←Back to Research
Research #17 · Research Report · Token Economics

The Gigawatt P&L: Bringing the AI Trade Back from Narrative to Unit Economics

Three equations, one transfer price, three industrial forms: this paper calibrates the unit economics of the AI economy against two landmark 2026 filings, builds up from single-company decompositions to an industry-level P&L, identifies the information limits of capital markets, and proposes a new valuation KPI matrix with a tiered disclosure standard.

ByYumei Dou
PublishedNovember 19, 2026
FormatResearch Report · Token Economics
Statusv1.0 · published 2026-11-19 (data through Anthropic's final prospectus and first-day close)

Part I · The Mathematical Spine: Three Equations, One Transfer Price

The AI economy has two layers: an infrastructure layer that turns electricity into compute, and an application layer that turns compute into intelligence and charges for it. Each layer has one revenue equation, and the two layers are joined by a single transfer price. These three relationships form the spine of everything in this paper; every number that follows hangs on one of them.

Equation 1: Infrastructure-layer revenue (expanded in energy terms)
Revenue per GW·yr = utilization × 3.15×10¹⁶ J × (tokens / joule) × ($ / token)
One gigawatt running flat-out for a year is about 3.15×10¹⁶ joules — a physical constant. Tokens/joule is set jointly by model architecture and silicon; it is a physical-layer variable (the subject of Research #15 and #16). $/token is set by market pricing. The numerator of a financial metric already contains the physical-layer metrics: promoting physical-layer indicators into valuation indicators is not a forecast — it is algebra.
Equation 2: Application-layer cost (the Fabless cost equation)
Compute cost per GW-equivalent·yr = (GPUs per GW) × ($ / GPU·hr) × 8760 × reservation rate
The physical bridge: tokens / GPU·hr = tokens/joule × GPU power × 3600. The GPU·hr basis and the GW energy basis convert into each other exactly through power. Tokens/GPU·hr varies with hardware generation and model throughput, and $/token varies with model capability — any citation must carry its spec.
Equation 3: Cross-layer closure (the transfer-price identity)
Infrastructure-layer revenue ≡ application-layer compute cost; application-layer revenue = ρ × application-layer compute cost
The same dollar is revenue to the seller and cost to the buyer: $ / GPU·hr is this industry's transfer price. ρ is revenue per compute dollar — the application layer's core efficiency ratio. Solve the three equations together and the entire industry's P&L can be derived from unit data.

These three equations are the entire grammar of what follows. Each gets a one-line job description, then a specific assignment in each part of the paper:

Table 1: What each equation does, and where

EquationOne-line jobRole in Part IIRole in Part III
Equation 1 (infrastructure revenue)The industry's macro revenue ruler: a bottom-up measure of industry-wide AI revenue and its outlook, built from installed capacityPrices the Foundry and the IDM; the yardstick for single-company decompositions; multiplied by installed base, it gives the industry's revenue lineIts physical-layer variables (tokens/joule, bytes/token) are precisely what the market does not observe today: the physical layer of the KPI matrix comes from here
Equation 2 (Fabless cost)A two-way ruler between layers: the same formula read by the buyer is a compute-cost account, read by the seller is a lessor's revenue model (reservation rate and utilization mirror each other)Prices the Fabless player; its GPU·hr bill, bucketed by contract tenor, is where application-layer risk livesThe tenor buckets are not disclosed — which is exactly why the "premium compute" risk the S-1 itself admits cannot be audited: the disclosure matrix comes from here
Equation 3 (closure / transfer price)The single-company monetization ruler: application revenue = ρ × compute cost; ρ and compute input are both revenue multipliers, measuring model competitiveness and the success of commercial monetizationThe reconciliation identity of the industry P&L: the Foundry's output is the Fabless player's input, and they cancel on consolidation; the IDM folds the identity inside its own booksEach of the three sources of information failure touches it once: source one misuses it (extrapolating an industry base from a few companies' readings), source two makes it unusable (folding leaves no separable variables), source three makes it unreliable (closed-loop financing contaminates the transfer price)

Calibration: Measured Values from Two Filings

The 2026 filing season gave every variable in the three equations an auditable anchor for the first time. Table 2 collects the calibrated values for both layers, with sources:

Table 2: Calibrated values and sources for the three equations' variables

VariableCalibrated valueBasis and sourceHangs on
Revenue per GW·yr≈ $16.5B (prior quarter: $6.3B)SPCX Q3 8-K: SI infrastructure revenue $7.022B annualized / 1.7 GW end-of-periodEquation 1, left side
Build cost per GW$40–50BIndustry all-in basis (Altimeter: ≈ $50B for the Vera Rubin generation; NVIDIA: $40B+ of sellable content per GW)Cost benchmark
Depreciation per GW·yr$8–10BFive-year straight-line, InAI estimateCost benchmark
tokens / joule0.1 to 2.5 (boundary conditions required)TokenPowerBench and John Snow Labs 2025–26 benchmarks: ≈ 2.56 for a 70B-class model at FP8, large batch; long context and deep reasoning fall to 0.1–0.14; energy per token roughly triples from 2K to 10K contextEquation 1; the physical bridge
bytes / token2.3× to 48× reduction, generation over generationResearch #16: measured across thirty-plus models from ten labsThe architectural elasticity behind tokens/joule
ρ (revenue per compute dollar)1.77 (first nine months of 2026; 1.72 in Q3 alone)ANT S-1/A, as filed: nine-month revenue $33.242B ÷ compute and infrastructure expense $18.808B. Media readings of 2.08 used a narrower Q2 pure-compute basis; this paper standardizes on the auditable wide basisEquation 3
Application-layer operating cash flow+$6.6B over nine monthsANT S-1/A (−$2.0B in the same period a year earlier)The cash signature (Part II)

Two of these calibrations need immediate warning labels. First, the denominator warning: SPCX's revenue per GW jumped from $6.3B to $16.5B in a single quarter while capacity grew only 21%. Per the 8-K, the primary driver was new cloud service agreements commencing billing, with the Cursor consolidation (closed mid-August) contributing roughly $1.4B of the unit-level increase. A single-quarter reading with end-of-period capacity in the denominator is inherently volatile — this is a direct argument for the disclosure standard in Part III. Second, the spec warning: the 25× range in tokens/joule is not measurement error; it is the true span of boundary conditions (chip-level versus facility-level with PUE, prefill versus decode, batch and context specs) and architectural elasticity. Any citation must carry its spec.

Closure Test: Three Independent Sources Cross-Verify

Whether Equation 1 actually holds can be tested by inversion. Take the SPCX-calibrated $16.5B/GW·yr, assume roughly 1 token/joule at facility level and 50% utilization, and the implied blended realized price is about $1.05 per million tokens — squarely inside the observed range of blended API pricing. Filings (SPCX 8-K), energy benchmarks (the TokenPowerBench family), and market prices (Artificial Analysis) close on each other: this set of equations is not just algebra; it is verified by three mutually independent data sources. Connecting the application layer through Equation 3: application-layer revenue supported per GW ≈ 16.5 × 1.77 ≈ $29B per GW (an illustrative chain multiplication, InAI estimate).

Figure 1: The Mathematical Spine and the Three-Source Closure Test

Three equations chain the physical layer to the application layer; each link is anchored by an independent source; the inverted price lands inside the market range

Equation 1: infrastructure-layer revenue 1 GW·yr 3.15×10¹⁶ J × tokens/joule 0.1–2.5, with boundaries TokenPowerBench, R15/16 × $/token × utilization market pricing, ops variable Artificial Analysis Revenue per GW·yr ≈ $16.5B SPCX Q3 8-K, measured Equation 3: cross-layer closure (same dollar — seller's revenue, buyer's cost) Infra revenue ≡ application compute cost Transfer price: $/GPU·hr (bridge: tok/GPU·hr = tok/J × power × 3600) App revenue = ρ × compute cost; ρ = 1.77 (ANT S-1/A, measured) App revenue supported per GW ≈ 16.5 × 1.77 ≈ $29B (illustrative) Closure: Equation 1 inverted ≈ $1.05 / M tokens (1 tok/J facility-level, 50% utilization) Inside the blended API price range: filings · energy benchmarks · market prices cross-verify
The closure test is an illustrative InAI estimate: tokens/joule is a facility-level approximation and utilization is assumed; changing the boundary conditions moves the implied price in the same direction. ρ uses the EDGAR S-1/A as-filed basis (nine-month 2026 revenue ÷ compute and infrastructure expense).
Field note · 2026-11-08: The election shock — this framework's first live stress test

Between November 3 and 6, an election outcome the market read as slowing the AI buildout landed — and the cross-section sorted itself exactly along this paper's dividing lines. The collectors of scarcity rents were sold (memory −19.3%, optics and interconnect −16.8%, data-center physical layer −12.0%, semiconductor equipment −11.4%); the rent payer was bought (Apple +3.2%); the hyperscalers were roughly flat (losing growth, gaining capex relief); and the non-AI control group was +0.1% (InAI calculation, equal-weighted groups on unadjusted closes, 2026-11-03 to 11-06).

Three implications for this framework. First, Equation 1's variable attribution played out in real time at the industry level for the first time: the physical terms (tokens/joule, installed base) did not move — what got repriced were the commercial terms and duration. Second, Table 9's growth requirement moves with the pace of construction: on a fleet basis, if 2027 starts are frozen, the required revenue CAGR falls from roughly 30% to roughly 17% — a slowdown lowers the industry's endogenous floor before it hurts the demand side. Third, Table 11's backlog coverage ratio becomes the discriminator of who benefits in that scenario: low-coverage, short-contract players (0.47) capture the most repricing as supply tightens, while high-coverage, locked-price players (12) are protected but capped.

Update · 2026-11-10: the window did not get longer; it is opening on schedule. Anthropic filed Amendment No. 2 to its S-1 the same day and launched the roadshow: a $70–78 range on 825 million shares, implying roughly $920B–1.02T of market value on shares outstanding after the offering and up to ≈ $1.40T on the fully diluted media basis. The amendment added three incremental readings: end-October run-rate revenue of $81B, up roughly $10B in one month (from ≈ $71B at end-September), a +14.1% month, level with the Q2→Q3 sequential pace of ≈ +13.9%/month and the largest monthly addition on record; total commitments A updated to $900.9B (from $895.4B); and a $26B concurrent private placement of Class N stock to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC: suppliers entering the capital structure as shareholders, giving Table 5's Ld a directly readable new component from day one of trading.

