A flow-of-funds map of the AI build-out, drawn to one dollar scale: who pays, who spends, where the money lands and who keeps it, for 2025, 2026 and a 2027 scenario, with what the market pays for each node. Click any bar for the detail and its sources.
Flows compiled 2 September 2026 (2025 and 2027 columns 8 September) · market values at the 4 September 2026 close (Asia 8 September) · published 16 October 2026
Anthropic and OpenAI had about $105B of combined annualised revenue in August 2026, against $733B of 2026 capex guided by Microsoft, Amazon, Alphabet and Meta: roughly 7 to 1, from 14 to 1 in 2025. In our 2027 scenario the gap narrows only to about 6.5 to 1.1,18,37,41,44,46
The four put about 99% of their annualised operating cash flow back into capex in 2026, against about 71% in 2025. Their bond issuance reached $219B by 20 August 2026, from $108B in all of 2025. In our 2027 scenario their combined free cash flow is about −$190B.37,39,42,45,57,58,138
Nvidia kept about 66 cents of each revenue dollar as operating profit in its latest quarter, TSMC about 60 and the memory makers 70–80 on a price spike. The two labs are the only node on the map with a net cash outflow.61,62,77,86,89,92
At the 4 September 2026 close, memory makers traded at 4.6× next year's consensus earnings, TSMC at 18× and Nvidia at 15× its fiscal 2028. The two labs' private marks stood at 17× their run-rate and the equipment makers at 29–30× earnings. Multiples are our calculations from public consensus.234,231,208,197,199,243,246
Use the year toggle or the arrow keys to compare years; the bars and ribbons keep the same scale, so widths compare directly. The italic line under a bar is what the market pays for it. Every figure has its basis and sources in the panel.
The map includes Anthropic, the developer of the AI model used to help compile it. See the note at the foot of the page.
2026 basis: lab run-rates (Anthropic end-Jul, OpenAI mid-Aug), hyperscaler CY2026 capex guides at midpoint, latest-quarter operating cash flow annualised, Nvidia Q2 FY27 annualised, TSMC near its +40% guide, memory Q2 annualised. Compiled 2026-09-02.
Esc closes the panel · Tab moves between bars · Enter opens · ← → change year
≈ marks our estimate. Capex composition per Alphabet (60/40) and Microsoft (two-thirds short-lived assets); custom-ASIC share 10–20% of accelerator dollars; Nvidia's payees from TSMC's 20-F and Morgan Stanley/Yonhap; memory shown AI-related only (phones and PCs excluded). Retained = revenue less cash costs and capex for TSMC and memory, less costs for Nvidia. Market values: US names at the 2026-09-04 close, Taiwan, Korea and Japan at the 2026-09-08 close, Amsterdam at 2026-09-07; the labs at their latest primary-round post-money. P/E and EV/sales use consensus for the fiscal year that covers calendar 2027, from public aggregators (MarketScreener, Yahoo Finance, stockanalysis), with the basis stated per company. The multiple on the map divides market value by the bar's own figure for the year shown: the cash it keeps, or the labs' revenue run-rate. Compiled 2026-09-09.
