Global hyperscale cloud and AI infrastructure: AI-cloud demand accelerated sharply in 2Q26, but returns on unprecedented infrastructure spending remain the central debate

Overview

This quarterly comparison examines the AI-driven cloud businesses of Amazon, Microsoft, Google, Alibaba, Oracle and GPU-cloud provider CoreWeave. Bernstein argues that demand and backlog remain robust, but the sector’s investment case increasingly turns on the cost, timing, financing and returns of bringing constrained data-center capacity online.

Core views

Bernstein frames hyperscale cloud as a $1.3–1.5T market opportunity spanning traditional cloud and generative AI. The report says AI has transformed the competitive discussion: Google’s vertically integrated models and TPU silicon have made it a much stronger AI-cloud competitor, Oracle is scaling rapidly from a smaller base, and AWS delivered a marked reacceleration. Across providers, the binding constraint has shifted from GPUs to powered, operational data-center capacity. This supports demand and pricing in the near term, but raises questions about capex returns, capacity-delivery speed, model competition, and whether slower frontier-model progress could curb AI adoption. Capex is the report’s central cross-sector tension. Cash capex at four hyperscalers was roughly $150B in 2Q26, and Bernstein models about $660B for 2026, or close to $800B including Meta. All but Microsoft have sought external funding this year. Revenue and RPO/backlog have accelerated and demand still exceeds supply, but cloud revenue must continue to accelerate to support the pace of investment. Bernstein distinguishes Microsoft by its guidance for positive free cash flow in FY27 and capex growth, excluding component-price effects, below Azure revenue growth. The report also notes that AI IaaS/PaaS mix may depress gross margins during the build-out, while workload mix, GPU useful lives, capacity utilization and cost pass-through will determine longer-run economics. AWS was the quarter’s standout on growth and incremental dollars. AWS revenue grew 37% year on year, its fastest growth in 18 quarters, with an approximately $169B annualized run-rate and about $4.6B of sequential revenue added. Backlog rose to $496B from $364B in 1Q26, helped by a roughly $100B Anthropic agreement, while much 2027 capacity is already committed. AI revenue reached a $25B annualized run-rate from more than $15B the prior quarter; custom silicon also exceeded a $25B run-rate. Bernstein argues that Bedrock, data gravity, custom chips and AI-driven spending on storage, databases, networking and security make AI a multiplier for the wider AWS platform. AWS operating margin reached about 39%, up roughly 170 basis points sequentially, although a $600M energy-contract credit helped; excluding it, margin still increased about 20 basis points. Management’s return framework cites servers with under-three-year paybacks against five-to-six-year contracts, while data centers are built about two years before monetization and can produce revenue for more than 30 years. Microsoft Azure grew 43% in constant currency, above 39–40% guidance, and management guided to 45% for the following quarter with further second-half CY26 acceleration expected. Bernstein attributes the beat to CPU/GPU efficiency, process improvements that brought capacity online earlier, quick monetization of that capacity, and stronger GitHub Copilot consumption. Paid Office 365 Copilot seats rose 50% quarter on quarter to 30M, while commercial RPO reached $678B, up 84% year on year. Capex including finance leases was $41.0B, up 70% year on year, with more than two-thirds on short-lived assets; the company maintained its $190B CY26 guide before lease reclassification, adjusted to $175B after the useful-life change. Bernstein believes concerns over unstarted leases and hardware commitments are overstated, while recognizing that depreciation remains a margin headwind. It estimates Azure gross margin at 50.7% versus 52.6% a year earlier, but sees cloud growth and cost discipline offsetting part of the pressure. Google Cloud revenue rose 82% year on year to $24.8B, operating income tripled to $8.8B, and operating margin reached 35.6%. Backlog increased to $514B, with just over half expected to be recognized over the next 24 months. Bernstein highlights the differentiated combination of Gemini models, proprietary TPU silicon and the start of third-party TPU system sales. Enterprise adoption was broad: nearly 90% of the Fortune 100 used Gemini Enterprise, more than 9M developers built with Google models monthly, and model APIs processed about 22B tokens per minute. However, model delays around Gemini 3.5 Pro and continued capacity scarcity are key concerns. Google plans to use third-party infrastructure as a temporary bridge, which Bernstein expects to pressure cloud margins modestly while supporting customer growth. Alibaba Cloud reported 44.9% year-on-year revenue growth to RMB48.4bn in the June quarter, while adjusted EBITA margin reached 11.6% and EBITA rose 132.7% year on year to RMB5.6bn. AI-related products grew at a triple-digit rate for a twelfth consecutive quarter and represented about 35% of external cloud revenue; AI-related ARR rose to RMB49.5bn from RMB36bn, with MaaS and AI-native subscription ARR at RMB16bn as of August. Bernstein notes incremental cloud EBITDA margins above 50%, supported by scale and tight compute supply, alongside group capex of RMB67.6bn versus a prior implied RMB30bn quarterly run-rate. The report views this as a stronger compute-investment commitment but notes management’s caution against annualizing the quarter. In-house T-head silicon could improve cost efficiency, with the M890 chip commercially deployed and supply expected to ramp in the second half. Oracle’s OCI revenue reached $7.4B in 1FQ27, up 121% year on year, while RPO rose $209B year on year to $664B. Bernstein emphasizes that much recent RPO growth came from prepaid or bring-your-own-hardware AI contracts, which require minimal incremental Oracle funding and may have similar or better margins than ordinary AI contracts. OCI capex was $28.5B, or $18B net of customer prepayments; FY27 guidance calls for $90–95B of capex, including $20–25B of customer prepayments and about $70B of net cash outlay. Oracle completed a $20B equity raise, expects funding needs and capex to peak in FY28, and has secured more than 10GW of data-center capacity to come online over three years. Bernstein argues the AI-cloud transition can be highly value accretive over five to 10 years, though it will pressure near-term margins and cash flow. OCI AI-infrastructure gross margin guidance is 30–40%, while distributed cloud and cloud-native offerings are guided at 40–60%. CoreWeave delivered $2.58B of 2Q26 revenue, up 112% year on year and 24% sequentially, with approximately $104B of backlog and more than $25B of net new commitments signed early in 3Q. It raised FY26 revenue guidance to $12.4–13.2B, adjusted operating-income guidance to $960M–$1.15B, and capex guidance to $35–39B. Active power reached 1.5GW after nearly 500MW was added in the quarter, with a year-end target above 1.85GW; contracted power was 3.7GW at quarter-end and about 4.2GW after quarter-end. Bernstein acknowledges improved pricing and expected contribution margins 5–10 percentage points above recent contracts, but maintains an Underperform view because the company benefits from current capacity scarcity and could be among the first and hardest hit when capacity eases. Its long-term 8GW active-power target by 2030 requires execution across development, interconnection, hardware, financing, deployment and customer demand. Shorter two-to-three-year enterprise contracts may raise ASPs and margins, but increase re-leasing, utilization, residual-value and technology-obsolescence risk.

