Executive Intelligence · Special Report · November 15, 2026
Central thesis. The next cloud wars will not be decided by software platforms alone. They will increasingly depend on who controls computing capacity, proprietary chips, energy, data centers, networking, and the capital required to operate them. AI is not creating an entirely new cloud industry; it is changing the economics of the existing one so fundamentally that the old competitive logic no longer applies.
1. The new cloud infrastructure race
For fifteen years, cloud computing meant renting commodity servers at high margins. Generative AI and inference are changing the shape of demand. AI workloads run on dense clusters of accelerators that draw several times the power per rack, and serving a model to millions of users is a continuous load that grows with adoption. The third quarter of 2026 shows the scale: Microsoft, Amazon, Alphabet and Meta together spent about $203 billion on capital projects in a single quarter, more than double the same quarter a year earlier, according to an analysis of their SEC filings.
Demand is real. Nvidia reported data center revenue of $89.0 billion for its quarter ended July 26, up 117% year over year, and its CEO said multiple frontier labs are now scaling in parallel. Three resources have become strategic in a way they were not in the software-led cloud era:
- Compute capacity. Microsoft’s CFO said customer demand continues to exceed available supply even as Azure grew 48%. Equinix’s CEO told investors there is not enough capacity to meet the demand the company sees.
- Electricity. Power decides where capacity can be built, how fast, and at what cost. Equinix management cited power availability, legislation, and community sentiment as industry headwinds.
- Data center shells. Buildings engineered for high-density AI are scarce; Equinix says its new facilities carry roughly twice the power and cooling capacity of older sites.
Scale has always mattered in cloud. What has changed is the kind of scale: it is no longer only engineering talent and software distribution, but the ability to secure chips, power, and financing years ahead of revenue.
2. The major players in the next cloud wars
Amazon Web Services remains the largest cloud and posted its fastest growth in nearly five years: revenue rose 39% to $45.98 billion and operating profit was $17.8 billion, a margin of roughly 39% on our arithmetic. Its backlog reached $578 billion, up from $496 billion at the end of June, and revenue commitments for its Graviton and Trainium chips have each grown more than 20x year over year. The cost shows in cash flow: Amazon’s trailing-twelve-month free cash flow swung from a $14.8 billion inflow a year earlier to a $42.4 billion outflow, which its own release attributes mainly to AI infrastructure.
Microsoft Azure leads on enterprise distribution. Microsoft Cloud revenue rose 30% to $63.9 billion, Azure grew 48%, and the company disclosed Azure’s dollar revenue for the first time: about $33.2 billion. Commercial remaining performance obligation rose 73% to $677 billion, a figure that covers all of Microsoft’s commercial business, not Azure alone. Microsoft was the only one of the big four to stay clearly free-cash-flow positive, at $18.6 billion, though that was down 27% from a year earlier.
Google Cloud posted the quarter’s headline: revenue up 103% to $30.7 billion, a $560.8 billion backlog, and a 36.2% operating margin versus 23.7% a year earlier. Two caveats apply. Its figures include Wiz, acquired for $29.5 billion in March, and revenue from selling TPU systems, neither of which is reported separately, so part of the growth is acquired or hardware revenue rather than cloud rental. Even so, its integration of chips, models, and cloud is the clearest full-stack strategy in the market.
Oracle Cloud Infrastructure is the most leveraged bet among the incumbents. In its fiscal first quarter, total revenue rose 30% to $19.3 billion, OCI grew 121% to $7.4 billion, and remaining performance obligations reached $664 billion. Total debt stands at $125 billion and quarterly interest expense rose 55% to $1.4 billion.
CoreWeave is the specialist test case: Q3 revenue rose 168% to $3.66 billion, and backlog reached $132.5 billion, built with a fraction of the hyperscalers’ resources (see Section 5).
The comparison at a glance
| AWS | Azure | Google Cloud | Oracle OCI | CoreWeave | |
|---|---|---|---|---|---|
| Latest quarter revenue | $46.0B | ~$33.2B | $30.7B* | $7.4B (OCI) | $3.66B |
| Year-over-year growth | 39% | 48% | 103%* | 121% | 168% |
| Backlog | $578B | $677B (all Microsoft commercial) | $560.8B | $664B (RPO) | $132.5B |
| Backlog ÷ annualized revenue (our calculation) | ~3.1x | n/a | ~4.6x | ~9x (vs. trailing revenue) | ~9x |
*Includes Wiz and TPU system sales. Backlog definitions differ by company; cross-company comparisons are directional. Ratios are The CODEW’s own calculations from reported figures.
