Q3 Concentration: Top-1 Holds 27.5% of Big-5 AI Capex — Top-3 Holds 74%

Absolute AI infrastructure dollar totals answer how large is the stack. They do not answer how concentrated is the system at the top of the distribution. That is the capital-markets question this Q3 concentration lens is built for.

Our theme already maps the vintage path: the July research spend map put Big-5 midpoints near $760B; the Aug post-Q2 update revised to about $802.5B; the mid-Q3 spend update now prints roughly $835B. The late-Aug concentration companion answered the distribution question on a hotter ~$858B late-Aug 202608 stack. This post re-asks the same core question on the mid-Q3 desk vintage that the absolute-dollar Q3 update uses — so shares and levels stay on one tape.

The interactive dashboard above is a concentration lens, not another KPI strip. Top-1 share is 27.5% (Amazon, ~$230B). Top-3 share is 73.6% (Amazon + Alphabet + Microsoft, ~$615B of ~$835B). Approximate Big-5 HHI is ~2,211 — still elevated versus a five-way equal split (~2,000). Pair this distribution cut with the intensity / cash-flow frame for affordability, and with the absolute-dollar updates for the vintage path.

The Q3 concentration scoreboard

Read the table as a ladder. Absolute dollars rose nearly fourfold from 2024 (~$231B) to mid-Q3 2026 (~$835B). Concentration did not equalise. Top-1 share drifted from the low thirties toward the high twenties, then stabilised as Amazon’s Aug→Q3 raise held the lead. Top-3 stayed in a tight band around the mid-to-high seventies — a larger oligopoly, not a broad industrial base.

Top-1 and top-3: the mid-Q3 company ladder

On the mid-Q3 desk vintage the ranked Big-5 ladder is:

1. Amazon — $230B (27.5%)
2. Alphabet — $210B (25.1%)
3. Microsoft — $175B accounting CY (21.0%)
4. Meta — $145B (17.4%)
5. Oracle — $75B net of prepayments (9.0%)

Cumulative share hits 53% after two names and 74% after three. That is the visual the Lorenz panel is built to show: the equal-split diagonal is a straight 20/40/60/80/100 path; the actual path bends hard toward the top-left.

Two definition notes matter immediately. Microsoft’s $175B is the accounting calendar print after lease reclassification — economic campus spend is still commonly modeled nearer ~$190B. Toggle the Microsoft control in the dashboard: under the economic frame, top-1 eases slightly to ~27.1%, top-3 edges to ~74.2%, and the Big-5 sum moves to ~$850B. Oracle’s $75B is net of customer prepayments; gross cash outlays remain nearer $95–100B. Neither caveat erases the concentration story. Both change the second decimal, not the shape of the ladder.

Alphabet’s second-place share is itself a concentration signal. Serial guidance raises through the year — from the low-$180Bs toward a $205–215B band with midpoint $210B on this vintage — mean the number-two program is not a static runner-up. It is a rising second pole. Amazon and Alphabet together hold ~53% of the mid-Q3 stack before Microsoft is even counted.

Growth is more concentrated than the stock

Stock concentration (who holds today’s dollars) and raise concentration (who captured the Aug→Q3 path) are related but not identical. From the Aug post-Q2 print (~$802.5B) to mid-Q3 (~$835B), the Big-5 midpoint sum rose about +$32.5B / +4.0%.

Decompose the positive company raises and the picture skews further:

• Amazon +$10B
• Alphabet +$10B
• Meta +$7.5B
• Oracle +$5B
• Microsoft $0 (accounting print flatlease reclass still the economic story)

Amazon + Alphabet alone account for roughly 62% of positive dollar raises. The donut panel makes that explicit: incremental authorizations remain more oligopolistic than an equal five-way split would imply. For equity and credit desks that treat “the AI capex cycle” as a diversified industry factor, that raise skew is the practical rebuttal. Factor exposure is not five equal names — it is two programs doing most of the upward rewriting on this vintage step, with Microsoft’s flat accounting print masking a still-large economic build.

Compared with the late-Aug concentration companion, where Amazon+Alphabet captured about two-thirds of a larger Jul→Aug raise wave, the Aug→Q3 step is smaller in dollars but still top-heavy. Meta and Oracle absorb a real minority of the positive delta — enough to keep the tail alive, not enough to break the top-two raise oligopoly.

Multi-year path: dollars up, equalisation incomplete

Track top-1 and top-3 across vintages:

Two readings sit in that table. First, as more names authorized mega-programs (Oracle’s step-up; Meta’s band; Alphabet’s serial raises), top-1 share eased from the low thirties even as top-1 dollars exploded — Amazon went from $75B (2024) to $230B (mid-Q3) while its percentage of a much larger pie fell then stabilised. Second, top-3 never left the mid-to-high seventies. The system did not become a broad industrial spend base; it became a larger oligopoly.