| From | To | $B / yr | Kind | Basis and note | Sources |
|---|---|---|---|---|---|
| End customers of the labs | Anthropic + OpenAI | 105 | Cash paid for goods and services | Subscriptions, API and enterprise revenue: $105B combined run-rate (Anthropic end-Jul, OpenAI mid-Aug 2026). Bases differ: Anthropic books marketplace revenue gross. | [1][18] |
| Lab equity raised | Anthropic + OpenAI | 217 | Capital-markets funding: equity, bonds | Primary lab equity in 2026: OpenAI $122B at $852B post (Mar); Anthropic $30B Series G (Feb) + $65B Series H at $965B (May). | [27][8][9] |
| Core franchises | Big-four hyperscalers | 580 | Cash paid for goods and services | Big-four operating cash flow ≈ $690B/yr (latest quarter annualised) less ≈ $110B of estimated lab compute spend already inside it. | [37][39][42][45] |
| Bonds issued | Big-four hyperscalers | 219 | Capital-markets funding: equity, bonds | Investment-grade bonds 2026 to date, five names: JPM AM $219B (2026-08-20). FactSet: incremental debt = 32% of capex over the last twelve months. | [58][57] |
| Anthropic + OpenAI | Big-four hyperscalers | 110 | Cash paid for goods and services | Estimated annual lab compute spend to the clouds (revenue plus burn). Commitments are multi-year: OpenAI about $1.4T over 8 years; Anthropic more than $300B disclosed. | [26][25][12][13] |
| Anthropic + OpenAI | Lab cash reserves | 212 | Internal costs, reserves, buybacks | Equity not yet spent: cash reserves and burn cover (OpenAI cash and securities $73B at Q1-26; plan to burn $25B in 2026). | [23][22] |
| Big-four hyperscalers | Buybacks, dividends, cash | 176 | Internal costs, reserves, buybacks | Cash in less capex: buybacks, dividends and cash build (derived). | [37][39][42][45] |
| Big-four hyperscalers | DC shells, power, grid | 293 | Cash paid for goods and services | ≈ 40% of capex: data centres, power and networking (Alphabet 40%; Microsoft about one-third long-lived assets; FactSet compute ≈ 60%). | [44][37][57] |
| Big-four hyperscalers | CPUs, storage, assembly | 95 | Cash paid for goods and services | Residual of the compute bucket: CPUs, storage, networking silicon and assembly (estimate). | [57] |
| Big-four hyperscalers | Server memory | 100 | Cash paid for goods and services | Memory content of AI servers, HBM plus server DRAM (estimate; Amazon raised its 2026 capex by $20B on memory cost alone). | [41][94] |
| Big-four hyperscalers | Custom ASIC + AMD | 50 | Cash paid for goods and services | Custom ASIC + AMD ≈ 10–20% of accelerator dollars (Morgan Stanley); Broadcom AI ≈ $58B FY26 across all customers (raised on 2026-09-02). | [68][69][70] |
| Big-four hyperscalers | Nvidia | 195 | Cash paid for goods and services | Nvidia Hyperscale segment $48.7B in Q2 FY27, annualised (10-Q, 2026-08-26). | [62] |
| AI clouds, enterprise, sovereign | Nvidia | 161 | Cash paid for goods and services | Nvidia ACIE segment (AI clouds, industrial, enterprise, sovereign) $40.3B in Q2 FY27, annualised. Includes Oracle and CoreWeave. | [62][63] |
| Server memory | Memory makers | 100 | Cash paid for goods and services | Server memory bought inside servers flows to the three memory makers. | [94][95] |
| Nvidia | Nvidia retains | 224 | Retained by the node | Revenue less cost of goods and expenses: operating margin 66% (Q2 FY27). Buybacks + dividends ≈ $26B per quarter; equity purchases $42B in H1 FY27. | [61][62] |
| Nvidia | Nvidia R&D, SG&A, capex | 43 | Internal costs, reserves, buybacks | R&D $7.1B and SG&A $1.4B per quarter, own capex ≈ $2.7B per quarter. | [62] |
| Nvidia | ODMs | 23 | Cash paid for goods and services | Assembly: Foxconn ≈ 50%, Quanta ≈ 31%, Wistron ≈ 19% of NVL72 racks (Morgan Stanley estimate). | [76] |
| Nvidia | TSMC | 33 | Cash paid for goods and services | Wafers + CoWoS: Nvidia was 19% of TSMC 2025 revenue (≈ $23B); 2026 estimate ≈ $33B. | [75][82] |
| Nvidia | Memory makers | 33 | Cash paid for goods and services | HBM: SK hynix ≈ two-thirds of HBM4 for Rubin; Nvidia supply commitments $279B, primarily memory (annual spend estimated). | [74][62] |