Update · 2026-11-19 (listing day): priced above the range, and the disagreement over the multiple is a disagreement over share count. Anthropic priced on November 18 at $83 a share (range $70–78) and upsized the offering from 825 million to 980 million shares (about 956 million new shares from the company and about 24 million from its seven co-founders), raising about $81.3B, the largest IPO on record by base offering size; underwriters hold an option on a further 147 million shares; the stock lists on Nasdaq as ANT on November 19 and the offering is expected to close on November 20. On the roughly 13.22 billion shares outstanding after the offering (final prospectus; includes the concurrent placement shares, excludes the over-allotment option), market value at the offer price is about $1.10T, about 13.5× end-October run-rate revenue; on the fully diluted media basis, which counts options and RSUs, it is about $1.5T, about 18.5×. The stock closed its first day at $97.17, up 17.1% from the offer price, which puts market value on shares outstanding at about $1.28T, about 15.9×; yet what could trade that day was mainly the 980 million shares of the public offering, about 7.4% of shares outstanding, and still under a tenth with the employee shares eligible for sale from day one (up to 315 million): that valuation was set by trading in less than a tenth of the company.

Three implications for this paper. First, basis discipline for the multiple. One company on one day has three share counts: the 980 million shares of the public offering, about 13.22 billion shares outstanding, and about 18 billion fully diluted with options and RSUs (implied by the $1.5T media figure). At the offer price "13.5×" and "18.5×" are both correct; the only difference is which count is used. The $1.5–2T discussed in the media before the roadshow was typically a fully diluted figure, and the deal priced at its low end. The decline from the 22–30× of that discussion (against Q3 annualized revenue) comes mainly from the denominator: run-rate rose from $67.8B to $81B within a quarter. Second, supplier shareholding is now real. At the offer price the $26B placement is about 313 million Class N shares: about $10B each from Google and Nvidia, about $2.5B from AMD, about $1B each from Broadcom, Micron and Microsoft, and about $0.5B from TSMC (Amendment No. 3), about 2.9% of total commitments A; once closed it is the first grade-three (funds disbursed) supplier money on the demand side (Table 5). Third, A has a time profile for the first time. Amendment No. 3 (2026-11-17) discloses that about $451.6B of the $900.9B is expected to be incurred before the end of 2030, with substantially all of the rest falling in 2031–2045; about $70.7B is due within the 12 months after 2026-09-30, about 1.04 times Q3 annualized revenue and about 0.87 times end-October run-rate revenue. The same document defines run-rate revenue as consumption fees over the trailing 28 days multiplied by 13, plus monthly subscription fees multiplied by 12: a four-week annualization, not realized annual revenue. The offering was completed on its original timetable after the election shock: Table 12's disclosure-format window is not extended; it is now. Body figures in this paper have been refreshed to Amendment No. 3.

Part II · From Single Companies to an Industry P&L: Three Industrial Forms

With the spine in place, individual companies can be decomposed. But to go from companies to the industry, one structural question has to be answered first: whose books do the three equations sit on? The answer depends on the company's degree of folding. The semiconductor industry spent thirty years unfolding itself from vertical integration (IDM) into a Fabless/Foundry division of labor, and capital markets learned to price the layers separately: foundries on capacity and utilization, fabless designers on design and channel — nobody applies the same multiple to both TSMC and Apple. The AI economy is in the early innings of the same unfolding, and all three forms already coexist. Table 3 is this paper's core framework table:

Table 3: The three-form framework (revenue equation, cost equation, 2026 evidence, cash signature)

FormRevenue equationCost equation2026 evidenceCash signature
Foundry
builds compute, sells compute
Azure · AWS · Oracle · neoclouds
GPUs/GW × $/GPU·hr × utilization
(= the downstream's compute cost, Equation 3)
Build cost per GW ($40–50B)
Depreciation per GW·yr ($8–10B)
Pure-rental-basis revenue ≈ $6–12B/GW·yr (InAI working-paper estimate)Heavy capex; deeply negative FCF through construction; 2.7–4.2-year payback; duration and refinancing risk
Fabless
rents compute, sells intelligence
OpenAI · Anthropic
tokens × $/token
= ρ × compute cost (ρ = 1.77)
Equation 2: the GPU·hr bill, bucketed by tenor (spot / short / long)ANT: nine-month revenue $33.2B (≈ 13× YoY); operating cash flow +$6.6B; total commitments ≈ $900.9B (S-1/A Amendment No. 3; the 11-02 amendment showed $895.4B, of which fixed/minimum-guarantee $313.8B, tenors 1–20 years)Asset-light but commitment-heavy: take-or-pay-style contracts (Google, Amazon, Microsoft — "regardless of actual usage," per the filings) are debt-like fixed costs; fast cash conversion, extreme operating leverage
IDM
builds and operates (the folded form)
SPCX SI · Google (partly)
Reports revenue per GW directly ($16.5B)
The transfer price vanishes inside the books
Both cost equations stacked on one balance sheetSPCX: Q3 SI revenue $7.56B, capex $16.4B (84% of the whole company); nine-month FCF ≈ −$35BThe two signatures blended — the hardest books to read; when the folding boundary moves, unit revenue jumps

The table yields three immediate corollaries. First, the gap between the Foundry pure-rental basis ($6–12B/GW·yr) and SPCX's reported $16.5B comes from bundled solutions and services: companies with different degrees of folding have unit economics that cannot be compared directly. Second, within this framework ANT's enormous long-dated contracts earn a precise analogical name: capacity reservations (our term; the S-1 says only "commitments"). Their economic substance is isomorphic to Apple prepaying TSMC to lock future wafer supply — which means the disclosure and analysis conventions the semiconductor industry built around capacity reservations can be transplanted wholesale. Third, today's market prices all three forms (Table 3's IDM, Fabless, Foundry) with the same generic toolkit: revenue multiples. SPCX (IDM) trades at roughly 35× annualized sales; ANT priced its IPO on 2026-11-18 at $83 a share, above the filed $70–78 range, implying about $1.10T of market value on the roughly 13.22 billion shares outstanding after the offering (final prospectus), about 13.5× its end-October run-rate revenue of $81B; on the fully diluted media basis, which counts options and RSUs, about $1.5T and about 18.5×. The two bases differ by nearly 40%, so any quoted multiple must state its share count; the $1.5–2T of pre-roadshow media discussion was 22–30× Q3 annualized revenue, the deal priced at its low end, and the decline in the multiple comes mainly from growth in the run-rate denominator rather than from price. It is the same revenue-multiple language; CoreWeave and the independent compute providers (Foundry) are likewise priced on revenue multiples plus a backlog narrative. But the three forms have completely different cash-conversion structures (Table 3, last column): the same $1 of revenue comes with positive operating cash flow in Fabless hands and with tens of billions of quarterly external financing needs in IDM hands. Pressing one revenue multiple onto three different cash structures is the structural root of Part III's "information limits."

The IDM Decomposition: SPCX and the "Six-Month Payback" Verdict

SPCX's SI business is the specimen IDM — and the only filer in the entire market from which revenue per GW can be read directly. The calibration ($16.5B/GW·yr) and the denominator warning were given in Part I; here we complete the cost-side comparison and deal head-on with one claim that demands it.

First, use Equation 1 to attribute the jump — its first empirical application. When the single-quarter jump from 6.3 to 16.5 happened, the physical terms on Equation 1's right side did not move: chips and model architecture were unchanged (tokens/joule constant), and utilization's quarterly movement was limited and amplified by the end-of-period denominator. By elimination, the jump can only sit in the commercial terms: $/token and billing scope (new cloud service agreements commencing billing; the Cursor consolidation enlarging the billable entity) — which cross-verifies precisely with the drivers disclosed in the 8-K. Equation 1 turns "why did unit revenue jump" from a narrative question into variable elimination — and shows again why denominator and billing-scope disclosure (Table 12) are indispensable.

$16.5B of annual revenue against a $45B midpoint build cost: payback is 2.7 years on a revenue basis and about 4.2 years on an EBITDA basis (Q3 SI adjusted EBITDA margin 65.5%) — enough to cover $8–10B of annual depreciation with room to spare. That is a healthy but not extraordinary level of unit economics. The market, however, is circulating a far more aggressive claim: Musk's "new compute pays back in six months," from the SPCX Q3 earnings call. We take the claim seriously and complete the test: six-month payback requires roughly $90B of annual revenue per GW — more than five times the company's own filing-implied value ($16.5B) and above the most optimistic number on any public basis. Verdict: the claim fails on both the revenue basis and the EBITDA basis. The only self-consistent reading is "cash operating recovery excluding equipment cost" — which is not what "payback" means to investors. In Figure 2 we print the verdict directly on the chart rather than leaving readers to guess.

With the verdict in, one dimension is still missing: build speed (τ, the lag from capex outlay to revenue generation). SemiAnalysis's measured timeline: SPCX's first cluster live in 122 days, the second cluster's first phase in 91 days, MiniHard built to 450MW in about 5 months, current cadence roughly 450–500MW activated per month across four parallel projects; the industry norm is 8–12 months. The 2–3-quarter gap changes three line items (InAI estimates). One: full-cycle payback = τ + revenue-basis payback — about 3.0 years for SPCX versus roughly 3.5–3.7 years at industry pace, a 15–20% shortening. Two: construction-period financing cost ≈ capex × τ × cost of funds — about $2.0–3.5B saved per GW (assuming 8–12% cost of funds). Three: generational rent capture: GPU rents fall about 25% a year (SemiAnalysis index: blended H100 $6.62/GPU·hr in H2 2023 down to $2.80 in March 2026), so each extra quarter of τ costs about 6% of first-year revenue and swaps the highest-rent quarters of the depreciation life for the lowest. The defensible core of Musk's claim is a speed advantage — but it shows up as a 15–25% shortening of full-cycle payback, not an 80% one. The verdict stands; the advantage has a different name: τ. τ is also a financing-structure variable: every quarter shaved off the lag is one quarter's less capex that external money must bridge — the structural reason SPCX can sign large compute contracts that begin billing in the month of delivery.

Figure 2: Revenue per GW — Measured, Misused, and a Claim That Fails

Units: $B / GW·yr. Two common basis errors are adjudicated directly on the chart

$B / GW·yr 6.3 SPCX Q2, measured 16.5 SPCX Q3, measured this paper's calibration 40–50 Build cost (one-time) ⚠ Basis error: misread as annual revenue ≈ 90 "Six-month payback" implies ✕ Verdict: fails — 5×+ measured Musk, SPCX Q3 earnings call depreciation hurdle 8–10 Full-cycle payback (incl. build-time τ, InAI estimate) SPCX: τ ≈ 1 quarter (measured 91–150 days) → ≈ 3.0 yrs Industry norm: τ ≈ 3–4 quarters (8–12 months) → ≈ 3.5–3.7 yrs
The two SPCX quarters are 8-K measured annualizations (end-of-period GW denominator); the build-cost band is the industry all-in basis (attributed statements by NVIDIA and Altimeter); "six-month payback" is the claim's implied requirement (InAI estimate) — the verdict's basis is in the text; the τ timeline follows SemiAnalysis (October 2026 reporting), and full-cycle payback is an InAI estimate.

