Charted: AI Capex Hits $760B in 2026 Guidance — and $1.4T in the Bull Case

The number everyone quotes is not one number

Ask how large AI infrastructure spending is and you will hear $725 billion, $750 billion, $765 billion, $1.14 trillion, $1.4 trillion, $5.2 trillion, and $7.6 trillion — often in the same briefing deck. Those figures are not disagreements about arithmetic. They are answers to different questions: which companies, which year, which layers of the stack, and whether the total is an annual run-rate or a multi-year cumulative.

The interactive dashboard above is a spend map across those scopes. Toggle 2026 / 2027 / 2028, switch between gross hyperscaler capex and an AI-attributed (~75%) slice, and compare research-house fans against McKinsey’s cumulative scenarios. The point is not to pick a single “true” total. It is to keep the scopes honest so markets can debate substance instead of mixing labels.

What company guidance actually says for 2026

After Q1 2026 earnings, the five largest cloud / AI infrastructure spenders disclosed or reaffirmed calendar-ish guidance that midpoints near:

Sum those midpoints and you land near $760B of gross company capital expenditure for 2026 — not a pure “AI-only” ledger. CreditSights’ post-earnings aggregate (~$750B) sits in the same neighborhood. Apply the commonly used ~75% AI-attributed factor and the AI-specific slice of that stack is roughly $545–570B.

That company stack is the cleanest near-term observed number markets have: it is grounded in guidance, not in a top-down silicon model. It is also incomplete. It excludes non-hyperscaler buyers, sovereign AI programs, much of the colocation / power ecosystem, and the full global compute + facility + generation stack that Goldman Sachs Global Institute models separately.

Annual scenarios diverge hard by 2027

Where guidance ends, research scenarios take over — and the spread widens.

Goldman Sachs Investment Research (hyperscaler gross scope) has circulated a 2027 base near $1.14T against Street consensus near $920B, with a bull path around $1.4T if cloud backlogs and token demand keep supply short into the second half of 2027. That is the origin of many “trillion-dollar AI capex” headlines. It is still a hyperscaler frame: five (or so) balance sheets, gross PP&E, not every watt and every accelerator worldwide.

Goldman Sachs Global Institute’s Tracking Trillions baseline — covered in depth in our chips-and-data-centers breakdown — is a different object. It is an all-in AI infrastructure scenario (compute + data centers + power) of $765B in 2026, $1.01T in 2027, and $1.22T in 2028, compounding toward roughly $7.6T cumulative from 2026–2031. GS is explicit that the Global Institute product is a sensitivity framework, not a house forecast from Investment Research.

Put the two Goldman families side by side and the 2027 comparison looks paradoxical until you read the scopes: IR’s hyperscaler base ($1.14T) can sit above GI’s all-in AI total ($1.01T) because IR is counting entire company capex programs, while GI is counting AI infrastructure layers with a different perimeter and methodology. Mixing them produces fake contradictions.

Street and CreditSights prints for 2026 cluster with company guidance. Into 2027–2028, consensus paths that apply Dell’Oro-style growth rates on top of today’s run-rate typically land below Goldman’s IR base — which is exactly why IR argues consensus is too conservative.

Cumulative frameworks answer a longer question

McKinsey’s Cost of Compute work (April 2025) does not try to pin a single calendar-year hyperscaler total. It frames global data-center capital intensity through 2030 under constrained, base, and accelerated demand:

• Constrained: ~78 GW incremental capacity · ~$3.7T AI-specific · ~$5.2T total DC
• Base: ~125 GW · ~$5.2T AI-specific · ~$6.7T total
• Accelerated: ~205 GW · ~$7.9T AI-specific · ~$9.4T total

These are multi-year cumulative dollars. Annualizing the base AI figure naively (~$5.2T / 6 years ≈ $870B/year) produces a useful order-of-magnitude check against 2026–2027 run-rates — but it is not a substitute for company guidance or for Goldman’s year-by-year GI path. The dashboard’s McKinsey panel is there so readers can see scenario width without pretending the units match an Amazon 10-K line item.

GS Global Institute’s ~$7.6T (2026–2031) cumulative sits in a similar “era-scale” conversation as McKinsey’s base/accelerated band, again with different layer definitions. The right use of both is direction and sensitivity, not false precision to the nearest $10B.

Intensity and financing sit next to the totals

Dollar totals alone do not tell you whether the cycle is sustainable. Capex as a share of revenue for the same companies has moved into ranges last associated with telecom build-outs. Meta’s FY25 intensity near 35% of revenue is the vivid example; Microsoft and Alphabet have also stepped well above their early-2020s norms. On the physical side, our global AI data-center build tracker shows how announced megawatts and live campuses still diverge — spend authorized is not the same as capacity energized.

Funding is shifting with the scale. Public reporting around Goldman’s financing work has pointed to investment-grade hyperscaler issuance rising toward roughly a third of capex by 2026–2027 as free cash flow cannot stretch as fast as the build plan. That does not invalidate the demand signal — signed cloud backlogs and multi-year GPU / power commitments are real — but it does mean credit markets now co-price the cycle alongside equity narratives.

Physical constraints bind the other side. US data-center capacity shortfalls measured in tens of gigawatts by late-decade (Goldman / Morgan Stanley prints in market commentary), CoWoS and HBM bottlenecks on the silicon side, and skilled-labor shortages on the construction side all argue that dollars authorized ≠ megawatts energized on the same schedule. Spend totals can keep rising even while delivered capacity lags the press-release curve.

How to read the dashboard without mixing scopes

Use this checklist when a headline throws a big AI spend number:

1. Company or system? Big-5 guidance / CreditSights / GS IR ≈ hyperscaler gross. GS GI ≈ global AI infra layers. McKinsey ≈ cumulative global DC.
2. Year or era? $760B is a 2026 run-rate family. $5.2T / $7.6T are cumulative decade-edge figures.
3. Gross or AI-attributed? A 75% haircut is a research convention, not a line item in the 10-K.
4. Base or bull? GS IR’s $1.14T vs $1.4T 2027 pair is a scenario fan, not a point forecast with error bars.
5. Sensitivity levers? In the GI framework, $/MW data-center cost and silicon useful life dominate; training-vs-inference mix matters more for returns than for aggregate capital required.

If two numbers fail any of those checks, they are not in conflict — they are simply not comparable.

Caveats and methodology

• 2026 company figures are guidance midpoints and post-earnings aggregates, not final audited totals. Ranges (especially Meta and Alphabet) mean the true year-end print can miss the midpoint.
• 2027–2028 company paths in the stacked area are directional consensus / projected values for visualization continuity; treat them as softer than 2026 guidance.
• AI-attributed (~75%) is a CreditSights-style factor applied uniformly here for interactionactual AI shares differ by company and year.
• GS Global Institute figures are a scenario framework, not Goldman Sachs Global Investment Research forecasts. GS IR hyperscaler figures come from public secondary reporting of research notes.
• McKinsey scenarios are cumulative through ~2030 and include traditional IT as well as AI loads in the “total DC” series.
• Totals may not sum across houses because perimeters differ (leases, power, non-hyperscaler buyers, geographic coverage).
• This post is explanatory data journalism, not investment advice.

Primary synthesis sources: company Q1 2026 earnings guidance; CreditSights aggregate commentary; Goldman Sachs Global Institute Tracking Trillions (April 2026); Goldman Sachs Investment Research hyperscaler scenario figures via public reporting (June 2026); McKinsey & Company, The Cost of Compute (April 2025).