Analysis framework

Bernstein compares reported and estimated cloud revenue, growth, backlog/RPO, capex, cash flow, margins and capacity across providers on a calendar-quarter basis. It then evaluates each vendor’s AI monetization, infrastructure differentiation, funding needs, contract duration, capacity build-out and margin path, while noting that reported cloud segments are not fully comparable because they include different non-IaaS/PaaS revenues.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI-cloud capacity supply and demand analysis

  The report assesses demand, GPU and powered-data-center constraints, capacity additions and the consequences of potential overbuilding for pricing and utilization.

• Industry AnalysisVolume-price decomposition

  Cloud growth, incremental revenue and pricing analysis

  Bernstein compares year-on-year growth, sequential net revenue additions, backlog and pricing to judge changes in competitive momentum.

• Corporate Fundamentals and FinanceFree cash flow analysis

  Capex, funding and free-cash-flow analysis

  The report evaluates whether capex can be funded internally, the role of prepayments and external financing, and expected free-cash-flow outcomes.

• Valuation methodsEV/EBITDA valuation

  CoreWeave EV/EBIT valuation

  Bernstein states that it values CoreWeave on a 25.5x EV/EBIT basis.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Microsoft (MSFT)

  Covered hyperscaler and AI-cloud provider

  Strengths

  Azure growth, Copilot adoption, large RPO and guidance for FY27 positive FCF.

  Weaknesses

  AI infrastructure depreciation pressures Azure and cloud margins.

  Comparison

  Azure grew 43%, behind Google Cloud’s 82% but above AWS’s 37%.

  Risks

  Capacity constraints, rising capex and margin pressure.

