S&P Global is a regulatory franchise the market is pricing as an AI casualty. Five segments — Ratings (NRSRO designation, three-sided network, bond indentures), Indices ($5.48T AUM, contractual), Market Intelligence (data feeds + desktop), Commodity Insights (Platts benchmarks), and Mobility (spinning Q2 2026). The AI disruption narrative applies to ≈3% of segment profit (MI desktop layer). The other 97% is moat-protected by regulation, contract, or network effect.

V-Score Card

Arithmetic

V = (0.25 x 5) + (0.22 x 5) + (0.18 x 5) + (0.12 x 3) + (0.15 x 5) - (0.06 x 2)
  =  1.25 + 1.10 + 0.90 + 0.36 + 0.75 - 0.12
  =  4.24

Gates

G1 = 1[E > 1] = 1[5 > 1] = 1    PASS
G2 = 1[A > 1 OR C+E+U >= 12] = 1[3 > 1 OR 15 >= 12] = 1    PASS (both conditions)

V = 4.24 x 1 x 1 = 4.24    FORTRESS

Adversarial Stress Test (B0)

The C=5 and E=5 scores were challenged. Key arguments:

C=5 challenge: Ratings methodology is published (Dodd-Frank transparency). Moody's and Fitch independently derived equivalent frameworks — the cognition is deep but not irreproducible. Index methodology is rules-based and trivially derivable. The score conflates data accumulation (E) and market trust (M) with cognitive complexity (C).

E=5 challenge: c_l = infinity holds for existing bond indentures and commodity contracts, but for new issuance and regulatory capital the switching cost is high but finite. MI data infrastructure competes with Bloomberg and FactSet — stickiness, not irreducibility.

Assessment: C=4 is the strongest challenge. The cognition has been independently derived by competitors, and frontier models will approximate the judgment layer within 2 years. But "approximate" is not "replace" — the regulatory stamp is the moat, not the methodology. E=5 holds at company level: Ratings + Indices (65% of segment profit) dominate the blended score. The honest range is V = 3.99-4.24, straddling the FORTRESS/EMBEDDED boundary.

Published score: V = 4.24, acknowledging B0 range [3.99, 4.24].

Dimension Analysis

C=5: Crystallized Cognition — The Methodology Stack

SPGI's cognition is layered across segments. Ratings carries 150 years of credit methodology — default studies, recovery analysis, sector-specific criteria, qualitative management assessments. This isn't a model you retrain; it's institutional knowledge encoded in thousands of analyst decisions, calibrated against actual defaults across multiple credit cycles.

The inter-module dependencies are genuinely superlinear. Ratings data feeds MI credit analytics. MI entity resolution feeds Indices construction. Platts commodity pricing informs Ratings on energy sector credits. $200M in intersegment revenue flows are the observable trace of deeper analytical dependencies. Replacing one module orphans data pipelines in adjacent segments.

B0 concession: Moody's and Fitch independently derived equivalent credit methodologies. The cognition is deep but not unique — three firms hold it. The moat is the oligopoly structure (network effect, M=5) plus the regulatory designation (E=5), not the methodology alone. Index methodology (S&P 500 construction) is rules-based and trivially derivable. AI could approximate the ratings judgment layer within 1-2 years. But approximation doesn't matter — the bond indenture says "S&P rating," not "equivalent credit assessment."

E=5: Ecosystem Entrenchment — The Regulatory Fortress

NRSRO designation is the keystone. Ten registered globally. SEC-supervised. Bond indentures contractually specify S&P rating triggers — changing the reference agency requires bondholder consent or new issuance. Basel III risk-weighting requires NRSRO ratings for bank capital calculations. Solvency II and RBC reference agency ratings for insurance capital. The three-sided network (issuer pays for rating, investor requires rating, regulator mandates rating) creates a self-reinforcing loop where each participant's switching cost is amplified by the other two.