The last row tells its own story: the newer and more concentrated the AI cloud business, the longer the delivery runway behind it, and the more the revenue depends on infrastructure that has not yet been built.
3. The economics of AI infrastructure
Capital expenditure at unprecedented scale
After the Q3 reports, 2026 capex guidance across the big four sums to roughly $720–740 billion: Amazon about $220 billion, Alphabet $195–205 billion, Microsoft about $170 billion including finance leases, and Meta $135–145 billion. These are management projections, not results, and Alphabet has said it expects capex to increase significantly in 2027. One tracker calculates Q3 capex as a share of revenue at 52.5% for Meta, 42.7% for Microsoft, 40.3% for Alphabet and 38.3% for Amazon. These are third-party calculations, but they illustrate the shift from mid-teens percentages of revenue to well over a third.
Free cash flow is the pressure valve
Amazon, Alphabet and Meta posted negative free cash flow in Q3; Microsoft did not. Oracle shows the same pattern: Q3-equivalent capex of $28.5 billion exceeded record operating cash flow of $23.1 billion, with the gap funded by debt and equity, including a $20 billion at-the-market stock sale.
Margins: strong today, with a depreciation lag
Reported cloud margins remain healthy, with AWS near 39% and Google Cloud at 36%. The risk is timing. Capex lands on the balance sheet first and reaches the income statement over years as depreciation, so today’s margins reflect yesterday’s smaller asset base. Alphabet has already said depreciation will rise with infrastructure investment. Margins on the 2026 vintage of capacity are not yet visible.
Utilization, return on capital, training versus inference
Return on invested capital depends on how fully and how long hardware is used. Early evidence is encouraging: Oracle reported 97.9% AI infrastructure utilization and a 20% average premium on GPUs it renewed or resold, and CoreWeave pointed to firm pricing on older GPUs. If older accelerators keep earning, depreciation schedules look conservative; if a new chip generation sharply devalues them, they do not. Training is lumpy and concentrated among a few labs; inference is recurring and grows with usage, so the financial case for the buildout increasingly rests on inference demand, whose pricing is still being set.
Key question: can cloud providers keep attractive margins? Our assessment: probably yes for the largest providers, but with lower returns on capital than the software-era cloud earned. Contracted backlogs, supply scarcity and funding from profitable core businesses support margins; rising depreciation, financing costs and the eventual arrival of excess capacity threaten them. AI infrastructure spending is not automatically profitable. It pays off only if utilization, pricing and customer credit hold for the full life of the asset.
4. The chip and computing power advantage
Nvidia remains the center of gravity. Beyond GPUs, its networking business has become a franchise, with data center networking revenue of $14.8 billion in the quarter ended April, up 199% year over year. Analyst estimates put its share of data center AI accelerator revenue around 70–75%. That is an estimate, not a reported figure, and revenue share flatters Nvidia because custom chips are largely internal costs for the companies that build them.
Custom silicon is the hyperscalers’ hedge. AWS (Trainium, Graviton), Google (TPU), Microsoft (Maia) and Meta (MTIA) all design their own chips for lower cost per unit of compute and leverage in negotiations with Nvidia. Google has gone furthest by selling TPU systems directly, with most of that revenue expected in 2027. Analysts estimate custom silicon at roughly 15–20% of the market and growing fast.
AMD is a credible second source. It reported record data center revenue of $8.5 billion in Q3, up 97%, and holds a 6-gigawatt multi-generation GPU agreement with OpenAI beginning with MI450 GPUs in the second half of 2026. Nvidia’s data center business is still about ten times larger, but AMD gives buyers a negotiating alternative.
The cluster, not the chip, is the unit of competition. Networking, high-bandwidth memory and advanced packaging are the hidden bottlenecks. Microsoft attributed roughly $25 billion of its 2026 capex to higher component pricing, a reminder that supply constraints raise costs for buyers.
5. The rise of specialized AI cloud providers
CoreWeave rents dense, optimized GPU clusters under multi-year contracts to AI labs and enterprises. Its customer mix is improving: through 2027, more than 75% of committed capex is tied to investment-grade and enterprise customers. Q3 revenue rose 168% to $3.7 billion, full-year revenue guidance was raised to $13.0–13.3 billion, and backlog reached $132.5 billion, up 138% year over year. The company expects 36% of that backlog to be recognized within 24 months and 26% after 48 months.
