V-Score Card

TICKER:             EPAM
V-SCORE:            0.00 (Raw: 1.38, zeroed by Gate 1)
VERDICT:            COLLAPSED
κ (conviction):     0.00
GATE 1 (E>1):       FAIL — E = 1. No irreducible infrastructure.
GATE 2 (A>1∨Σ≥12): PASS (A=2>1, moot)
FAST SCREEN:        0/3 — No proprietary data, no regulatory mandate, not embedded in transactions.

Dimension Scores

Arithmetic

  0.25 × 2 = 0.50   (C: Cognition)
  0.22 × 1 = 0.22   (E: Infrastructure)
  0.18 × 2 = 0.36   (U: Breadth)
  0.12 × 2 = 0.24   (A: Distribution)
  0.15 × 2 = 0.30   (M: Gravity)
 −0.06 × 4 = −0.24  (F: Friction penalty)
            ──────
  Raw       = 1.38

  G₁ = 𝟙[E > 1] = 𝟙[1 > 1] = 0   ← BINDING
  G₂ = 𝟙[A > 1 ∨ C+E+U ≥ 12] = 𝟙[2 > 1 ∨ 5 ≥ 12] = 1

  V = 1.38 × 0 × 1 = 0.00

Stress Test: Maximum-Bull Scenario

Even granting E=2 (some switching friction is "infrastructure") and C=3 (30 years of organizational capital is "superlinear"):

  Raw = 0.75 + 0.44 + 0.36 + 0.24 + 0.30 − 0.24 = 1.85
  V = 1.85 → DEAD ZONE. κ = 0.

The gap between best-case (1.85) and survival threshold (3.0) is 1.15 points. Structurally unrescuable.

Dimension Analysis

E = 1: The Kill Dimension

The V-Score framework asks a precise question: does there exist any task τ in EPAM's domain where the cost of doing it locally with AI is infinite? The answer is no, and EPAM's management wrote the proof into their own SEC filing.

The entire task domain — coding, testing, cloud migration, data engineering, consulting — is computable. For every τ ∈ 𝒟(s): c_ℓ(τ,t) = g(τ)/M(t) → 0 as model capability M(t) → ∞. Since c_s(τ) ≥ c_s_min > 0 (irreducible cost of a human engineer), there exists t*(τ) such that c_ℓ < c_s for all t > t*(τ). This is the Tool Death Theorem, and it applies a fortiori to services companies — a SaaS company at least has crystallized cognition in its product. A services company IS the routing layer that collapses.

The financial evidence shows displacement is already materializing. Operating margins compressed 200bps (11.5% → 9.5%) despite 15% reported revenue growth. Utilization declined 270bps intra-year (78.1% → 75.4%). Three consecutive years of restructuring programs totaling ≈$100M+ in severance. Organic growth decelerating toward stall speed (7.1% → 5.6% → 3%). Rate increases of 2-3% against 30-50% AI productivity gains imply revenue-per-project is compressing even as bill rates hold.

The bull counterarguments were tested and fail:

Jevons Paradox ("cheaper code = more EPAM demand"): Historically rare for labor services. ATMs didn't increase teller demand. Desktop publishing didn't increase typesetter demand. The incremental demand is met by the cheaper tool, not through the old labor model.

Enterprise complexity ("AI makes coding easy but integration is hard"): True today, false in the limit. V-Score estimates lim as t → ∞. And even today, margins are compressing and utilization is declining — complexity isn't generating pricing power.

DIAL platform ("EPAM becomes agent orchestration layer"): Open-source, zero revenue, anyone can fork it. Building tools for the ecosystem is not capturing the ecosystem.

AI-native revenue $600M ("EPAM is riding the wave"): 11% of 2026 guided revenue. Still T&M billing. Consulting "how to adopt AI" is itself automatable.

C = 2: Moderate but Linear

Thirty years of enterprise delivery across six verticals is genuine accumulated knowledge. The AIRON agent scoring top-5 on SWE-Bench proves real engineering capability. But the compounding test fails: ∂²c_derive/∂n² ≈ 0. Projects are "singular in nature" (10-K L1516). Banking methodology doesn't depend on healthcare methodology. The accumulation is linear across independent verticals, held in people with 8.5% attrition, building code that clients own.

A frontier model re-derives EPAM's delivery methodology within 12-18 months. The specific client knowledge (how Bank X's systems work) is M-dimension gravity, not C-dimension cognition. It's specific to the relationship, not to EPAM's compound IP.

A = 2: Agents Route Around, Not Through

DIAL 3.0 with MCP support is the strongest A-dimension asset — technically agent-accessible. But it generates zero revenue and is open-source. When an AI agent needs code written, it routes to Copilot, Cursor, Claude Code — not to EPAM. The engagement model requires human sales, scoping, SOW negotiation. Agents can't programmatically discover, evaluate, or procure EPAM's services.

The Cursor partnership (Jan 2026) makes EPAM a customer of AI tools, not a platform agents route through. Management positions agentic AI as competitor, not distribution channel.

Regime Context

15-Week Factor Regression (2025-12-31 to 2026-04-13)

IT services is in a high-correlation selloff. The AI displacement narrative is moving the entire sector as a single factor, crushing idiosyncratic signal.

