gives cloud providers opportunities to attach AI services to existing workloads. Independent data platforms can also benefit by helping customers operate across clouds or avoid tying every layer to one provider.
Current market debates — September 2026
Amazon’s second-quarter 2026 results reported AWS growth of 36.7% year over year. Microsoft’s fiscal fourth-quarter release and Alphabet’s second-quarter filing provide complementary evidence from other major platforms. Reporting definitions differ: AWS, Azure-related disclosures and Google Cloud are not identical business boundaries.
The immediate debate is whether accelerating demand translates into attractive returns on the much larger infrastructure base. The constructive case is that capacity constraints ease into sustained consumption and higher-value services. The countercase is that depreciation, energy, financing and price competition absorb more of the growth than expected. Follow cash generation and the marginal cost of serving new workloads alongside headline revenue.
Oracle’s June 2026 results highlight infrastructure and application growth, while CoreWeave’s second-quarter results provide a specialist comparison. The relevant distinction is how each business finances, commissions and monetises capacity. Large contracts are valuable only to the extent that they convert into profitable service delivery and collectible cash.
Structural debates: portability and platform power
Multi-cloud strategies can reduce dependence on one provider, but they also introduce operational complexity. Portability requires more than running the same application container: databases, identities, networking, monitoring and commercial commitments all matter. Customers may use several clouds while still maintaining significant switching costs within each workload.
Open models and frameworks can make parts of AI more portable. Providers can respond by competing on price, performance and operations, or by attaching differentiated data and application services. The long-run question is which layer remains difficult to substitute after model access becomes more widely available.
Another structural issue is the relationship between provider investment and customer financing. Cloud commitments can support model developers and large deployments, but concentration and reciprocal commercial relationships require careful interpretation. Evaluate the ultimate source of demand, contract terms and cash collection rather than treating every announced agreement as independent proof of market expansion.
What to watch
Track consumption growth, capital expenditure, depreciation, service mix, customer concentration and cash returns. Compare contract announcements with operational capacity and revenue recognition. Monitor whether customers expand into managed data and AI services or primarily seek lower-cost infrastructure.
The strongest cloud franchise turns physical capacity into a service customers repeatedly choose because it improves their economics and reduces operational burden. AI increases the opportunity, while making disciplined capital allocation more important.
Explore this sector
AI infrastructure & data centres — sector overview
Related sectors: Semiconductors & chipmaking · Data platforms & analytics