The Fabless Case: ANT's S-1 and GAAP's Two-Way Distortion

The application-layer (Fabless) evidence comes from the Anthropic S-1. The core efficiency ratio ρ = 1.77 (first nine months of 2026: $1.77 of revenue per $1 of compute and infrastructure expense, on the EDGAR S-1/A as-filed basis; the quarterly path is 1.82 in H1 and 1.72 in Q3 alone, easing as capacity ramps), and the company states continued improvement as a goal. Nine-month operating cash flow was +$6.6B, adjusted operating profit +$4.1B (a 12% margin). The Fabless cash signature is confirmed here: revenue growth keeping pace with compute expansion, and revenue covering compute cost — that is the definition of "cash discipline," and it shows up in the application layer, not the infrastructure layer.

The same filing also supplies an accounting-decomposition case study. Of the $104.039B nine-month net loss, $94.324B sits in one line: "Change in fair value of convertible notes, derivative liabilities, and SAFEs" ($21.936B of it in Q3 alone). The mechanism, per the S-1: the company's valuation rises → "an increase in the fair value of our outstanding convertible promissory notes" → the conversion rights embedded in instruments issued in historical financing rounds (converting into Class N common stock before the offering completes; major holders include Amazon) gain value → the liability is remeasured at fair value → the income statement books a loss. The enormous book loss is, mathematically, a direct consequence of the valuation surging; its bearers are the legacy shareholders (via conversion dilution); and it is non-cash by nature. Strip it out and the GAAP operating loss is just $1.260B (a −4% operating margin); strip out stock-based compensation and the non-cash donation-matching expense as well, and adjusted operating profit is +$4.081B.

Set beside the IDM, the distortion is symmetric: more than half of SPCX's $4.4B Q3 net profit came from non-operating items (books better than reality); ANT shows a $104B book loss while operating near breakeven (books worse than reality). GAAP misleads in both directions at once; unit economics is fooled in neither. That is the direct evidence for Part III's claim that a KPI matrix must replace any single income-statement reading.

Where the Transfer Price Lives: Four Accounting Quantities, Loop Contamination, Two New Signals

Equation 3 says infrastructure-layer revenue and application-layer compute cost are the same money. This section examines the two deformations that money undergoes in practice: disclosure conventions slice it into accounting quantities of wildly different widths (the A-to-D problem), and financing arrangements stop it from being an independently discovered market price (the loop-contamination problem). Pin the viewpoint first: the main thread examines exactly one pair of entities — upstream seller SpaceX (selling compute; the contracting entity is its wholly owned subsidiary SpaceXAI, formerly X.AI LLC, which the S-1/A calls the SpaceXAI Arrangements) and downstream buyer Anthropic (buying compute). Commitments are the buyer's act; orders and revenue recognition are the seller's. ANT's total commitments to all suppliers (call it A) enter only as denominator context. Figure 3 gives the panorama first: the market consists of three kinds of entities and two kinds of edges, and the paper's three equations, four accounting quantities, and L-signals can all be located on it; then we push the lens in on one edge.

Figure 3: Market Entities and the Cross-Sectional Caliber Map

Three kinds of entities, two kinds of edges: orders point right, financing points left; the four quantities A–D and the L-signals are located on the map

Contracts & orders: buyer → seller Financing: funder → recipient Entity 1: Infra as Foundry (builds & sells compute) AI Superscalar Amazon Azure Oracle Google Cloud SpaceX SI IDM: edge folded into books Equation 1: $16.5B/GW·yr NeoCloud CoreWeave Nscale Nebius Entity 2: LLMs as Fabless rents compute, sells intelligence (Eq. 2) Anthropic OpenAI NVIDIA critical silicon: orders in, financing out External equity & bonds +$110.4B (non-supplier sources) External revenue (the real exit) Equation 3: = ρ × compute cost, ρ 1.77 Compute purchases: B 120.5 the Eq. 2 / Eq. 3 edge Direct chip deals (OpenAI ≈ 5 GW committed, etc.) Chip purchases Note facilities $77B (zero drawn) L_d numerator (Google, Amazon, Broadcom) Equity + backstop $6.3B L_u in the field $30B equity + $105B campus guarantee: L_u in Table 5 Cross-sectional caliber chain: four quantities, wide to narrow, on one "compute purchase" edge (Table 4): A 900.9 buyer's total orders B 120.5 this edge (mostly cancelable) C 14.1 seller's backlog, all customers D 7.56 / qtr revenue (valuation cashes here) B ÷ A ≈ 13.4% C ÷ B ≤ 12% C ÷ annualized D ≈ 0.47
How to read the map: orange edges are contracts and orders (buyer pointing to seller); blue edges are financing (funder pointing to recipient). L_u = the blue edge between a pair ÷ the opposing orange edge; L_d = the share of an entity's blue in-edges that come from its own suppliers (denominator: all blue in-edges — for ANT, $110.4B from non-supplier sources plus $0 drawn from suppliers, so L_d ≈ 0 on a drawn basis). An orange-blue cycle between the same pair is a closed loop (Table 5). Entity grouping and edge directions follow Yumei Dou's sketch (Nov 2026). Sources: ANT S-1/A as filed (A, B, note facilities, ρ, external financing); SPCX Q3 8-K (C, D); NVIDIA–OpenAI is on the bases updated in 2026-11 ($30B of funded equity and the $105B PORTS-Pike campus guarantee, see Table 5; the earlier $100B letter of intent never became a contract); CoreWeave's backstop purchase is its filed 8-K. Dollar amounts in $B (A–D strip shown in $B). Google appears both in Entity 1 (cloud) and as ANT's financier — the on-map form of credit entanglement (Table 10, source three).

Four accounting quantities, each with its own subject. The single purchasing relationship "ANT buys compute from SpaceX" leaves four progressively narrower accounting quantities across the two companies' filings; this paper labels them A through D. A, total commitments: buyer ANT's future purchase commitments plus leases to all compute suppliers combined (the S-1/A commitments note: includes options and future minimum payment obligations, "up to" language, tenors as long as twenty years). B, commitment to a single seller: the portion of A allocated to SpaceXAI — an obligation to pay in the future, not a transaction that has occurred. C, backlog: seller SpaceX's signed contracts that will convert to revenue over a foreseeable window (the accounting name is remaining performance obligations) — note it aggregates all customers. D, revenue recognized: what the seller books to the income statement under ASC 606 — the only accounting quantity where valuation finally cashes out. The four quantities form three adjacent ratios, each with its own subject's viewpoint; Table 4 gives the measured values:

Table 4: Four accounting quantities (A–D) on one purchasing relationship, and three adjacent ratios

QuantitySubject and definitionAmount ($B)Adjacent ratio (whose viewpoint, what it reads)
A — Total commitmentsBuyer ANT → all compute suppliers: future purchase commitments plus leases (S-1/A Amendment No. 3: entered into as of and subsequent to 2026-09-30; the 11-02 amendment showed 895.4, of which ≈ $213.8B carries termination rights or contingencies)900.9(denominator context)
B — Commitment to one sellerBuyer ANT → seller SpaceXAI: purchase cap (through 2029, "up to"; ≈ $3.1B/month once capacity fully delivered, expected complete December 2026; non-cancelable only through January and March 2027, thereafter either party may cancel on 90 days' notice)120.5B ÷ A ≈ 13.4% — buyer ANT's view: dispersion of commitments; nearly an even three-way split with Google (121.3) and Amazon (110.0)
C — BacklogSeller SPCX: signed cloud-service contracts (all customers, remaining performance obligations)14.1C ÷ B ≤ 12% — the ceiling on commitment-to-contract conversion: C still includes other customers, so the ANT-only conversion rate is lower (equivalently, B ÷ C ≥ 8.5×)
D — Revenue recognizedSeller SPCX: SI-segment revenue recognized in the quarter (annualized ≈ 30.2)7.56 / qtrC ÷ annualized D ≈ 0.47 — seller SPCX's view: how freely contracts flow into revenue, and how far ahead the book sees (the backlog coverage ratio)

Each ratio belongs to a specific viewpoint, and mixing viewpoints is exactly how the market misreads. B ÷ A ≈ 13.4% — buyer ANT's view: a reality test of the commitment structure. SpaceXAI received only about one-eighth of ANT's total commitments, split almost evenly with Google (121.3) and Amazon (110.0). Two facts together make A look more like an options portfolio than precisely calibrated demand. First, large commitments cost little at the margin (≈ $213.8B carries termination rights or contingencies; the SpaceXAI tranche becomes 90-day-cancelable once past the early-2027 window). Second, in Equation 3's grammar, compute input is a revenue multiplier (application revenue = ρ × compute cost): the larger the commitment, the larger the revenue outlook the market extrapolates through ρ — a large commitment is itself narrative leverage. If SpaceX's capacity under construction will matter enormously to U.S. compute over the next three years, yet ANT allocates it only an eighth of its commitments, then either ANT's true demand is far smaller than its total commitments, or the allocation does not yet reflect true capacity dependence. Both readings point to the same conclusion: A (900.9) must not be read as demand. To be clear, this is an interpretation, not a finding — the paper states it as a testable hypothesis. If commitments are mostly narrative, expected conversion (C ÷ B) stays low, allocations stay even, cancellation rights stay in place; if they are real demand, conversion should rise quarter by quarter. The test is on the 2027 checklist.

C ÷ B ≤ 12% — the viewpoint switches from buyer to seller: the ceiling on commitment-to-contract conversion. Of the $120.5B commitment, less than an eighth has so far crystallized into signed contracts countable in the seller's backlog (and the 14.1 still includes other SPCX customers, so the ANT-only rate is lower). Part of the gap is created by the accounting definition itself: backlog (remaining performance obligations) by definition counts only the non-cancelable portion, and most of B carries cancellation rights and never enters the line. The gap between the three quantities is therefore not merely "not yet converted" — it is the accounting developer-fluid that makes commitment hardness visible. Headlines live at B ("up to $120.5B mega-deal," "nearly $900B of compute commitments"); the order book lives at C; between them stretches a distance of at least 8.5×, mostly cancelable — and any re-marking of that distance is released as volatility.