The Dodd-Frank natural experiment (2010, Section 939A) proves the moat. Congress directed federal agencies to remove NRSRO references from regulations. Sixteen years later: Ratings revenue $2B to $4.7B. The law changed. The market didn't. Bond indentures are private contracts — Congress can't rewrite them.

Indices: $5.48T tracks SPDJI indices. ETF prospectuses reference specific indices. Changing index requires prospectus amendment, SEC filing, investor communications. The CME put on the SPDJI JV implies $24.5B+ standalone valuation — the market values the contractual lock-in at 10x+ Indices segment revenue.

B0 concession: MI data infrastructure (Capital IQ, Xpressfeed) competes with Bloomberg Terminal, Refinitiv, FactSet. This is enterprise data stickiness (E=3-4), not regulatory entrenchment (E=5). Blended across segments, the case for E=4 exists if you weight MI infrastructure equally with Ratings/Indices lock-in. We don't — Ratings + Indices represent 65% of segment profit.

U=5: Usage Breadth — The Workflow Octopus

SPGI touches every department of a financial institution. Trading desk uses Platts for commodity pricing. Risk uses Ratings for capital calculations. Compliance uses regulatory feeds. Portfolio managers use Indices benchmarks. Research uses Capital IQ and Credit Analytics. Treasury uses market data. IR uses reference data. Legal uses bond indenture information. Insurance uses Ratings for RBC. Commodities uses Platts assessments.

35+ products across 5 segments. Cross-departmental data dependencies mean the switching cost is superlinear — replacing Capital IQ for research doesn't help if Risk still needs RatingsXpress and Compliance still needs regulatory feeds, all of which share entity resolution and reference data from the same platform.

A=3: Agent Accessibility — The Gap

SPGI has the data. Every major AI platform licenses it. But SPGI hasn't built the agent-native interface layer. AlphaSense is shipping GraphQL APIs, developer portals, custom AI agents, and due diligence workspaces. SPGI's CEO called it "early days" on the Q4 2025 call.

This is the one dimension where SPGI is genuinely behind. The risk: AlphaSense or a similar platform becomes the agent-native interface to SPGI's own data, capturing the interaction layer and commoditizing the data substrate. The mitigation: SPGI's data licensing model means they get paid regardless of interface — but the margin and customer relationship shift to the interface owner.

A=3 is correct. Not agent-default (A=5), not agent-absent (A=1). Functional APIs exist, third-party integration works, but no SPGI-built agent layer.

M=5: Market Power — The Numbers

$15.3B revenue. 42% operating margin. $5.1B FCF. $14.66 diluted EPS (+19% YoY). 83% subscription/recurring revenue.

The market power metrics are extraordinary:

• Ratings: Big Three hold 97-98% share after 16 years of deregulation attempts
• Indices: $5.48T AUM is contractual and fee-generating (basis point fees on AUM)
• Platts: Benchmark pricing referenced in physical commodity contracts worldwide
• MI: 6.5-7% subscription ACV growth, accelerating

$5.0B in buybacks in 2025 (+52% YoY), with management front-loading $1B in Q1 2026 citing "stock price levels." New 30M share authorization. OSTTRA sold to KKR for $3.1B. Six acquisitions deepening data/AI moat (With Intelligence $1.8B, ProntoNLP, Visible Alpha).

F=2: Friction — Low and Managed

SPGI actively minimizes friction. "Flexible distribution policy" — clients can use data wherever they prefer, including in LLMs. Clean APIs. 83% subscription model. No consultant dependency for implementation. Active app rationalization (10%+ eliminated, SSO enabled).

Enterprise sales cycle prevents F=1. Complex procurement for $15.3B enterprise contracts is inherent, not fixable. But within enterprise constraints, SPGI is among the lowest-friction data providers.

Thermodynamic Summary

SPGI is a low-entropy system. The structural moats are thermodynamically stable — they resist perturbation because the energy required to displace them exceeds the energy available to any attacker.

Energy barriers by segment:

97% of revenue sits behind energy barriers that exceed the available activation energy of AI disruption. The MI desktop layer (≈$2.0-2.6B, 13-17% of revenue, ≈5-6% of segment profit) is the only surface area where AI can apply sufficient energy to displace SPGI. And even there, the data feed infrastructure (as opposed to the analytics layer) has its own switching costs.

