Naive deployment-weighted 'world prices' for data-center SKUs can invert cost rankings and mislead procurement; two cost-preserving operators—a normalized two-way fixed-effects decomposition and a convex common-weight benchmark—cut ranking violations markedly while preserving total costs.

Main Finding

Blending campus-level unit prices into system-level "world prices" for identical SKUs in multi-campus AI data centers should preserve total cost but also be robust to Simpson-type ranking reversals driven by SKU-specific location mixes. The paper proposes two practical, distributed-implementable operators that (1) exactly preserve portfolio cost and (2) substantially reduce ordering violations compared with naive deployment-weighted blending: (i) a two-way fixed-effect (FE) estimator with a scalar cost-normalization, and (ii) a constrained convex common-weight aggregator solved via a small quadratic program (with an active-set projection for nonnegativity). Together these form a detect → correct → validate workflow that is inexpensive to run in production pipelines.

Key Points

• Problem and risks

  • Firms need a single system-level price per SKU for budgeting, chargeback, procurement and benchmarking.
  • Naive deployment-weighted blends preserve total cost but can produce Simpson-type reversals (ranking flips) because each SKU uses its own location mix as weights.
  • Such reversals distort procurement trade-offs, capacity planning, and fairness in chargeback systems.

• Desirable properties for a blending operator

  • P1: Exact cost preservation (Σ p̄_i Q_i = realized total cost C).
  • P2: Location-robustness in pairwise comparisons (minimize/eliminate dominance reversals).
  • P3: Implementability in distributed systems using low-dimensional aggregates.

• Proposed operators

• Two-way fixed-effect (FE) operator - Model: p_ij = α_i + γ_j + ε_ij, estimated by quantity-weighted least squares (q_ij as weights) with the constraint Σ γ_j = 0. - Produces product fixed effects α̂_i as pre-normalized world prices; then apply a single scalar shift δ to ensure exact cost preservation: p̄_i(FE) = α̂_i + δ. - Properties: exact cost preservation after normalization; preserves dominance in near-additive/balanced settings; interpretable campus premia γ̂_j; straightforward to implement (group-by + regression). - Failure modes: sparse SKU×campus coverage, strong SKU×campus interactions or weak identification can increase ordering violations.

• Constrained convex common-weight operator (W) - Idea: use a single nonnegative campus weight vector w (same for all SKUs) so blended prices are p̄_i(W) = Σ_j w_j p_ij — this prevents Simpson reversals when local dominance holds. - Choose w as the projection of a baseline weight vector w̃ (e.g., global quantity shares) onto the simplex intersected with the cost-preservation hyperplane Σ_j w_j A_j = C, where A_j = Σ_i Q_i p_ij. - Solve min ||w − w̃||^2 subject to w ≥ 0, Σ w_j = 1, and cost constraint. Closed-form projection available ignoring nonnegativity; enforce nonnegativity with a small active-set loop. - Properties: exact cost preservation, dominance-robust ordering under nonnegativity, cheap post-processing after distributed aggregation; numerical complexity trivial for typical J (tens).

• Diagnostics and workflow

  • Introduces diagnostic metrics: Ordering Violation Rate (OVR) to quantify pairwise ranking reversals, and Cost Distortion Ratio (CDR) to quantify extent of cost misallocation from ranking changes.
  • Recommends a detect → correct → validate pipeline: detect naive reversal risk (OVR/CDR) → correct using FE or W operator → validate reduced OVR/CDR and exact cost preservation.

• Empirical evaluation

  • Simulations and a realistic AI data-center OPEX case study show substantial reductions in ranking violations versus naive blending, while preserving total cost to machine precision.
  • Both methods integrate easily into distributed frameworks (e.g., SQL / PySpark) via a few GROUP BY aggregations and a small driver-side computation.

Data & Methods

• Data setup

  • Observations: unit prices p_ij ≥ 0 and deployed quantities q_ij ≥ 0 for product i at campus j.
  • Derived aggregates: product totals Q_i = Σ_j q_ij, campus exposures A_j = Σ_i Q_i p_ij, total realized cost C = Σ_i Σ_j p_ij q_ij.

• Method details

• FE method - Quantity-weighted least squares estimate of α_i (product effect) and γ_j (campus effect) with Σ γ_j = 0. - Scalar normalization δ chosen to satisfy Σ_i (α̂_i + δ) Q_i = C. - Implementation: GROUP BY (product,campus) to build panel → weighted regression (q_ij weights) → compute δ from aggregated α̂_i and Q_i → store p̄_i(FE).

• Constrained common-weight method - Baseline w̃ (e.g., global quantity shares) and exposure vector s (s_j = A_j). - Solve QP: minimize ||w − w̃||^2 subject to w ≥ 0, Σ w_j = 1, and s^T w = C. - Closed-form Euclidean projection exists when nonnegativity ignored: w = w̃ − A^T (A A^T)^{-1} (A w̃ − b) with A = [1^T; s^T], b = [1; C]. - If negatives appear, use active-set: fix negative entries to zero and re-project on remaining indices until all w_j ≥ 0. - Implementation: compute campus-level aggregates in distributed manner, solve small QP on driver, broadcast w and compute p̄_i(W) by weighted sums.

• Diagnostics

  • OVR: fraction (or count) of product pairs where local dominance is reversed in blended prices.
  • CDR: a measure of cost distortion associated with ranking violations (paper defines formally; used to detect severity).

• Complexity and production considerations

  • Heavy aggregation step is linear in observed (i,j) pairs; post-processing is O(J) or at worst O(J^2) in active-set; negligible when J is small.
  • Integrates with standard data engineering (GROUP BY, broadcast small tables, join + group-by to compute p̄_i).

Implications for AI Economics

• Robust technoeconomic benchmarking

  • World prices that are both cost-preserving and location-robust improve the reliability of SKU-level comparisons used in procurement, vendor negotiations, and benchmarking across regions.
  • Reduced Simpson-type reversals leads to clearer decision signals for which SKUs are truly cheaper/higher cost.

• Better budgeting, chargeback and fairness

  • Exact cost preservation ensures reconciliability with financial reporting and OPEX systems (critical when energy/cooling vary across campuses).
  • Location-robust blends reduce perverse incentives and perceived unfairness in internal chargeback allocations caused by aggregation artifacts.

• Improved capacity planning and workload allocation

  • More stable SKU rankings support better prioritization for capacity upgrades, workload reallocation, and energy-cost-aware placement decisions.

• Practical adoption and governance

  • The methods are lightweight to implement in existing distributed data stacks; they can be run regularly to keep benchmarks synchronized with evolving deployments and prices.
  • The detect→correct→validate workflow (using OVR/CDR diagnostics) supports governance: surface when naive blends are risky, apply correction, and validate that accounting identities hold.

• Limitations and caveats

  • FE operator performs best when SKU coverage is reasonably balanced and campus premia are approximately parallel across SKUs; strong SKU×campus interactions or sparse coverage can erode its robustness.
  • The common-weight operator requires a feasible projection; if the cost constraint is tight or baseline weights are far from feasible, the active-set may force substantial weight shifts (the paper includes diagnostics and slack-based fallback).
  • Neither operator can globally eliminate all possible paradoxes for arbitrary data matrices, but both materially reduce practical reversal risk while preserving cost.

Recommendation (practical): instrument your cost pipeline to compute OVR/CDR for current naive blends, and—when risk is material—apply the FE and/or constrained-common-weight operator (both easy to implement in SQL/PySpark) to produce cost-preserving, more interpretable world prices for budgeting, chargeback, and procurement analyses.