The financing is where the model is most fragile. CoreWeave spent $14.5 billion on capex in Q3 and guides to $37–39 billion for the year, roughly three times its guided revenue. Total debt was about $46.9 billion at quarter-end, more than double year-end 2025, and Q3 interest expense of $846 million exceeded its $207 million operating profit. GAAP net loss widened to $709 million.
Management is moving toward cheaper, contract-linked funding: the weighted-average cost of debt is 7.9%, down 250 basis points year over year, and the CFO said the company does not intend to issue more parent-level unsecured debt at current yields through 2027. Instead, it uses delayed-draw loans matched to specific customer contracts that can finance 70–100% of the associated capex. Management also noted that if it limits investment to committed backlog, 2027 capex could stay roughly flat while revenue and EBITDA grow significantly.
Sustainability. Specialists can build durable businesses where they secure power and anchor customers early, but their margin for error is far thinner than the hyperscalers’, who can fund the same GPUs from operating cash flow. Oracle sits between the two worlds: a legacy vendor running a debt-funded buildout, though most of its newest backlog arrived through prepayments or bring-your-own-hardware deals that need no incremental Oracle capital. That is evidence customers, not just lenders, are increasingly financing the buildout.
6. Enterprise cloud spending: the next battleground
Hyperscaler backlogs are the best available proxy for enterprise and AI-lab commitments, but they include a significant share of AI-lab contracts, not only traditional enterprise spending. For enterprise decision-makers, five dynamics matter:
- Workload migration. Microsoft reports Foundry has more than 120,000 customers, with revenue more than tripling over the past year, a sign that enterprises are building AI applications on hyperscaler platforms rather than only experimenting.
- Multi-cloud as a default. Oracle has expanded its multi-cloud database footprint to 70 regions and made Oracle Interconnect for AWS generally available. Even competitors are building bridges because customers demand them.
- Lock-in and switching costs. AI raises switching costs through data gravity, model-specific tooling, and chip-specific optimization. Trainium and TPU workloads are not portable without rework, so price advantages must be weighed against lost flexibility.
- Private AI versus public cloud. Capacity scarcity pushes some buyers toward colocation and owned infrastructure. Hybrid designs, with steady-state inference on private infrastructure and bursty training in public cloud, are increasingly practical.
- Cost discipline. Buyers must actively manage AI consumption. Cost optimization is becoming a core enterprise capability, not an afterthought.
7. The data center and energy equation
The physical layer is where the cloud wars become industrial competition. Equinix offers a useful read on the colocation market: record annualized gross bookings of $533 million in Q3, up 35% year over year, with pricing on new deals up 12%. It expects to deliver about 40% more cabinets in 2026 than planned at the start of the year, with over 20% of the capacity arriving in Q4 is already pre-sold, and it controls roughly 3 GW of developable capacity. Its CEO described this as the early stage of durable pricing strength. Digital Realty is pursuing a similar strategy of large capacity blocks for hyperscale customers.
Power is the binding constraint. Equinix signed a 20-year agreement in Georgia to pay for new power lines and a substation, an example of operators financing utility infrastructure to secure supply, and acknowledged that power limits, regulation, and local opposition have delayed some projects, including one in Dubai. CoreWeave still expects 1.85 gigawatts of active power by the end of 2026. Even the best-funded buyers depend on suppliers’ schedules: Microsoft said some third-party capacity deliveries expected in its first fiscal quarter were delayed.
Nuclear, renewables, cooling and fiber. Long-term power purchase agreements, including nuclear and renewables, are becoming standard strategic instruments; liquid cooling has moved from niche to requirement as rack densities climb; and fiber and interconnection matter because AI clusters need low-latency links to data and to each other. Equinix and Digital Realty sit where scarce, powered, connected buildings meet demand from both hyperscalers and enterprises. They carry less GPU depreciation risk than cloud providers but share the same grid constraints.