| Custom ASIC + AMD | TSMC | 15 | Cash paid for goods and services | Broadcom, AMD, TPU and Trainium wafers at TSMC (estimate; Broadcom ≈ 15% of 2026 CoWoS). | [73] |
| Custom ASIC + AMD | ASIC vendor margin, other | 35 | Internal costs, reserves, buybacks | Vendor margin (Broadcom adjusted EBITDA ≈ 68%), memory and assembly inside custom systems. | [68] |
| Other TSMC customers | TSMC | 119 | Cash paid for goods and services | TSMC 2026 revenue ≈ $167B (≈ +37% on 2025; the slightly-above-40% guide implies ≈ $171B) less the AI accelerator wafers shown. | [77][79][82] |
| Memory makers | Memory makers retain | 75 | Retained by the node | AI-related memory revenue less cash costs and capex (SK hynix op margin 76%, Micron 80%, Samsung DS ≈ 70%). | [86][89][92] |
| Memory makers | ASML | 8 | Cash paid for goods and services | Memory makers' EUV purchases (estimate; memory = 34% of ASML FY25 system sales). | [85] |
| Memory makers | Applied, Lam, TEL, KLA, fabs | 25 | Cash paid for goods and services | Memory capex to Applied, Lam, TEL, KLA and fab construction (AI-attributable share; SK hynix ≈ $33–35B and Micron ≈ $27B in total). | [87][89] |
| Memory makers | Memory cash costs | 25 | Internal costs, reserves, buybacks | Cash costs of the AI-related memory revenue (estimate). | [86][89] |
| TSMC | TSMC retains | 65 | Retained by the node | Revenue less cash costs and 2026 capex of $60–64B (cash basis; reported operating margin 60%). | [77][79] |
| TSMC | ASML | 12 | Cash paid for goods and services | ASML share of TSMC capex ≈ a fifth (upper bound ≈ 22% from ASML's Taiwan-region sales). | [85] |
| TSMC | Applied, Lam, TEL, KLA, fabs | 50 | Cash paid for goods and services | The rest of TSMC's $60–64B capex: Applied, Lam, TEL, KLA, testers, fab construction. | [79] |
| TSMC | TSMC labour, materials, opex | 40 | Internal costs, reserves, buybacks | Labour, materials, utilities and opex (cash costs excluding depreciation; estimate). | [77] |
| Bar | Market value | Basis | P/E | EV / sales | FCF yield | 12-month | × 2025 | × 2026 | × 2027E | Sources |
|---|---|---|---|---|---|---|---|---|---|---|
| Big-four hyperscalers | $12.21T | MSFT $3.71T + GOOGL $4.14T + AMZN $2.79T + META $1.57T, 2026-09-04 close | 23× | 5.3× | 1.2% | −18% to +46% | 44× kept | 69× kept | 76× kept | [181][182][183][184][185][186][187][188][189][190] |
| Anthropic + OpenAI | $1.82T | OpenAI $852B post-money (2026-03-31; August tender at the same mark) + Anthropic $965B post-money (Series H, 2026-05-28) | — | — | — | — | 63× run-rate | 17× run-rate | 11× run-rate | [197][198][199][200][201][202][203] |
| Nvidia | $5.55T | 24.1B shares × $230.36, 2026-09-04 close | 15× | 8.2× | 2.3% | +34% | 50× retained | 25× retained | 14× retained | [206][207][208][209] |
| Custom ASIC + AMD | $2.68T | Broadcom $1.70T + AMD $0.78T + Marvell $0.20T, 2026-09-04 close | 22× | 9.7× | 1.8% | AVGO +17% · AMD +195% · MRVL +249% | — | — | — | [210][211][212][213][214][215] |
| CPUs, storage, assembly | $610B | Dell $340B + Arista $244B + Super Micro $26B, 2026-09-04 close; the listed server and networking names behind this bar | 21× | 2.0× | 1.1% | DELL +314% · ANET +37% · SMCI −3% | — | — | — | [216][217][218][219][220][221] |
| DC shells, power, grid | $712B | a sample, not a sum: GE Vernova $251B + Schneider $194B + Eaton $160B + Vertiv $108B; contractors, utilities and land are unlisted or not AI-attributable | 30× | 4.5× | 3.7% | VRT +123% · GEV +57% · SU +31% · ETN +18% | — | — | — | [222][223][224][225][226][227][228][229][253] |
| TSMC | $2.03T | 25.93B shares × NT$2,470 (2026-09-08 Taiwan close) at 31.50; the ADR implies $2.22T | 18× | 8.4× | 1.8% | +109% Taiwan · +82% ADR | 47× retained | 31× retained | 29× retained | [230][231][232][253] |
| Memory makers | $3.4T | Samsung $1.28T (common ex-treasury plus preferred) + Micron $1.15T + SK hynix $0.98T; Korea at the 2026-09-08 close | 4.6× | 2.4× | 5.9% | MU +718% · Hynix +547% · Samsung +285% | — | — | — | [233][234][235][236][237][238][239][240][241][253] |
| ASML | $670B | 384M shares × €1,500.40 (2026-09-07 Amsterdam close) at 1.1622; the ADR implies $659B | 29× | 11× | 1.7% | +127% | — | — | — | [242][243][244][253] |