• Amazon (AMZN)

  Covered hyperscaler and AI-cloud provider

  Strengths

  AWS growth reacceleration, $496B backlog, AI and custom-silicon run-rates above $25B.

  Weaknesses

  Large ongoing infrastructure investment precedes monetization.

  Comparison

  AWS added about $4.6B sequential revenue, close to Google Cloud’s roughly $4.7B.

  Risks

  Returns on AI capex and sustainability of AI demand.

• Alphabet (GOOGL)

  Covered hyperscaler and AI-cloud provider

  Strengths

  Vertical integration of models and TPUs, 82% cloud growth, strong backlog and margin expansion.

  Weaknesses

  Reliance on third-party bridge capacity and model-execution scrutiny.

  Comparison

  Cloud growth exceeded AWS and Azure, with sequential net revenue additions near AWS.

  Risks

  Gemini 3.5 Pro delays, capacity allocation and cloud-margin pressure.

• Alibaba (BABA, 9988.HK)

  Covered China hyperscaler

  Strengths

  Accelerating cloud growth, improving EBITA margin, AI ARR expansion and in-house silicon development.

  Weaknesses

  Sharp capex increase and reported revenue includes internal transfers.

  Comparison

  Cloud growth was about 45%, above Azure and AWS but below Google and OCI.

  Risks

  Compute investment scale, supply ramp and pace of in-house-chip efficiency gains.

• Oracle (ORCL)

  Covered hyperscaler and AI-cloud provider

  Strengths

  Rapid OCI growth, $664B RPO, prepaid-contract economics and expanding data-center capacity.

  Weaknesses

  Large investment and financing requirements weigh on near-term cash flow and margins.

  Comparison

  OCI’s 121% growth was fastest among the providers discussed.

  Risks

  Funding execution, capacity delivery, OpenAI-related demand and capex peak in FY28.

• CoreWeave (CRWV)

  Covered GPU-cloud provider

  Strengths

  Strong revenue growth, backlog, customer commitments and active-power expansion.

  Weaknesses

  High capital intensity, customer concentration and reliance on constrained-capacity conditions.

  Comparison

  A specialized neocloud rather than a scaled incumbent hyperscaler.

  Risks

  Capacity easing, residual-value risk, technology obsolescence, financing costs and delivery delays.

Key data

• Four-hyperscaler cash capex~$150B in 2Q26; ~$660B in 2026EBernstein estimate for four hyperscalers; close to $800B including Meta.
• Cloud growth in 2Q26Azure 43%; AWS 37%; Google Cloud 82%; Alibaba ~45%; OCI 121%Year-on-year growth comparison.
• AWS backlog$496BUp from $364B in 1Q26.
• Google Cloud backlog$514BMore than half expected to be recognized as revenue over the next 24 months.
• Oracle RPO$664BUp $209B year on year, or 46%.
• CoreWeave FY26 capex guidance$35B–$39BRaised following strong capacity delivery and customer wins.

Impact & implications

The report portrays AI cloud as a large and demand-constrained growth market, but argues that provider outcomes will diverge according to capacity access, workload mix, proprietary infrastructure, contract economics, financing and the ability to protect margins while scaling. Incumbent platforms show improving monetization evidence, whereas specialized GPU-cloud providers face greater sensitivity to normalization in capacity and pricing.

Risks

• AI capacity may be overbuilt, which could reduce pricing for bare-metal or low-software infrastructure.
• Slower frontier-model innovation or AI adoption could weaken demand and the value of current infrastructure investment.
• AI IaaS/PaaS build-out may pressure gross margins, with GPU useful life and workload mix remaining uncertain.
• Powered-data-center delivery, hardware procurement, funding and customer deployment create execution risks, especially for CoreWeave.
• Google faces model-execution delays and temporary margin pressure from third-party capacity; Oracle faces significant financing and cash-flow demands.

What to watch

• Whether cloud revenue and backlog growth continue to support the pace of capex.
• Availability and speed of bringing powered data-center capacity online.
• Evidence of AI infrastructure returns, including payback periods, margins and free-cash-flow conversion.
• Google’s frontier-model execution and Gemini 3.5 Pro progress.
• Oracle’s capex, financing and cash requirements ahead of its October Financial Analyst Day.
• CoreWeave’s conversion of contracted power into active revenue-generating capacity, customer concentration and shorter-contract exposure.