C ÷ annualized D ≈ 0.47 — seller SPCX's view: how freely contracts convert, and how far ahead the book sees. Annualized revenue recognition (≈ 30.2) has overtaken backlog (14.1): revenue is being generated mostly by short-cycle, newly activated contracts converting on the spot — conversion is frictionless and explosive, but the order book sees no further than half a year, which explains the extreme dispersion in sell-side estimates for this company. Contrast the other extreme: CoreWeave's FY2025 revenue of $5.1B against a $60.7B backlog — backlog ÷ annual revenue ≈ 12× — a valuation suspended almost entirely on unexecuted contracts, with highly concentrated counterparties. The same metric moved to an application-layer subject: ANT's backlog (RPO) as a seller of intelligence is $13.8B, about 75% recognizable within 12 months (per the S-1/A), or ≈ 0.20 against $67.8B annualized revenue — lower than SPCX, but for a different reason: under usage-based billing the order book is structurally thin, and forward visibility rides on retention and usage rather than contracts. A very low reading therefore has two distinct causes (structural thinness from the billing model, versus thinness from short contracts converting immediately), and interpretation must carry the business-model context. The two extremes (0.47 and 12) become the two alert boundaries of the backlog coverage ratio: too low, insufficient forward visibility; too high, valuation suspended on unexecuted contracts. The healthy band lies between. The S-1/A also makes commitment "hardness" directly readable in tiers for the first time: Google ($121.3B), Amazon ($110.0B), and Microsoft ($44.9B) are take-or-pay "regardless of actual usage"; Nscale ($44.6B) and Lambda ($41.3B) bill only on delivery and acceptance; SpaceXAI ($120.5B) becomes cancelable on 90 days' notice once past the early-2027 window. Three completely different strengths of obligation lie in the same "commitments" line. One more directly readable number: ANT's commitments due within the next 12 months are about $70.7B (S-1/A Amendment No. 3; $69.8B in the 11-02 amendment) against $67.8B of Q3 annualized revenue — a ratio of about 1.04. One year of committed bills roughly equals one year of revenue.

Now the loop. Cross-layer mutual financing is already common talk in 2026. What this paper adds is not another description but three things: placing it inside the equation system (it attacks Equation 3 — when the supply side funds the demand side and then backstops its capacity, $/GPU·hr is no longer an independently discovered market price; the transfer price still holds in accounting but is contaminated in its pricing sense); giving it two symmetric, filings-computable definitions; and calibrating the red line with a same-basis historical precedent.

Table 5: The two loop signals — definitions and same-basis calibration

Loop signal (viewpoint)Definition2026 same-basis calibrationTelecom-precedent red line (1999–2001, same basis)
Vendor financing ratio Lu
(supply-side view: money I give my customer ÷ money my customer gives me)
A supplier's financial support to a single customer (equity investment + backstop purchase commitments + loan guarantees) ÷ that customer's purchases from the supplierThe largest case is still NVIDIA and OpenAI, and its basis changed twice within a year. Old basis (void): the September 2025 letter of intent, an investment cap of $100B released in tranches against 10 GW, gave Lu ≈ 25–30%; the letter never became a contract. Funded-equity basis (grade three): NVIDIA took a direct $30B equity stake in the round that closed in March 2026, against the ≈ 5 GW of NVIDIA systems OpenAI committed to at the same time (≈ $35–40B per GW): Lu ≈ 15–17%. Signed-guarantee basis (grade two, not yet in effect): NVIDIA provides up to $105B of credit support for the land, power and shell of the first 4.25 GW of the PORTS-Pike campus in Ohio, which SB Energy builds and owns, leases to OpenAI for 20 years, and which hosts NVIDIA systems exclusively; against the ≈ $149–170B of NVIDIA product those 4.25 GW imply: Lu ≈ 62–71% (InAI estimate; each building's guarantee takes effect on delivery, and no guarantee payments had been made as of 2026-10-25). The ANT S-1/A adds three facility-basis data points (supplier-provided note facilities ÷ purchase commitments to that supplier, all undrawn as of 2026-09-30): Google $20B ÷ $121.3B ≈ 16%; Amazon $15B ÷ $110.0B ≈ 14%; Broadcom $42B ÷ $161.2B of leases ≈ 26%. Every reading remains below the 130% red line, but the largest is now halfway to itNortel's financing to downstream carrier customers reached 130% of the equipment price those customers were buying: the financing exceeded the sales contract itself, with the excess effectively funding the customers' operating expenses. Nine equipment vendors carried ≈ $25.6B of such exposure in total (McKinsey, end-2000)
Supplier capital share Ld
(demand-side view: how much of my money comes from the people selling to me)
The portion of a buyer's external financing that comes from its own suppliers (convertibles, equity, backstop commitments) ÷ total external financing; read jointly with the company's coverage ratio (revenue ÷ capex)ANT: with amounts blank in the original S-1 the ratio could not be computed; the S-1/A completes it and makes it readable from a single document for the first time: three supplier note facilities totaling $77B, zero drawn as of September 30 (Ld ≈ 0 on a drawn basis), against $110.4B of equity financing in the same period; the historical Amazon convertible notes convert to common stock before listing. The November 2026 concurrent placement adds a directly readable component: seven suppliers buy $26B of Class N equity at the offer price, about 25% of the money the company raises at listing (about $79.4B from the public offering plus the $26B placement). CoreWeave: supplier NVIDIA both holds equity and provides a $6.3B backstop purchase (filed by 8-K), while the company's coverage ratio (revenue ÷ capex) stays below 1 — a triple overlayThe same facts from the demand side: CLEC carriers that sustained purchasing on vendor financing — roughly four dozen bankruptcies, 2000–2003; their supplier Lucent provisioned ≈ $8.1B of customer loan commitments and ultimately wrote off ≈ $3.5B — a realized loss rate of ≈ 43%

Two clarifications are mandatory. First, caliber discipline — and the caliber itself has a hardness ladder. The same Lu comes in three grades, by how real the money is. Grade one, letters of intent: unsigned intended amounts. NVIDIA's $100B for OpenAI was of this kind, and it never became a contract; this paper keeps it only as a lesson in basis: a grade-one reading can vanish entirely. Grade two, signed facilities and guarantees: contracted but not yet used. ANT's three note facilities (zero drawn as of 2026-09-30) and NVIDIA's $105B PORTS-Pike guarantee (not yet in effect, nothing paid) sit here. Grade three, funds disbursed: NVIDIA's $30B of OpenAI equity, the $26B of equity in ANT's concurrent placement (once closed), and Nortel's 130%, which was lending that actually went out and was ultimately partly written off. Hardness rises by grade, and cross-grade comparison must state the grade; within grade three, equity and credit must also be kept apart: what AI suppliers have disbursed so far is equity, where losses are absorbed at market value by the investor and trigger no default, while the telecom red line was credit, written off when the customer defaulted. That is both the basis of the "below the red line" judgment and its qualifier. The only numbers compared against 130% in the table are structurally matched (financing support ÷ corresponding sales). Other circumstantial readings (NewStreet's estimate that ≈ $10B of NVIDIA investment ties to ≈ $35B of revenue, about 27% of the prior year's revenue; NVIDIA's own statements on 2026-11-17 that frontier labs are about 15% of its data center revenue, that total guarantees stand at $118B, and that it has a further $32B of conditional commitments to neoclouds; the top two customers at 39% of revenue; $2B to Nebius, ≈ $2B to xAI) use varying calibers and serve only to show the loop's breadth — they are not compared to 130%. Second, what the red line means: when Lu and Ld are high at the same time and the buyer's coverage ratio is below 1, the marginal dollar of demand is effectively funded by the supply side itself. Reported revenue and unit economics (including this paper's $16.5B/GW·yr calibration) then contain a self-dealing component, and the transfer price loses its price-discovery function. On the same-basis calibration, AI's current largest Lu has risen from 25–30% on the letter-of-intent basis to about 62–71% on the guarantee basis (grade two, not yet in effect); it is still below telecom's terminal 130% (grade three, funds disbursed), but the structure is isomorphic and moving one way: within a year the largest reading has travelled half the distance to the red line. These two signals therefore must enter the dashboard (Part III, Table 11). Disclosure improves things immediately: with the original S-1's blanks, Ld could not be computed; once the S-1/A disclosed supplier facilities, commitments, and drawn status side by side, both signals became readable from a single document for the first time. Every step of disclosure removes one blind spot from the dashboard.

Figure 4: The A-to-D Exchange Rates and the Loop's Caliber Ladder

Top: four accounting quantities and three adjacent ratios on one purchasing relationship, bar width ∝ log₁₀(amount), each ratio tagged with its viewpoint; bottom: the hardness ladder of Lu

Four quantities A–D: one purchasing relationship projected onto two sets of books ($B; bar width drawn in log₁₀) A total commitments: buyer ANT → all suppliers 900.9 B commitment: ANT → SpaceXAI (mostly cancelable) 120.5 C backlog: SPCX, all customers 14.1 D revenue recognized (annualized ≈ 30.2) 7.56 B ÷ A ≈ 13.4% (buyer's view: commitment dispersion) Near three-way split with Google, Amazon: A is not demand C ÷ B ≤ 12% (ceiling on commitment-to-contract conversion) 14.1 includes other customers; ANT-only rate is lower C ÷ annualized D ≈ 0.47 (seller's view: contract to revenue) Revenue overtakes backlog: instant conversion, <6-month visibility Headlines live at B, the order book at C, valuation cashes at D Loop signal L_u's caliber-hardness ladder: 2026 readings vs the telecom red line (1999–2001) L_u (financing support ÷ matching sales), soft to hard Grade 1, letter of intent (void): NVIDIA–OpenAI 25–30% Grade 2, undrawn notes: Google 16%, Amazon 14%, Broadcom 26% Grade 2, guarantee (not in effect): NVIDIA–OpenAI ≈ 62–71% Grade 3, funded equity: NVIDIA–OpenAI ≈ 15–17% Grade 3, funds disbursed — telecom's endgame: Nortel up to 130% Financing exceeded equipment sales; nine vendors, $25.6B exposure Outcome: Lucent committed $8.1B, wrote off $3.5B (≈ 43% loss rate) Telecom was customer credit, written off on default; AI suppliers have so far disbursed equity, marked to market
Top chart: bar width ∝ log₁₀($100M units): the linear proportions of 900.9 : 120.5 : 14.1 : 7.56 ($B) are about 119:16:1.9:1 — drawn linearly, the last two bars would vanish, so the chart uses logs and prints the multiples as text. The commitment total is ANT S-1/A Amendment No. 3 as filed (EDGAR accession 0001628280-26-074338, 2026-11-17) and the per-supplier amounts are the 2026-11-02 amendment as filed (accession 0001628280-26-068676); backlog and revenue recognition are SPCX 8-K as filed; CoreWeave is its FY2025 filing ($5.1B revenue, $60.7B backlog); telecom data is an industry-history compilation from Lucent/Nortel annual reports, SEC files, and McKinsey estimates.

