North American Cloud CapEx Revised Up to $851 Billion for 2026
Taylor Wilson
Bank of America raised its 2026 global hyperscaler CapEx forecast to $851 billion; JPMorgan on the same day projected North America's Big Four cloud spenders growing ~80% — both point to the same conclusion: 2027 is not a cliff but a high-base extension.
How big is $851 billion, really?
BofA on July 9 lifted 2026 global hyperscaler CapEx to $851 billion, with 2027 rising further to $1.15 trillion — still ~35% year-on-year growth.
JPMorgan issued a separate estimate the same day: North America's Big Four cloud firms will grow data-center CapEx by ~80% in 2026, with at least another 50% in 2027.
This means → two banks using different scopes reached the same verdict — the 2027 spending curve does not roll over; the slope eases slightly but the absolute number keeps climbing.
Q2 alone is expected to hit $197 billion, up 83% year-on-year. In plain terms = the money is already being spent and equipment already being delivered — this is not a far-off projection.
A $2 trillion order book — can it be read as revenue?
Microsoft, Oracle, Google and Amazon have disclosed or estimated remaining performance obligations and backlog totaling over $2 trillion: Microsoft ~$627 billion, Oracle ~$638 billion, Google >$460 billion, Amazon ~$364 billion.
Contract definitions and recognition cycles differ across the four, so this cannot be treated as future revenue directly — but it is enough to underpin multiyear procurement commitments for power, campuses, servers and networking gear.
This means → the order book proves construction certainty, not profit certainty — the money will be spent, but whether it converts into high-margin revenue is a separate question.
How are the four companies spending differently?
Microsoft leans on Azure contracts and leases to accelerate delivery; Amazon uses its in-house chip Trainium — a custom AI inference chip — to drive down per-unit inference cost.
Google is trying to turn its Tensor Processing Unit (TPU) — an AI chip originally built for internal use — from an internal asset into a product it can sell externally; Meta is exploring leasing surplus compute capacity and layering model services on top.
This reflects a clear divergence: Microsoft bets on contract scale, Amazon on cost efficiency, Google on chip commercialization, Meta on compute monetization — similar CapEx totals, entirely different business logic.
Can cash flow handle the pressure?
From 2026 to 2028, cloud CapEx may reach 100%–115% of operating cash flow, compressing the industry's free-cash-flow margin by roughly 1, 5 and 4 percentage points respectively versus the prior baseline.
Since 2026, the leading companies have raised ~$244 billion through long-term debt, equity and structured finance — near-term construction funding is not the bottleneck.
In plain terms = cash is not short today, but spending is running almost as fast as earning — if revenue growth fails to keep pace with rising depreciation, free cash flow gets squeezed thin.
Has the valuation logic shifted — from "how much you spend" to "how much you earn"?
The market's valuation watershed for cloud companies has moved from "CapEx size" to "how much revenue each dollar of CapEx generates."
Three variables will be decisive: depreciation growth rate, compute utilization and revenue per watt — all three need to improve in tandem.
This means → simply spending the most no longer wins. Put simply = the market used to reward "willingness to spend"; now it is grading "whether the spending pays off."
Custom chips rising — will the supply-chain winner set widen?
JPMorgan projects 16.3 million AI accelerator shipments in 2026, rising to 23.3 million in 2027.
Custom ASICs/XPUs — AI chips designed to a specific customer's specification — are expected to climb from 32% unit share in 2025 to 42% in 2026 and 53% in 2027.
This means → Nvidia still commands high-performance training and the general-purpose software stack, but incremental spending will flow more broadly into custom chips, high-bandwidth memory (HBM), advanced packaging, high-speed interconnects, test equipment and fab tools.
This reflects the cycle's most critical falsification signal: as long as order conversion, cloud revenue and utilization outrun depreciation, CapEx revisions can keep supporting the supply chain; but if excess capacity and price erosion appear simultaneously, long-term return assumptions face downward pressure.
Content is for reference only, not financial advice.