| Applied, Lam, TEL, KLA, fabs | $1.15T | Lam $385B + Applied $361B + KLA $243B + Tokyo Electron $159B; Tokyo at the 2026-09-08 close | 30× | 9.4× | 1.5% | LRCX +206% · AMAT +187% · TEL +162% · KLAC +113% | — | — | — | [245][246][247][248][249][250][251][252][253] |
| ODMs | $152B | Hon Hai NT$3.53T + Quanta NT$1.26T at 31.50, 2026-09-08 Taipei close; aggregator data, medium confidence | 12× | 0.4× | −1.6% | Hon Hai +23% · Quanta +27% | — | — | — | [254][255][253] |
| AI clouds, enterprise, sovereign | $568B | ORCL $457B + NBIS $62B + CRWV $49B, 2026-09-04 close | — | 5.4× | −7.8% | ORCL −29% · CRWV +2% · NBIS +249% | — | — | — | [191][192][193][194][195][196][204][205] |
Three sources of cash enter the system: end customers (≈ $105B/yr to the labs plus the clouds' own AI lines), the hyperscalers' non-AI franchises (≈ $690B/yr of operating cash flow), and capital markets (≈ $219B of hyperscaler bonds so far in 2026, ≈ $217B of lab equity). The first source is an order of magnitude smaller than the build it is supposed to justify; the other two are what actually fund it.
Anthropic and OpenAI convert customer revenue and investor equity into compute contracts with the clouds, and increasingly into their own sites. Anthropic just turned a small adjusted operating profit; OpenAI plans to burn $25B this year. Their revenue is the only line on the map growing faster than the capex above it.
Four companies commit $720–745B of 2026 capex against ≈ $690B of operating cash flow: Microsoft and Amazon with visible backlogs ($678B and $496B), Alphabet and Meta with none to show externally. Oracle and CoreWeave rent capacity they borrow to build.
About 60% is compute, 40% shells, power and networking. Of the compute, Nvidia's hyperscale segment (≈ $195B annualised) is roughly a quarter of the four's capex; custom silicon is 10–20% of accelerator dollars and growing faster. Nvidia keeps 66 cents of every revenue dollar and spends 25 on cost of goods.
Nvidia's dollar goes to TSMC (wafers and CoWoS; Nvidia was 19% of TSMC's 2025 revenue) and to memory (supply commitments jumped from $119B to $279B in one quarter, primarily memory). TSMC keeps 60 cents; the memory oligopoly keeps 70–80 on a price spike that is now the largest input-cost shock in the chain.
TSMC reinvests ≈ 37% of revenue as capex, about a fifth of it with ASML. Memory capex flows to the same toolmakers. The dollar stops here: ASML (54% gross margin), the deposition, etch and inspection vendors, and construction are the last payees.
2025 is actual: company filings for calendar 2025, Microsoft calendarised, Nvidia's fiscal 2026. 2026 uses the latest annualised run-rates and company guides. 2027 is our scenario: company guides where given, consensus and sell-side elsewhere, reconciled at about $1.1T of big-four capex (range $0.93–1.3T).
≈ marks our estimate. Lab revenue is annualised run-rate, not recognised revenue; recognised figures are in the notes. Memory is shown AI-related only.
For TSMC and the memory makers, revenue less cash costs and capex. For Nvidia, revenue less operating costs.
US names at the 4 September 2026 close, Taiwan, Korea and Japan at 8 September, Amsterdam at 7 September; the labs at their latest primary-round valuations. P/E and EV/sales use public consensus for the fiscal year covering calendar 2027. Market value over a bar's own figure (the cash it keeps, or the labs' run-rate) is our calculation.
Each bar's panel lists the sources behind it. Numbering has gaps where a source was consolidated.
These observations are general market commentary, published for information only. They are not investment advice, a recommendation, or an offer or solicitation to buy or sell any security, and they do not take into account any reader's objectives or circumstances. Mentions of companies are not recommendations. Figures are as dated and sourced; our own calculations and estimates are labelled as such. Views may change without notice.
Compiled with AI assistance (Claude, developed by Anthropic) from public sources. Anthropic is one of the companies discussed.