The Industry P&L: External Revenue Covers Only 20–30% of the Build

Multiply the three equations by installed capacity, and single-company decomposition scales up to an industry P&L. Table 6 first pins down the scale basis (public estimates diverge wildly — itself evidence for Part III):

Table 6: The scale basis (the current state of divergent public estimates)

Scale basisEstimateSource
U.S. installed base (end-2027, powered)35–50 GWInAI estimate (U.S. adds ≈ 12–18 GW in 2027)
Global additions 2026 (ex-China)≈ 25 GWSemiAnalysis (2026-10)
Global additions 2027 (ex-China)51.2 GW net IT capacity (previously 43)SemiAnalysis; publicly challenged by Gerstner, who calls ≈ 25 GW more realistic: top-tier shops differ by 2×
Under construction60+ GWSemiAnalysis; flags the 2027–28 bottleneck possibly rotating from power back to chips, advanced packaging, and HBM
Global AI-dedicated base, 2026≈ 20–44 GW (bases mutually irreconcilable)Public-estimate roundup: direct evidence of measurement failure

Before compiling, nail down the cost basis — it sets the denominator of every ratio that follows:

Capex basis definition: Foundry- and IDM-layer capex means capitalized construction investment (land, shell, power and cooling, servers and GPUs purchased in the year, including finance leases) — it goes on the balance sheet and depreciates annually; the four hyperscalers' $720–745B guidance is this basis. Two leakages run in opposite directions: facilities built by developers or SPVs and leased by the clouds (Hyperion-type projects) sit on the developer's books, outside big-four capex (understatement); big-four capex also contains non-AI routine infrastructure (overstatement) — the net direction is understatement. The Fabless layer has no material capex: its compute cost is annualized rental and usage expense booked to cost of revenue (ANT's $5.6B Q2 compute spend is exactly this), and long-dated commitments are locked-in future expense (debt-like), not capex. Consolidation rule: construction investment is counted once, on the layer that holds the asset; Fabless rent is an internal transfer, an elimination item. The industry denominator can only be build-side capex in total — adding rental fees would double-price the same infrastructure. On a per-GW basis the two relate as: $40–50B build (one-time) versus ≈ $16.5B annual rent, about 2.7–3× — the 2.7-year payback written another way.

The industry P&L must be compiled unfolded, by the three forms, because unfolding the folding is the whole point: the Foundry sits midstream — its revenue is this form's output and simultaneously the Fabless form's cost input (Equation 3's reconciliation); the IDM folds that reconciliation inside its own books. Table 7 unrolls along the chain:

Table 7: The industry P&L unrolled along the reconciliation chain

Form (position in the chain)Revenue (output)Cost (input)Coverage basis and 2026 readingRepresentative readings
Fabless (downstream)
the industry's main external-revenue exit
tokens × $/token: real sales to outside the industrythe GPU·hr bill (= midstream's output)ρ = revenue ÷ compute and infrastructure expense: 1.77, above 1 — healthyANT: ρ 1.77, nine-month OCF +$6.6B; OpenAI revenue debated at $50–70B (annualized)
Foundry (midstream)
output = downstream's input (elimination item)
GPUs/GW × $/GPU·hr × utilizationbuild cost per GW + depreciation $8–10B/GW·yrRevenue ÷ capex, split in two: hyperscalers' builds are covered by cloud and advertising cash flows; independent neoclouds stay below 1, dependent on debt and supplier financingCoreWeave FY2025: revenue $5.1B, net loss $1.2B, backlog $60.7B (12×); the supplier both holds equity and backstops
IDM (the folded form)
the reconciliation vanishes inside the books
reports revenue per GW directly ($16.5B)both cost equations stackedOperating cash flow ÷ capex ≈ 0.27 (SPCX nine months: $13.1B ÷ $48.1B)SPCX: nine-month FCF ≈ −$35B; external financing needs $30–55B per quarter (InAI three-scenario model)

Only after consolidation and elimination does the industry's true coverage show — and it must be read on two bases. Merge the three forms into one industry statement and Equation 3's transfer price eliminates: what midstream charges downstream for compute is an internal transfer, not external revenue. The industry's genuinely external revenue is only the Fabless layer's sales (plus infrastructure sales to non-AI tenants), while the build investment all stays in. Numerator: on Gerstner's BG2 basis, frontier labs plus inference companies will book about $200B of external revenue this year.

Basis one, the funding gap (cash-flow-statement view): external revenue ÷ current-year build investment. Take the four hyperscalers' 2026 guidance of $720–745B as the denominator (a conservative subset by construction) and get ≈ 0.27; with the denominator broadened, roughly 0.2–0.3. Reading: of what the industry will spend on building this year, only two to three tenths is covered by this year's external revenue — the rest is advanced by equity and credit markets. This basis answers "who is paying," not "is it profitable," because this year's build serves the next four or five years of revenue.

Basis two, the income statement (depreciation view): external revenue ÷ annual depreciation. Swap the denominator to depreciation. Against the steady-state annual depreciation implied by this year's build ($720–745B ÷ 5 ≈ $144–149B), coverage is ≈ 1.3–1.4; against depreciation on the in-service base (20–44 GW globally × $8–10B per GW·yr ≈ $160–440B), it is 0.5–1.2. Reading: current external revenue roughly just pays the depreciation itself, with zero left for power operations and zero return on capital. The two bases together form the complete judgment — two to three tenths self-funded on a cash basis, bare depreciation coverage on a profit basis. Each layer's own books can look good simultaneously (internal transfer revenue counts in each), which is the accounting reason "every layer's statements hold individually while the industry as a whole depends on external financing" can coexist.

What this level means needs a mature division-of-labor industry as the benchmark: semiconductors. The derivation: TSMC's 2025 revenue was $122.4B, capex $40.9B, D&A ≈ $22.1B (4Q25 management report, as published), about 70% of global pure-play foundry revenue (TrendForce); extrapolating by share, global foundry annual depreciation ≈ $32B and annual capex ≈ $58B. Downstream, global fabless revenue in 2025 was about $320–400B (NVIDIA alone ≈ $200B; total to be re-verified before final). TSMC skews to leading edge with higher depreciation intensity than peers, so extrapolating from it overstates the denominator — the semiconductor readings below are conservative:

Table 8: The semiconductor benchmark (extrapolated from TSMC actuals)

MetricSemiconductors (2025 actual & extrapolated)AI (2026, this paper's basis)Gap and meaning
Downstream revenue ÷ upstream annual depreciation10–12× (320–400 ÷ ≈ 32)0.5–1.4About an order of magnitude: in a mature division of labor, downstream revenue runs 10×+ upstream depreciation
Downstream revenue ÷ upstream current-year capex5.5–7× (÷ ≈ 58)0.27About 20×
Equipment accounting life5 years (TSMC equipment; buildings 20)4–5 years (GPUs)Identical: the depreciation period is not where the difference lives
Economic life multiple (economic life ÷ accounting life)3–4: equipment keeps producing on mature nodes for 10–20 years after depreciating to zero, entering a post-depreciation profit era (mature fabs become cash cows)≈ 1: generational performance-per-watt jumps drive retirement; economic life roughly equals or undershoots accounting lifeThe root variable behind the order-of-magnitude coverage gap: semiconductors' 5-year depreciation is front-loaded conservative accounting; AI's 5-year depreciation is real wear
Long-lived assets' share of capexHigh: 20-year shells and cleanrooms weigh heavily, reused across tool generationsLow: 70–80% is 4–5-year-life IT equipmentThe structural difference amplifies the row above

The endogenous CAGR requirement: a growth rate forced out by accounting structure. An economic life multiple of about 1 means the AI industry has no semiconductor-style post-depreciation cushion and cannot crawl toward high coverage over twenty years; revenue must catch up with the build within the first asset generation's depreciation life. If construction holds its current pace (the big four starting around $730B/yr), steady-state annual depreciation reaches that same order within five years, and the requirement on external revenue starting from $200B is:

Table 9: The endogenous CAGR requirement (build pace maintained)

Scenario: coverage external revenue must reach in five yearsCoverage meaning (multiples are of the AI industry's own annual depreciation)Implied external revenueImplied 5-yr CAGR
Scenario 1: breakeven line (depreciation coverage = 1)External revenue exactly equals industry annual depreciation: zero power opex, zero return on capital≈ $730B≈ 30%: the floor
Scenario 2: sustainability line (coverage 1.5–2)Covers depreciation plus power operations plus a minimal return: the industry starts self-sustaining$1.1–1.5T≈ 45–50%
Scenario 3: the semiconductor line (coverage 5–10)Reaches a mature industry's downstream-over-upstream coverage: Table 8 measures semis at 10–12×; the conservative band here is a requirement placed on AI, not a statement about semis$3.7–7.3T≈ 80–100%

This table explains a market behavior that previously lacked one. When Gerstner pressed the labs on BG2 — "$200B this year, $400B next, $800B the year after; can you double every year?" — it sounded harsh, but a CAGR around 100% is exactly what the asset structure "5-year depreciation × 1× economic life" endogenously requires for a semiconductor-grade margin of safety. The market's aggressive revenue expectations are not sentiment pricing; they are the algebraic consequence of the asset structure. Semiconductors were never asked for such growth, because their assets keep earning after depreciating to zero. Conversely, anything that pushes the economic life multiple away from 1 (a used-GPU market, older generations rotating to inference, long-term CUDA software support) directly lowers the growth rate the whole industry is required to deliver — which is why the economic life multiple must enter the KPI matrix.

Finally, rewrite Table 9's revenue requirement as a physical requirement via Equation 1. Summed across the industry, revenue still obeys Equation 1's structure: revenue = GW × utilization × tokens/joule × $/token. Measured token prices keep falling (same-capability API prices decline generation over generation, Artificial Analysis basis); call the annual decline X%. For revenue still to grow at a 30% CAGR (Table 9, Scenario 1), the product GW × utilization × tokens/joule must grow at roughly "30% + X%" per year. Thus rewritten, the revenue requirement becomes a growth requirement on three physical quantities — installed base, utilization, and energy efficiency. The argument for physical-layer metrics entering the valuation system is now complete: the basis is not analogy but the algebraic decomposition of the revenue-growth requirement onto these three variables. Research #15 and #16 measured architecture-side tokens/joule improving by several-fold to tens-fold per generation; tracking that pace of improvement is tracking whether the industry can meet Table 9's requirement.