Phase transition risk: The only path to structural degradation requires simultaneous erosion of the regulatory moat (NRSRO unwinding) AND the cognitive moat (frontier models replicating 150 years of cross-domain credit methodology) AND the network moat (bond indenture migration). These are independent events with low individual probability. The joint probability is negligible.

Regime Context: IR and delta

The 15-Week Regime (2025-12-15 to 2026-03-27)

The measurement window is dominated by a single latent factor: the "AI disruption" selloff across data/analytics names.

PC1 loadings (the AI disruption factor, all negative):

The market loads SPGI at 83% of FDS's AI-disruption sensitivity (0.435/0.526). Structurally, SPGI's AI-exposed surface area is ≈3% of segment profit versus FDS's ≈60%+. The market is mispricing the loading by roughly 25x.

IR (regime-contaminated, NOT alpha-diagnostic)

IR_SPGI (2-factor: SPY+XLF)      = -0.59
IR_SPGI (3-factor: SPY+XLF+PC1)  = -1.45

SE(alpha_ann) ~ 26.7%  -->  95% CI includes zero
71 trading days cannot distinguish signal from noise

IR measures the regime, not the stock. When rho_intra = 0.963, the "idiosyncratic" residual is not idiosyncratic — it contains the latent AI-disruption factor that hits all names indiscriminately. SPGI's negative IR reflects its loading on a narrative selloff, not a deficiency in the business.

delta: The Structural Discount

V_SPGI          = 4.24  (scored from C, E, U, A, M, F -- no price inputs)
V_market,SPGI   ~ 3.0   (implied by PC1 loading: market prices SPGI
                          as generic data/analytics, not regulatory franchise)

delta = V_SPGI - V_market = 4.24 - 3.0 = 1.24
B0 range: delta in [0.77, 1.24]

The delta is the edge. The market applies a uniform "AI disruption" discount to data/analytics names. V-Score, computed from structural properties orthogonal to price, says SPGI's moat is intact. The gap between structural reality and market perception is the trade.

Conviction Weight

kappa = (V - 3.0)+ = (4.24 - 3.0)+ = 1.24
B0 range: kappa in [0.77, 1.24]

kappa is regime-invariant. It does not depend on the 15-week return, the IR, the rho_intra, or the latent factor. It depends on C, E, U, A, M, F — structural properties that update on evidence (8-Ks, 10-Qs, regulatory changes), not price.

Sizing within the basket:

w_SPGI = W_S x kappa_SPGI / sum_j(kappa_j)

Where W_S is the basket-level allocation and sum_j(kappa_j) runs across all scored names.

Basket Verdict

V = 4.24 | FORTRESS | kappa = 1.24 | KEEP

SPGI sits at the structural ceiling of the V-Score framework. The only dimension below 5 is Agent Accessibility (A=3), which represents a distribution gap — not a survival risk. The adversarial stress test (B0) finds that even under the strongest challenges to C and E, V remains at 3.77-4.24. Only simultaneous degradation of the regulatory moat AND the cognitive moat AND the market power network drops SPGI below EMBEDDED — and the base rate for that triple failure is effectively zero.

The regime is doing the selling. Not the fundamentals. When rho_intra = 0.963, the market cannot see SPGI. It sees "data/analytics company, AI-exposed, sell." IR = -0.59 in this window reflects the latent AI-disruption factor, not structural deterioration. V-Score, scored from NRSRO filings and 10-K structural analysis, is invariant to the regime.

The edge is the delta: the market applies a uniform discount where the structural moat is non-uniform. SPGI's regulatory franchise (Ratings + Indices = 43% revenue, 65% segment profit) is structurally closer to ICE (exchange clearing infrastructure, PC1 loading -0.181) than to FDS (desktop analytics, PC1 loading -0.526). The market loads SPGI at -0.435. That gap is the trade.