Microsoft / Azure vs AWS / Amazon vs Google Cloud vs Meta: What AI Capex Is Actually Made Of

Amazon leads in absolute spend at $131.82B for fiscal 2025, but Microsoft and Meta each commit 35% of revenue to capital expenditure—nearly twice Amazon's 18%—and within those totals the composition is diverging sharply across custom silicon, long-dated power contracts, and submarine cables.

Amazon spent more than any other hyperscaler in absolute terms—$131.82B in fiscal 2025, up 58.8% from $83B in fiscal 2024—but that figure flatters its conviction: at 18% of revenue, Amazon is the lightest spender by intensity in the peer group. Microsoft and Meta each commit 35 cents of every revenue dollar to capital expenditure, with Microsoft posting $115.95B in its fiscal 2026 (ending June 2026, up 79.6% from $64.55B) and Meta $69.69B in calendar 2025 (up 87.1% from $37.26B—the sharpest acceleration in the group). Google/Alphabet sits at 23% of revenue on $91.45B in fiscal 2025, up 74.1%. Together, a late-September 2026 estimate puts the four companies combined 2026 AI capital outlay at $725B; Goldman Sachs projects that five major buyers will reach $1.2T by 2027, approximately 50% above the 2026 pace.

The largest single line inside any AI capex budget is accelerated compute. Google is farthest along in monetizing its own silicon: Alphabet began recognizing TPU system revenue for the first time last quarter, with Piper Sandler projecting that line could reach $104B by 2028. To hedge interconnect lock-in as cluster sizes scale, Alphabet is simultaneously backing Nvidia's NVLink Fusion and AMD's UALink—an explicit dual-supplier bet at the fabric layer. AWS has pursued the most aggressive third-party silicon diversification: its September 2026 multi-generation agreement with Qualcomm, reported at up to $60B, for custom AI inference accelerators is the first deal of its scale by a Western hyperscaler outside the Nvidia ecosystem; AWS is already deploying Qualcomm hardware for inference workloads while Qualcomm uses AWS Bedrock to co-design subsequent generations. Meta's approach is the structural outlier: its Muse AI agent runs on AMD EPYC Turin CPU hosts—two cores and 8GB of memory per user—on the thesis that recommender-feed inference costs approximately one-hundredth of generative chat per workload, meaning a large fraction of Meta's AI traffic never touches GPU silicon. Microsoft's silicon composition is less detailed in available disclosures, but its $115.95B fiscal 2026 outlay—and one September 2026 analyst projection of $190B in calendar-year AI infrastructure spend—implies GPU procurement at a volume that shadows the Nvidia data-center segment's total revenue.

Physical construction is the longest-lead capex component and therefore the clearest indicator of five-year conviction. Microsoft is targeting a 38-gigawatt global data center network by 2032, has annexed 3,500 acres in Wyoming for a new facility, and pledged more than $10B across UAE, Saudi Arabia, Qatar, and Kuwait by 2030. Meta's Hyperion campus in Louisiana has expanded to 5 gigawatts and a reported $50B in total spend, with full completion targeted for 2032—one of the largest single infrastructure investments on record. AWS secured approval in September 2026 for a 235-megawatt development near Sydney and committed $5.3B for 50 megawatts at Saudi Arabia's AI Zone by 2028, but its Middle East position took an unplanned hit: after drone strikes damaged Bahrain and UAE facilities, AWS instructed clients to permanently abandon those resources with no recovery timeline. Alphabet is reported to be pursuing an off-grid energy and commercial strategy—co-locating generation capacity at campus rather than relying solely on grid connections—but specific new campus announcements are not in current coverage.

Power is increasingly the binding constraint on deployment, not silicon. Microsoft's 20-year power purchase agreement with Chevron for its Texas data center—natural gas rather than purely renewable—signals that reliability is winning over optics. Both Microsoft and Google are co-signed on supporting Australia's flexible clean energy rules rather than hard zero-carbon mandates, an explicit statement that neither intends to let regulatory rigidity govern deployment pace. Meta is running both playbooks simultaneously: a large Texas solar PPA coexists with gas generation for some AI data centers, and the company has reportedly restructured certain facilities as experimental research models to access more favorable federal tax treatment. A New Jersey facility linked to Microsoft was fined $1.07M in September 2026 for 62 unpermitted backup generators—a detail that illustrates how rapidly physical power buildout is outrunning permitting infrastructure at a stage when Microsoft's 38-gigawatt target still lies six years ahead.

Networking has moved from overhead to strategic asset. Meta's Petal submarine cable, spanning roughly 4,300 miles between the US and France, delivers a petabit of raw bandwidth—a new order of magnitude for a hyperscaler-owned cable. AWS is building a 420-terabit-per-second transpacific fiber route from Japan to Washington. The Qualcomm-AWS collaboration extends into optical interconnects, targeting 1.6 terabits per second for within-cluster AI networking—a figure that matters because inter-accelerator bandwidth now governs model scale as directly as compute FLOPS. Alphabet's simultaneous funding of both NVLink Fusion and UALink is partly a networking hedge: both are high-speed accelerator-to-accelerator fabrics, and architecture optionality matters most as cluster sizes push toward hundreds of thousands of accelerators.

The clearest near-term constraint on all four companies is not capital or silicon but TSMC's CoWoS advanced-packaging capacity, which is already limiting Google's TPU deployment pace and caps GPU supply across the group. Three signals will determine whether the current intensity is sustainable: whether TSMC's CoWoS output expands materially through 2027, setting the supply ceiling for accelerators; whether Google Cloud's contracted backlog of $513.9B converts to revenue fast enough to validate Alphabet's raised 2026 guidance of $195B to $205B; and whether the Qualcomm-AWS deal at up to $60B forces Nvidia to reprice at the hyperscaler tier, reshaping everyone's silicon cost structure. Analysts flagged in late September 2026 that the S&P 500 earnings tailwind from hyperscaler AI spending may already be fading—and one analysis argued that if the capex cycle cracks, Nvidia's order book would show the stress before Meta's, a direct consequence of Meta's CPU-heavy inference architecture providing a structural buffer that Amazon, Microsoft, and Google do not have.