Historical calibration: why high-speed buildouts turn disastrous. The comparison basis first: telecom's industry-wide coverage in 2000 was not low — the huge installed voice and data revenue base far exceeded that year's capex (aggregate coverage above 1). The disaster did not strike the industry as a whole; it struck the newly built sub-sector with coverage far below 1, wholly dependent on external debt and vendor financing (CLECs and new fiber networks). Catastrophe was three things stacking in that sub-sector: vendor financing inflating demand (Nortel's financing at 130% of equipment sales; nine vendors, $25.6B exposure), revenue-side cash flow breaking (roughly four dozen CLEC bankruptcies), and credit losses realizing (Lucent: $8.1B committed, $3.5B written off, a 43% loss rate). Against the present, one uncomfortable fact: the AI industry's post-elimination funding-gap coverage (0.2–0.3) sits exactly where telecom's new-build sub-sector alone sat. But two structural differences hold: the downstream Fabless layer already self-covers operationally (ρ above 1, positive operating cash flow — the CLECs never managed it), and the largest midstream builders fund construction out of cloud and advertising profits rather than debt alone. The systemic-risk verdict is therefore: risk is unevenly distributed, with stress concentrated in two places — independent neoclouds (coverage below 1 + backlog at 12× + a supplier closed loop, a triple overlay) and the IDM's credit channel ($30–55B of external financing per quarter). The dashboard's most critical moving indicators are the independent foundries' coverage ratio and the loop signals, then the IDM's financing cadence. Fabless ρ is the last line of defense: if it breaks below 1, the whole chain's external-revenue exit closes — and only then does the telecom new-build sub-sector's ending replay industry-wide.

Figure 5: The Industry P&L After Consolidation

Units: $B/yr. With the transfer price eliminated, industry external revenue versus capex: ≈ 0.2–0.3× today, and even the full-capacity ceiling falls short of 1

≈ 200 Current industry external revenue Frontier labs + inference cos. (Gerstner basis) 578–825 Full-capacity revenue ceiling U.S. 35–50 GW × $16.5B 720–745 Big-four cloud capex 2026 guidance, combined 250–300 IG issuance 2026E 200+ Private credit stock Implied annual depreciation of this year's build ≈ 144–149 Funding-gap basis ≈ 0.2–0.3 (green ÷ orange); depreciation basis ≈ 1.3–1.4 (green vs red dashed line): just pays depreciation, zero return on capital Midstream's compute charges to downstream are internal transfers — eliminated on consolidation, not external revenue
External revenue follows Gerstner's basis on BG2 (2026-10-19): frontier labs plus inference companies at roughly $200B/yr. The ceiling bar's light segment is the 50 GW case, the dark segment 35 GW (InAI estimates); the red dashed line is current-year build ÷ 5-year depreciation life. Bar heights and line positions are to scale.

Part III · Information Limits: The Three Sources of Volatility

The first two parts established how to measure; this part explains why the market currently cannot. The limits on how markets receive and analyze information produce volatility, and those limits have three sources. Each relates to the three equations differently — an equation misused, an equation unusable, an equation unreliable — and each has already been evidenced above:

Table 10: The three sources of volatility (mechanism, evidence, relation to the equations, this paper's remedy)

Source of volatilityMechanismEvidence in this paperRelation to the three equationsThis paper's remedy
One: observation mismatchFamous AI companies account for roughly a third of demand but all of the observation weight; their failures would make the market reprice the entire base — including the two-thirds that did not failHuang's decomposition (BG2): traditional computing and classic ML ≈ 1/3 (growing 30–40%/yr); 5–6 frontier labs ≈ 1/4–1/3; thousands of innovators ≈ 1/3 (80% on open-source models)Equation misused: for industry-wide revenue scale and outlook the right tool is Equation 1 (installed base × unit revenue, bottom-up); the market instead extrapolates the industry base from a few famous companies' Equation 3 readings (application revenue and ρ). The famous companies are indeed the most visible window with the most legible financials — but the window is not the baseGive the market tools to observe the base directly: the KPI matrix (Table 11)
Two: caliber folding and slicingTwo things stack: the IDM folds the transfer price inside its books, and disclosure slices one purchasing relationship into four separately written quantities A–D — headlines, order books, and income statements each live on a different oneFolding: in SPCX's 6.3-to-16.5 single-quarter jump, the Cursor consolidation (a folding-boundary move) contributed ≈ $1.4B of unit increase; slicing: Table 4's adjacent ratios 13.4%, ≤12%, 0.47 — reading one relationship whole requires splicing two companies' filingsEquation unusable: for a folded entity (SPCX SI), outsiders see only Equation 2-form contract revenue (rentals to ANT and others) and Equation 1-form macro estimates; the internal variables Equation 3 needs (application revenue ÷ internal compute cost) cannot be separated from the books. The equation isn't wrong — the applicable one can't be appliedThe A–D quartet and denominator disclosure standard (Table 12)
Three: credit entanglementCross-layer financial exposures are mutually held, so a single headline propagates across layers mechanically; the loop contaminates the transfer price's market characterANT's three suppliers are simultaneously standby creditors (note facilities totaling $77B, S-1/A); Amazon historically held its convertibles while selling it compute; NVIDIA earns ≈ 27% of revenue from customers it co-invests in (NewStreet estimate) and gives CoreWeave a $6.3B backstop purchase (unrented capacity leased by the supplier itself, filed by 8-K); the L-signals are calibrated (Table 5)Equation unreliable: related-party financing pulls $/GPU·hr away from an independently discovered market price, and ρ's denominator distorts with it. Equation 3 was meant to measure model competitiveness and monetization success (application revenue = ρ × compute cost, both factors amplifying revenue); contaminated, it still holds in accounting but is no longer a clean measure of successDeal-by-deal tagging of related financing and L-signal disclosure (Table 12)

The three sources share one property: none is a demand problem; all are information problems. Requiring a handful of famous companies' application-layer revenue to clear an extreme bar before the whole AI trade counts as safe — in an industry where thousands of companies are already diffusing into every sector — is a limit of information judgment, not of the industry. The direction of this paper's remedy is therefore not to predict who is right, but to supply capital markets with the observation variables they are missing.

Figure 6: A Third of the Demand, All of the Dashboard

The structure of observation mismatch (source one): economic weight versus observation weight

Economic weight (AI infrastructure revenue mix) Traditional compute & classic ML ≈ 1/3, growing 30–40%/yr 5–6 frontier labs ≈ 1/4 to 1/3 Thousands of innovators ≈ 1/3, 80% on open-source models Observation weight (market attention & news flow, illustrative) Famous AI companies' earnings and headlines A famous company fails → the market reprices the whole base (incl. the unfailed two-thirds) → an information failure, not a demand failure
The demand mix is Huang's attributed statement on the BG2 podcast (2026-10-19), using NVIDIA's revenue mix (self-described ≈ 80% of global AI infrastructure revenue) as proxy; the attention distribution is illustrative.

The New Valuation KPI Matrix

Table 10 diagnosed the three sources of volatility; the remedy comes in two steps — this section proposes the metric system, the next the disclosure requirements that make it computable. Table 11 is the AI valuation metric system this paper advocates to investors and analysts. None of its rows is proposed from thin air: the four financial-layer items were calibrated on public filings in the first two parts; the three signal-layer items carry alert boundaries or precedent red lines; the three physical-layer items have public benchmarks and measured ranges from InAI's own research. What this section advocates is turning these one-off research calculations into regular, quarterly-updated indicators:

Table 11: The new valuation KPI matrix

LayerKPIEquation variableCurrent calibrationQuestion it answersSource
FinancialRevenue per GW·yrEquation 1, left side≈ $16.5BIs capacity monetizing?SPCX 8-K
FinancialBuild cost per GW / depreciation per GW·yrEquation 1's cost counterpart$40–50B / $8–10BCapital at risk per unit; the profit hurdleIndustry attributed basis / InAI estimate
FinancialImplied payback periodEquation 1 ÷ cost terms2.7 yrs / ≈ 4.2 yrs; full-cycle (incl. build-time τ): SPCX ≈ 3.0 yrs vs industry norm 3.5–3.7 yrsDoes debt duration match the asset? Adjudicates false claims (Figure 2)InAI estimate
Financialρ (revenue per compute dollar)Equation 31.77 (nine months; 1.72 in Q3)Does the application layer cover compute (cash discipline)? Is the industry's external-revenue exit open?ANT S-1/A
SignalBacklog coverage ratio (seller backlog ÷ annualized revenue)Quality of Equation 3's accounting quantitiesANT 0.20 / SPCX 0.47 / CoreWeave 12 (the latter two are the alert boundaries)Too low: weak forward visibility (two distinct causes — structural thinness of usage billing, as ANT; instant conversion of short contracts, as SPCX — read with business-model context); too high: valuation suspended on unexecuted contractsSeller filings (Table 4); ANT S-1/A (RPO $13.8B)
SignalVendor financing ratio LuEquation 3's price purity (supply-side view)NVIDIA–OpenAI: ≈ 15–17% on funded equity (grade 3) and ≈ 62–71% on the campus guarantee (grade 2, not in effect); the earlier 25–30% letter-of-intent basis is void; ANT S-1/A three points: Google ≈ 16%, Amazon ≈ 14%, Broadcom ≈ 26% (grade 2, signed facilities, zero drawn); telecom red line 130% (grade 3, funds disbursed)Is the transfer price still an independent market price?Both parties' filings + same-basis precedent (Part II)
SignalSupplier capital share LdEquation 3's price purity (demand-side view)ANT now readable from the S-1/A alone: ≈ 0 drawn on note facilities ($77B undrawn); the $26B concurrent placement is supplier equity, about 25% of the money raised at listing; read jointly with coverageWho funds the marginal dollar of demand?ANT S-1/A
SignalEconomic life multiple (economic life ÷ accounting life)The time structure of Equation 1's cost termsWafer equipment 3–4 / GPUs ≈ 1Does the asset keep earning after depreciating out? It sets the revenue CAGR the industry is required to deliver (30% floor to 100%)TSMC depreciation policy + industry comparison (Part II)
PhysicalBytes per tokenThe architectural determinant of tokens/joule2.3–48× reduction in the new generationDirection of architectural efficiencyResearch #16
PhysicalTokens per jouleEquations 1 & 2, the physical bridge0.1–2.5 (with boundaries)The conversion rate from electricity to intelligenceTokenPowerBench et al.
PhysicalBytes per taskThe link between Equation 1 and the demand sidebytes/token × tokens/task; observable proxy: bytes/token × tokens/request (request ≠ task — an agent issues many requests per task, so the proxy systematically understates tokens/task but is directionally valid)The customer-side physical cost per task; discriminates whether token demand growth is task-driven or verbosity-drivenResearch #16 × OpenRouter (token volume and request counts both public)

All five signal-layer indicators are relational — undefined without naming the parties. The same signal has distinct readings and observability across the three kinds of entities (Table 3's three forms, with the chip layer folded into the supply side); Table 11a maps the current state of this 3 × 5 matrix: of 15 cells, 6 measured, 3 observed-at-zero (no arrangement seen), 4 not computable for lack of disclosure, 2 not applicable. The observability gaps are not a defect of the matrix — they are the brief for Table 12's disclosure advocacy: every dark cell corresponds to one disclosure; disclose it, and the cell lights up.

Table 11a: Entity × signal observability matrix (as of 2026Q3)

Signal (rows) × entity (cols)Entity 1 — Foundry (SPCX SI; CoreWeave as contrast)Entity 2 — Fabless (ANT)Entity 3 — Chips (NVIDIA)
Backlog coverage ratio0.47 (SPCX) / 12 (CoreWeave): measured0.20 (RPO $13.8B, S-1/A as filed): measured; order books are structurally thin under usage billingTotal backlog undisclosed: disclosure gap
Vendor financing ratio LuNo customer-financing arrangements observedNo customer-financing arrangements observed15–17% (to OpenAI, funded equity) and 62–71% (campus guarantee, not in effect); equity plus backstop to CoreWeave: the measured home turf
Supplier capital share LdCoreWeave's triple overlay observable but the ratio not computable: disclosure gap; SPCX undisclosedDrawn basis ≈ 0 ($77B of facilities); $26B of supplier equity in the listing placement: measuredNo supplier-financing dependence observed (upstream runs on prepayments and long contracts)
Economic life multiple≈ 1 (GPU assets on this layer's books): measuredN/A (asset-light; borne indirectly through long-dated commitments)N/A (but sets the whole industry's generational cadence)
12-month commitments-to-revenueCommitments to chip and equipment vendors not separately disclosed: to be spliced≈ 1.04: measured10-K supply obligations can be spliced: to be spliced

The dashboard's priorities follow the risk-distribution verdict at the end of Part II (the "historical calibration" passage): first priority is the independent foundries' coverage ratio and the L-series loop signals (where stress is most concentrated today); second, the IDM's external financing cadence ($30–55B per quarter); the last line of defense is Fabless ρ — if it breaks below 1, the industry's external-revenue exit closes and the consolidated statement's 0.2–0.3 funding-gap coverage loses its path to improvement. The long-horizon variable is the economic life multiple: if it moves away from 1 upward, the revenue growth required of the industry falls accordingly. Table 11 has been built as an interactive dashboard: each KPI auto-grades against the alert boundaries given in this paper and supports plugging in scenario values for instant re-grading, published alongside this paper and the per-GW quarterly readings page.

Where the physical-layer metrics slot into the stack is given directly by the decomposition at the end of Part II: the revenue-growth requirement lands on the product GW × utilization × tokens/joule — and the three variables face different kinds of constraints. GW is bound by power and interconnection (Table 6: ≈ 50 GW of announced chips, ≈ 35–40 GW of land-power-shell that can actually land); utilization has a physical ceiling of 100%; only tokens/joule has engineering-driven order-of-magnitude headroom. The "30% + X%" growth burden therefore falls mainly on it.

Who solves tokens/joule: three parties, with measurable leverage. The variable is determined jointly by three sides. Chip generations: H100 to B200 lifts bandwidth about 2.4×. Model architecture: Research #15 measured a single architectural change (MLA) compressing attention memory ≈ 57× and raising arithmetic intensity from ≈ 1 to 55 — letting the prior chip generation match the next one in long-context inference; Research #16 measured new-generation bytes/token falling 2.3–48× across ten labs. The serving stack: inference engines and batching cut energy per token 25–40%, FP8 quantization another ≈ 30% (TokenPowerBench basis). The magnitudes are wildly unequal: tens-fold per generation on the architecture side, 2–3× per generation on the chip side, tens of percent on the software side — the biggest lever on tokens/joule is in the model makers' hands, not the chip makers'. This adds a structural fact to the paper's industry conclusion: the "30% + X%" revenue-growth burden is carried by the downstream Fabless layer (Table 7), and the biggest tokens/joule lever happens to sit in that same layer — burden and lever on the same floor, so the industry has a self-improvement channel. The GW lever sits midstream and with the grid; the utilization lever sits with demand; neither is controlled by any single party.

What the measured −0.1 means in this paper: the quality of token demand. Research #16's raw findings were two: within a single lab, efficient-architecture siblings cut cost per task 6–7× (architecture works); across labs, the correlation between bytes/token and cost per task is only about −0.1 (architecture alone does not decide). The reason is the caliber chain: cost per task = bytes/token (architecture-determined) × tokens/task (how many tokens the model spends to finish a task — highly variable across labs) × pricing strategy. Put into Equation 1, the implication is: for tokens/joule improvements to convert into industry external revenue, they must pass through "cost per task falls → customer adoption rises"; if the industry's token-volume growth comes mainly from models getting wordier (tokens/task rising) rather than task volume growing, the quality of the revenue growth is low. The physical layer therefore needs two tiers of metrics, mutually non-substitutable (that is precisely what −0.1 means — you cannot infer one from the other): tokens/joule measures the supply-side efficiency of turning electricity into tokens; bytes/task (= bytes/token × tokens/task) measures the customer-side true physical cost per task — the quality of token demand. The physical layer is thus not an appendix to the valuation system; it is the directly tracked object of Table 9's growth requirement. tokens/task has no industry disclosure today, but an observable proxy exists: OpenRouter publishes both token volumes and request counts, so tokens/request can be computed by model and application (request ≠ task — an agent issues several requests per task, so the proxy systematically understates tokens/task but is directionally valid). Until disclosure catches up, bytes/token (Research #16, measured) × tokens/request (publicly measured) is the working basis for bytes/task.

The proxy's first application already has results (OpenRouter network-wide measurements, arXiv:2601.10088, late 2023 to late 2025, ≈ 20 months). Mean sequence length rose from under 2,000 tokens to over 5,400 (prompts ≈ 1.5K to 6K+, a 4× rise; completions ≈ 150 to 400, 2.7×; code requests routinely exceed 20K input tokens). Over the same window the platform's weekly token volume grew by hundreds of times; decomposing in logs, longer requests contributed about 15–20% of token-volume growth, and request-count growth (including user-base expansion) about 80–85% (InAI estimate). Three conclusions follow. First, the verbosity-driven component is real, measurable, and non-dominant in this window: extrapolating revenue from token volume should take a quality haircut in that proportion — neither full credit nor full dismissal. Second, growth's engine is on the input side: prompts (4×) outgrew completions (2.7×), driven by context stuffing and agents' repeated reads — and most input tokens hit cache (OpenRouter's October 2026 publication basis), so billed tokens and physically moved bytes separate further. That is exactly why bytes/task sits closer to true physical cost than token counts. Third, with agents becoming the majority of traffic in 2026 (OpenRouter estimates agent usage passed half from February 2026), tokens/request most likely accelerates upward; the decomposition needs recomputing on the 2026 request series — queued in the quarterly readings.

The Penetration Argument: After Models Converge, Deployment Economics Is All That Differentiates

Convergence is a measured fact, not a judgment. Research #16's measurements across thirty-plus models from ten frontier labs establish two facts side by side: new-generation architectural choices overlap heavily (attention mechanisms, sparsification, and mixture-of-experts configurations converging), and the generational 2.3–48× bytes/token reductions were achieved inside that convergent design space. The financial consequence of convergence is direct: same-capability-tier models substitute for each other, $/token gets pressed toward marginal cost, and the model layer's differentiation premium (ρ's numerator) narrows — same-tier API prices declining generation over generation (Artificial Analysis basis) and open-weight share rising are two faces of the same event.

Once the models themselves converge, the cost difference of the same weights in different operators' hands becomes the only remaining source of differentiation. The proposition has a P&L-shaped proof: a standardized decomposition of a pure token factory (an operator taking open weights and reselling API) shows a revenue midpoint of only about $1.05 per $1 of compute cost — roughly 5 points of gross margin. The efficiency spread of the serving stack (inference engine, batching, caching, quantization) is the entire gross margin: deployment economics is not one input to profit, it is the profit. The two tiers of physical metrics cover exactly this differentiation: bytes/token sets the architecture-given ceiling, and tokens/joule measures the conversion rate the operator actually achieves. When two companies serving identical weights post gross margins several-fold apart with identical line items, these two variables are the only explanatory ones left.

Why the promotion lands in 2027: three conditions are converging. First, necessity is now algebra rather than narrative: the end of Part II decomposes the industry's revenue-growth requirement onto GW × utilization × tokens/joule, with the burden mainly on tokens/joule — tracking the metric is tracking whether the requirement can be met. Second, measurability is filling in: the rising open-weight share makes bytes/token computable directly from public weights, and the tokens/request proxy makes demand quality trackable by quarter; the physical layer graduates from "lab metric" to "quarterly-updatable series." Third, the disclosure window: 2026–27 is the window of concentrated IPOs and refinancings for AI infrastructure and model companies, and disclosure formats are being set right now (Table 12). Analysts who must explain differentiation, metrics that have become computable, and disclosure formats being fixed — all meeting in one window: that is what dates the expectation that physical-layer metrics graduate into valuation metrics.

Finally, the correspondence: revenue per GW is to AI infrastructure what same-store sales is to retail — it splits growth into "opening stores" (new GW) and "store productivity" (revenue per GW), which is how the market learned to refuse companies that only report store counts. The operating drivers of store productivity are the two tiers of physical metrics: tokens/joule is the supply side of store productivity (the electricity-to-token conversion rate), and bytes/task is the cost side of the ticket (true physical cost per task). Same-store sales took one disclosure cycle to evolve from voluntary disclosure into an industry convention, with non-disclosers denied valuation premium; AI infrastructure sits at exactly that point. This paper expects physical-layer metrics to graduate into valuation metrics during 2027 — not by analogy but because the algebraic decomposition of the growth requirement, the completion of measurability, and the disclosure window coincide in time.

The Disclosure Standard: Transplant Semiconductors' Conventions, Complete the Half Already Begun

The second step of the remedy is disclosure. Table 12's audience is AI infrastructure and model companies, and the moment is now: 2026–27 is the window of concentrated IPOs and refinancings (SPCX listed, ANT in registration, more Fabless players and neoclouds behind them). Prospectus and periodic-report disclosure formats are being set at this moment, and conventions formed now will persist for years. Yet not one metric in Table 11 can currently be read off a single company's statements: Table 4's A–D quartet requires splicing both parties' filings, and Table 5's Ld could not be computed while the original S-1 left blanks (the S-1/A completed them — proving the disclosure is feasible and cheap). The application layer already has its first precedent of unit disclosure (Anthropic disclosing the relationship between compute spend and revenue in its S-1 — ρ's numerator and denominator). Table 12 lists the remaining gaps by layer; each disclosure fills exactly one missing input of Table 11:

Table 12: The mirror relationship between disclosure items and the KPI matrix

Disclosure item (by layer)Which Table 11 KPI it directly unlocksToday's workaround (and why it is unreliable)
Infra side: online GW (end-of-period and period average)The denominator of revenue per GW·yrAnnualizing on end-of-period point capacity: how SPCX's single-quarter reading jumped from 6.3 to 16.5
Infra side: GPU·hr sold, effective $/GPU·hr, utilizationThe transfer price (Equation 3) and Equation 1's utilization termEntirely unavailable: Foundry-layer pure-rental revenue can only be estimated on working papers (the $6–12B/GW·yr band)
Infra side: build cost per GW and depreciation per GW·yrImplied payback; adjudicating false claims (Figure 2)Depends on executive-interview bases ($40–50B is attributed statement, not disclosure)
Both sides: commitments, backlog, revenue recognized (the A–D quartet)The backlog coverage ratio (the 0.47 and 12 boundaries) and the three adjacent ratios; once in place, lights up Table 11a's Entity-3 backlog cell and the two "to be spliced" bill-ratio cellsEach side discloses some quantities; one relationship's A–D takes two companies' filings to splice (Table 4 is such a splice)
Both sides: deal-by-deal tagging of related financing (supplier-as-creditor, chipmaker-as-shareholder, backstop purchases)Vendor financing ratio Lu and supplier capital share Ld; once in place, lights up Table 11a's Entity-1 Ld cellLu could only be estimated at cap basis and Ld was uncomputable from blanks; the ANT S-1/A has demonstrated deal-by-deal disclosure (three suppliers' facilities, commitments, drawn status side by side) — pending adoption as convention
App side: tokens served and realized $/M tokensEquation 1's $/token term; third-party verification of the closure testOnly market list prices (Artificial Analysis) to approximate realized prices
App side: compute purchases bucketed by tenor (spot / short / long) and ρEquation 2's cost structure; continued observation of ρThe S-1 itself admits it "may pay premium prices for already-depreciated resources," but without bucket disclosure the risk cannot be audited
App side: tokens / GPU·hr (with hardware and model specs)The physical bridge; company-level measured tokens/jouleOnly public benchmark ranges (TokenPowerBench family) as substitutes — can differ from company reality by an order of magnitude

The precedents and the returns of disclosure are both on the record. On the semiconductor side, TSMC discloses revenue monthly and capex and capacity quarterly — which is why the market can stably price a company spending $40.9B a year on capex. On the retail side, same-store sales began as voluntary disclosure and ended as analysts refusing valuation premium to non-disclosers. For companies, disclosure buys a computability premium: estimate dispersion converges, financing costs fall. The cost of not disclosing is equally measurable: sell-side estimates of one reference asset's next-year revenue span $59.4B to $290.1B — a near-5× dispersion ultimately borne by shareholders as volatility and discount. For investors and regulators, disclosure dismantles the volatility mechanisms of the opening section: once the base is directly observable, one company's news no longer needs the entire industry repriced. Whoever discloses first collects the computability premium first.

When online GW, revenue per GW, GPU·hr volume and price, tenor-bucketed compute purchases, and deal-by-deal related financing enter periodic reports, every variable in the three equations becomes a direct line-item read, and the pricing of the AI trade returns from the narratives of two or three companies to the unit economics of the whole industry. That is what this paper's title means. Until disclosure arrives, InAI will publish spliced-and-computed tracking readings quarterly, on Table 11's basis and from public filings, as the transition period's public reference.

The readings are now being published: The Gigawatt P&L Tracker gives each Table 11 metric with its source, quarter by quarter; the AI Valuation KPI Dashboard grades them automatically against the alert thresholds in this paper and lets any reading be changed for a scenario test.

The 2027 Checklist

Every claim in this paper can be tested against data in 2027; readers can check quarterly:

  • Does SPCX's next-quarter SI revenue per GW hold the $16.5B basis; are folding-boundary (consolidation) changes disclosed separately;
  • ANT's post-IPO ρ trajectory (the S-1/A gives the starting points: H1 1.82, Q3 1.72), and whether the disclosure continues;
  • Does ANT's commitment conversion (C ÷ B) rise quarter by quarter; do allocations across suppliers stay near-even and cancellation rights stay in place — the direct test of commitments-as-demand versus commitments-as-narrative (the Table 4 hypothesis);
  • Does the dispersion of sell-side revenue estimates for AI infrastructure companies converge (the current reference-asset range: $59.4–290.1B);
  • Does a second company adopt per-GW or tenor-bucketed compute-purchase disclosure;
  • The SemiAnalysis–Gerstner disagreement on 2027 GW additions (51.2 vs ≈ 25) gets adjudicated by actual deliveries; does the bottleneck rotate back to chips and HBM as predicted;
  • Does industry external revenue grow at no less than 30% annualized (the endogenous floor); does evidence appear of the economic life multiple leaving 1 (used-GPU prices, utilization of older generations rotated to inference, long-term software support);
  • Does data-center credit-spread sensitivity to single-company news decline; do physical-layer metrics start entering sell-side models.

If most items hold, the AI trade migrates from narrative pricing to calibrated pricing — the very process this paper hopes to accelerate.

Sources

  • Filings, as filed: SpaceX (SPCX) 2026 Q3 / Q2 8-K EX-99 consolidated statements, Q3 10-Q, SEC XBRL; Anthropic, PBC Form S-1 (SEC EDGAR, accession 0001628280-26-067168, 2026-10-26; commitment amounts blank in the public draft) and Form S-1/A (accession 0001628280-26-068676, 2026-11-02; all amounts completed; this paper's ANT supplier decomposition, ρ, cash flow, and fair-value figures follow that amendment as filed, verified line by line against the original HTML); S-1/A Amendment No. 3 (accession 0001628280-26-074338, 2026-11-17: total commitments of $900.9B and their time profile, the run-rate definition, and the purchasers and amounts of the concurrent placement); the pricing announcement (2026-11-18: $83 a share, 980 million shares). The final prospectus (424B4, filed 2026-11-19) gives 13,219,642,276 shares outstanding after the offering, including the concurrent placement shares and excluding the over-allotment option (as cited in attributed reporting); first-day prices are Nasdaq closing data.
  • Media bases: private credit and issuance forecasts (Bloomberg 2026-02, citing Morgan Stanley, JPMorgan, IJGlobal); BIS Bulletin No. 120 (2026-01); hyperscaler capex from each company's mid-2026 earnings-call guidance.
  • Capacity and energy benchmarks: SemiAnalysis data-center forecasts (2026-10, Data Center Frontier reporting basis); SemiAnalysis SPCX build timelines and GPU rental index (2026-10 to 2026-11 reporting: clusters live in 91–150 days, industry 8–12 months; blended H100 rents $6.62 in H2 2023 to $2.80 in 2026-03); TokenPowerBench (arXiv:2512.03024); John Snow Labs tokens-per-joule benchmarks (2025–26).
  • Demand-side series: OpenRouter official data publications and rankings (weekly token volumes, request counts, cache basis, 2025–26); State of AI: An Empirical 100 Trillion Token Study with OpenRouter (arXiv:2601.10088): sequence-length trajectory (≈ 2,000 to 5,400+ tokens), prompt/completion decomposition, code-request basis; the tokens/request log decomposition is an InAI estimate.
  • Interviews and public statements (attributed): Jensen Huang, Brad Gerstner, Clark Tang, BG2 podcast (2026-10-19): the three-way demand mix, the ≈ $200B frontier-labs-plus-inference revenue basis, sellable content per GW, deployment constraints, Gerstner's challenge to the 2027 additions figure; Musk's "six-month payback" remark (SPCX Q3 earnings communication; adjudicated as failing in Figure 2).
  • Loop and precedent data: NVIDIA–OpenAI (the September 2025 letter of intent for up to $100B never became a contract; a $30B equity stake and ≈ 5 GW of NVIDIA systems in the round announced 2026-02-27 and closed 2026-03-31, per OpenAI's announcement and attributed reporting; the PORTS-Pike guarantee per NVIDIA's 2026-08-17 release and its 2026-11-17 earnings call: first 4.25 GW, guarantee cap $105B, total guarantees $118B, frontier labs about 15% of data center revenue), CoreWeave ($6.3B backstop purchase, filed by 8-K; IPO anchor ≈ $250M), Nebius ($2B), xAI (≈ $2B), and customer concentration (top two at 39%): compiled from company filings and attributed reporting; "≈ $10B of investment tied to ≈ $35B of revenue" is NewStreet Research's single-party estimate; CoreWeave FY2025 (revenue $5.1B, net loss $1.2B, backlog/RPO $60.7B) is its filing; the telecom precedent (Lucent's $8.1B committed and $3.5B lost, Nortel's 130%, nine vendors' $25.6B, four dozen CLEC bankruptcies) is an industry-history compilation from Lucent/Nortel annual reports, SEC files, and McKinsey estimates, each item sourced to primary documents before use.
  • Semiconductor benchmark: TSMC 4Q25 management report (2025 revenue $122.4B, capex $40.9B, D&A NT$688.1B) and annual-report depreciation policy (equipment 5 years, buildings 20); TSMC's ≈ 70% pure-play foundry share (TrendForce, 2025); Counterpoint's wider "Foundry 2.0" 2025 industry total of $320B (TSMC at 38%, wider basis, reference only); global fabless revenue totals and NVIDIA's share are industry compilations, to be re-verified before final; the share extrapolations, economic life multiple, and endogenous CAGR computations are InAI analysis.
  • InAI's own research and estimates (all labeled as estimates): Research #15, Research #16; the three-scenario model (starting from SPCX Q3 actuals, script reproducible); the U.S. end-2027 35–50 GW estimate; the industry P&L's two-basis coverage, the Fabless cost equation, and the identity closure test.