ρ_intra = 0.616 (HIGH-CORRELATION SELLOFF)
SPY: +0.1%  |  IT Services: −27.2%  |  Spread: −27.4%

Ticker   α_ann    σ_idio    IR     %Idio    γ      CumRet    t(α)
──────────────────────────────────────────────────────────────────
EPAM    −39.8%    34.7%   −1.15    35.2%   1.26    −39.4%   −0.59
CTSH    −11.7%    16.8%   −0.70    17.8%   0.91    −27.7%   −0.36
WIT     −27.6%    41.3%   −0.67    80.4%   0.45    −21.7%   −0.34
EXLS    −15.5%    30.9%   −0.50    41.3%   0.94    −29.6%   −0.26
G       −12.0%    24.9%   −0.48    39.1%   0.74    −24.3%   −0.25
INFY     +2.8%    29.5%   +0.10    44.5%   0.78    −22.6%   +0.05
GLOB    +31.5%    35.1%   +0.90    30.4%   1.32    −29.9%   +0.46

All t(α) values in [-0.59, +0.46] — no name has statistically significant alpha over this window. IR measures the regime, not the stock. When ρ_intra → 1, ε_i → 0 for all names. The entire IT services sector is being repriced by a single thematic factor, and stock-specific signal is unmeasurable.

EPAM Return Decomposition

Total:             −39.4%
  Sector (γ × IT): −31.7%   (80% of decline)
  Market (β × SPY): −0.0%
  Idiosyncratic:    −7.7%   (t = −0.59, not significant)

EPAM has the highest sector loading (γ = 1.26) of any name in the peer set. It's the leveraged expression of the AI-displacement-of-IT-services factor — 80% T&M, no product revenue, clients own all IP, management confirms substitution risk. The market treats it as the purest play on the theme, and the factor regression confirms this.

δ: V-Score vs Market-Implied

V(s) is orthogonal to r_sector(t). The V-Score is scored against structural properties — contract terms, task computability, IP ownership, regulatory mandates — not price. A 50% sector bounce wouldn't change V from 0.00. A sector crash wouldn't make V worse. The structural verdict is regime-invariant.

The delta tells the story:

δ_EPAM = V_structural − V_market_implied
       = 0.00 − ≈1.2
       = −1.2

V_market_implied ≈ 1.0−1.5 (from 9.2× forward P/E vs historical 20−30×;
  market prices structural impairment but not collapse — still assumes
  positive terminal value on stable earnings)

δ < 0: the market is still too generous. A forward P/E of 9.2× assumes the E is stable. But organic growth is decelerating from 7.1% to 3%, GAAP margins compressed 200bps despite revenue growth, and three consecutive restructuring programs signal demand erosion, not efficiency optimization. The market is pricing a melting ice cube at a discount to its current mass, not to its terminal mass.

Nineteen analysts carry median $193 target (+51% upside). This is analyst lag — pricing the EPAM of 2024, not the one whose 10-K reads like a V-Score rubric for E = 1. Citigroup's April downgrade to $145 with Neutral is the closest to structural reality.

Thermodynamic Summary

Nothing prevents intelligence from flowing around EPAM.

The company sells hours of human software engineering — T&M billing at hourly rates, contracts terminable at will, clients own all IP. This is the canonical Tool Death Theorem case: the entire task domain 𝒟(s) consists of computable tasks, therefore lim R(s,t) = 0.

The kill cycle is already running:

Durable revenue is ≈5% — complex enterprise architecture requiring human judgment plus liability acceptance. This is not separately reported and is insufficient to sustain a 56,600-person delivery model. The 95% exposed revenue is computable and facing active or imminent displacement.

Financial evidence of active displacement: operating margins 11.5% → 9.5% (-200bps), utilization 78.1% → 75.4% (-270bps intra-year), organic growth 7.1% → 5.6% → 3% (decelerating toward stall), three consecutive restructuring programs ($100M+ cumulative severance), $660M in buybacks at average $186.67 (now 31% underwater at $128).

Applicability note: EPAM is IT services, not SaaS. The V-Score was designed for software companies, but Intelligence Conservation applies a fortiori to services companies. A SaaS company at least has a product with crystallized cognition. A services company IS the routing layer — the exact thing that collapses as c_ℓ(τ,t) → 0. EPAM represents the strongest possible form of the Tool Death Theorem.

Conviction & Basket Verdict

V-SCORE:    0.00
TIER:       COLLAPSED
κ:          0.00   (max(0.00 − 3.0, 0) = 0)
w_EPAM:     0      (excluded from basket)
IR_EPAM:   −1.15   (not significant; t = −0.59; regime artifact)
δ_EPAM:    −1.2    (market still too generous)

EPAM is excluded from the SaaS survival basket. κ = 0 means zero weight allocation under the proportional rule w_i ∝ κ_i. The V-Score verdict is COLLAPSED — no irreducible infrastructure, all tasks computable, management confirms AI substitution in their own SEC filings. The regime analysis confirms the market is directionally correct but potentially insufficient in magnitude. IR is unmeasurable in the current high-correlation selloff and does not gate the structural verdict.