HHI on five Big-5 buckets stays above the equal-split benchmark of 2,000. That is a useful cross-post ranking tool, not a reconstructed global IT-capex census. The dual-axis panel — stack size rising while top-1 share stays sticky — is the capital-markets takeaway: scale and concentration can rise together.

From Jul research (~$760B) through mid-Q3 (~$835B), the cumulative raise is about +$75B / +9.9%. Shares moved less than levels. That is the Q3 punchline in one sentence: the pie grew; the slice math barely budged.

Perimeters: when shares do not travel

Company shares only travel inside a declared perimeter. The dashboard’s perimeter bars put mid-Q3 Big-5 (~$835B) next to Street (~$820B), CreditSights (~$830B), GS Global Institute all-in AI infra (~$765B), and GS Investment Research hyperscaler (~$800B). Those houses are answering related but non-identical questions.

• Big-5 / Street / Credit are roughly company-gross hyperscaler worldsshare math is meaningful.
• GS GI is an all-in AI infrastructure construct (compute + data center + power). You cannot allocate that total into Amazon/Alphabet/Microsoft shares without inventing a mapping the institute did not publish.
• GS IR is a hyperscaler research path that can sit above or beside GI depending on year and bull/baseanother reminder that “AI capex” is a family of stacks, not one ticker.

Apply the CreditSights-style ~75% AI-attributed haircut and the AI-specific slice of the mid-Q3 Big-5 stack is about $626B. Toggle it in the ladder view so both perimeters stay visible. Treat the haircut as a convention for comparing AI-heavy programs, not as a disclosed segment line.

Street still trails the desk midpoint stack on this vintage (~$820B vs ~$835B). That gap is a timing artifact as much as a disagreement: consensus catches official raises with a lag. Concentration shares computed on the desk ladder can look slightly different from shares implied by a lagging Street sum — another reason to declare the perimeter before quoting top-1 / top-3.

What would change the concentration story

Several shifts would rewrite this scoreboard:

1. A sustained non-Amazon top-1Alphabet or Microsoft clearing Amazon on a like-for-like gross (or economic) basis for consecutive vintages.
2. Top-3 share falling through the low-sixties while the Big-5 total still growsevidence that Meta and Oracle (or a sixth hyperscaler-scale program) are absorbing incremental dollars faster than the top three.
3. HHI drifting toward ~2,000 and staying there as the stack scalesequalisation, not just temporary raise noise.
4. Raise concentration cooling so Amazon+Alphabet no longer capture ~three-fifths of positive vintage deltas.
5. Meaningful spend outside the Big-5 perimeter that makes “Big-5 share of AI infra” itself a falling metricsovereign, neocloud, and enterprise build-outs large enough to move the industry distribution, not just the footnotes.

None of those are guaranteed by another earnings-week raise. They are competitive and financing outcomes. Until they arrive, desks should price AI infrastructure risk as concentrated program risk first and diversified sector beta second.

Caveats and reading notes

• Perimeter is Big-5 hyperscaler gross midpoints (with Oracle net-of-prepay as disclosed). This is not a full global AI-infrastructure census and excludes large non-hyperscaler and sovereign programs.
• Microsoft lease reclassification moves dollars off the capex line without canceling servers or campusesuse the accounting/economic toggle rather than treating a flat headline as a build pause.
• Oracle net vs gross and Meta finance-lease principal affect levels; they do not invert the top-1 / top-3 ranking on current vintages.
• HHI is computed on five company buckets inside the Big-5 setuseful for ranking concentration across theme posts, not a plant-level market definition.
• AI-attributed ~75% is a CreditSights-style convention applied uniformly; company-level AI mixes differ and are not fully disclosed.
• Vintage midpoints revise after every call and Street catch-updirectionally the concentration ladder has been stable even when point estimates move $5–15B per name on a Q3 step.
• The mid-Q3 desk stack (~$835B) and the late-Aug 202608 concentration companion (~$858B) are related but not identical vintagesdo not splice top-1 shares across those posts without noting the perimeter.
• Figures are USD billions; roundings apply. Do not mix GS GI all-in totals with company share percentages in the same sentence.

Bottom line

The AI capex system remains a concentration system on the mid-Q3 tape. On the Big-5 stack of ~$835B, top-1 holds 27.5%, top-3 holds 74%, and HHI sits near 2,211. The Aug→Q3 raise path is still skewed: Amazon and Alphabet captured about three-fifths of positive dollar raises while Microsoft’s accounting print stayed flat. Absolute dollars answered how large the cycle is. Shares answer who is writing it — and for now, that answer remains a short list.