8. The next competitive landscape
Over a five-year horizon, we assess each source of advantage:
| Source of advantage | Assessment |
|---|---|
| Hyperscaler scale and capital | The largest advantage. Roughly $730 billion of annual spending is mostly funded from operating profits, which favors the incumbents. |
| Proprietary chips and integrated platforms | Google has the clearest lead (TPUs, Gemini and cloud together); AWS’s Trainium is scaling; Microsoft’s edge is software distribution. The factor most likely to separate winners within the big three. |
| Pricing power | Strong while demand exceeds supply. When capacity catches up, pricing normalizes, and heavily indebted providers feel it first. |
| Specialized providers | Speed and focus, but thin margin for error. Likely outcomes: consolidation, deeper hyperscaler partnerships, or survival as niche suppliers to AI labs. |
| Enterprise distribution | Microsoft is strongest; AWS and Oracle are strong in specific workloads; Google has momentum but less entrenchment. |
| Energy access and physical ownership | The most underpriced advantage. Providers with secured power can grow while others queue for grid connections. |
If we had to rank who gains most strategically over five years: Google on integration, Microsoft on distribution and cash generation, and AWS on scale and silicon, with Oracle and CoreWeave as higher-variance outcomes. This is analysis, not a forecast of any company’s results.
9. Strategic outlook
Will AI strengthen the hyperscalers or create openings for new competitors? Both, but it strengthens the incumbents more. The capital requirement is a moat. New competitors win where they secure power, chips, and anchor customers quickly, but usually as suppliers to the hyperscalers’ ecosystem or to AI labs rather than as replacements for them.
Can specialized AI clouds build sustainable businesses? Yes, conditionally. CoreWeave’s results show real demand, improving margins and a less risky financing model, but its interest costs exceeded its operating profit in the quarter, and its capex is a multiple of revenue. Sustainability depends on contract quality, delivery schedules, financing costs, and whether GPU values hold. It is a business with a path, not yet one with a proven moat.
Will compute, electricity and data center capacity define the next era? Largely, yes. Software still matters, and Microsoft’s results show distribution is a real advantage, but the binding constraints are now physical: chips, power and buildings. The companies that lose will not necessarily have worse software; they will have been late to power, short of capital, or locked into the wrong hardware.
Bottom line. AI is changing cloud from a software-driven scale business into an infrastructure competition. Backlogs across the major providers now total well over $2 trillion, but a backlog is a promise, not revenue. The industry’s return on this spending will become clear only as the capacity built in 2026 and 2027 meets real utilization and real pricing.
Scope notes and limits of this edition
- Company figures come from reported results, filings, and earnings-call commentary, as relayed by the sources listed below, as of November 14, 2026. Capex totals for 2026 are management guidance, not results.
- Chip market-share figures and backlog ratios are analyst estimates or The CODEW’s own calculations and are labeled as such. Backlog and remaining performance obligation definitions differ across companies; Microsoft’s figure covers its entire commercial business.
- Google Cloud’s reported growth includes acquired Wiz revenue and TPU system sales.
- Digital Realty and several power, cooling and fiber topics are described at a strategic level; fresh quarterly data for them was not gathered for this edition.
Related Intelligence
AI Infrastructure Intelligence · Cloud & Infrastructure Intelligence · Enterprise AI Intelligence · Technology Intelligence · Industry Intelligence
Related Coverage
Sources
- Capex and cash flow: Mojo AI Services, Q3 2026 filings analysis; Value Add VC, 2026 capex guidance; Walnut, AI capex tracker; Alpha Agentic, hyperscaler capex tracker
- Cloud results: Cleus, Amazon Q3 2026; Redmond Magazine, Microsoft FY27 Q1; FourWeekMBA, Microsoft free cash flow; SEC-API, Alphabet Q3 2026 filing analysis; Midgard Finance, cloud comparison
- Oracle and CoreWeave: StorageNewsletter, Oracle fiscal Q1 2027; ERP Today, Oracle backlog; CNBC, CoreWeave Q3 2026; CoreWeave Q3 2026 call transcript; Vested, CoreWeave Q3; TradingKey, CoreWeave Q3
- Chips: NVIDIA Q2 FY2027 results (SEC); NVIDIA Q1 FY2027 results (SEC); AI Business Weekly, AMD Q3 2026; Alatirok, AI chip share estimates
- Data centers: Commercial Observer, Equinix Q3 2026; Equinix Q3 2026 call transcript; TradingKey, Equinix Q3 presentation; Panabee, Equinix Q3 2026
Primary documents to consult directly: company earnings releases, Form 10-Q and 10-K filings, and earnings-call transcripts for Amazon, Microsoft, Alphabet, Oracle, CoreWeave, Nvidia, AMD and Equinix.
Reviewed by Erwin Castro
on
Sunday, November 15, 2026
Rating:
No comments: