Definition
Arbitrage Concentration Risk (ACR) is the proposed metric for this: the degree to which AI-driven cost optimization has concentrated workloads in a single cloud provider, region, or availability zone, measured as the Herfindahl-Hirschman Index (HHI) of spend distribution across provider-region combinations, tracked over time. ACR borrows the HHI from antitrust economics, where it measures market concentration. The same mathematical property that makes HHI useful for detecting market concentration (it disproportionately weights the largest players) makes it useful for detecting workload concentration. A single provider-region combination at 80% of your compute spend contributes 0.64 to the index. The remaining 20% spread across five combinations contributes at most 0.04 total. High concentration registers unmistakably in a way that averaged metrics never would.
The Measurement Gap
The cost optimization results are good. Your multi-cloud placement engine (running on Spot.io, CAST AI, or your cloud provider's native recommendation service) has moved workloads steadily toward cheaper compute over the past two years. Spot instances where the risk profile allows. The cheapest region for non-latency-sensitive batch processing. Graviton instances for services that passed the compatibility test. The FinOps dashboard shows the results: cloud costs are down 24% year-over-year on a normalized basis. That number goes in the board slides.
What doesn't go in the board slides: 67% of your compute capacity now runs in a single AWS region. Two years ago it was 41%. The placement engine made a series of individually rational decisions (us-east-1 is consistently cheaper than us-west-2 for your workload profile) and those decisions accumulated into a concentration that nobody planned, nobody approved, and nobody is currently tracking.
The SRE team talks about resilience in terms of availability zones and multi-AZ configurations within us-east-1. The FinOps team talks about cost in terms of unit economics and commitment coverage. Neither team has a metric for the third dimension of risk: what happens to the business if us-east-1 has a significant event. The answer, at current concentration, is that most of your compute goes down. An AWS us-east-1 degradation in December 2021 affected large portions of the internet precisely because concentration had accumulated over time, through exactly the kind of individually reasonable operational decisions your placement engine has been making on your behalf.
Legacy metrics don't capture this because they weren't designed to track second-order effects of optimization. Cost per unit measures how efficiently you're spending: it doesn't measure where that spend is concentrated. Availability zone redundancy tracks distribution within a region: it doesn't capture cross-region or cross-cloud concentration. Spot instance interruption rates measure pricing-change-driven reclamations: they don't measure the blast radius if an entire provider or region has a structural outage.
The measurement gap for ACR is this: there is no metric that tells you whether AI-driven cost optimization is concentrating your workloads, or how that concentration has changed over time as a direct consequence of optimization decisions. A metric exists for whether optimization is saving money. No metric exists for whether optimization is compounding resilience risk in the process. Without ACR, you learn about the concentration problem the hard way: when the provider or region has an event and the blast radius is larger than anyone expected, because nobody measured the risk accumulation that cost optimization created.
Formula
Arbitrage Concentration Risk (ACR) is the proposed metric for this: the degree to which AI-driven cost optimization has concentrated workloads in a single cloud provider, region, or availability zone, measured as the Herfindahl-Hirschman Index (HHI) of spend distribution across provider-region combinations, tracked over time. ACR borrows the HHI from antitrust economics, where it measures market concentration. The same mathematical property that makes HHI useful for detecting market concentration (it disproportionately weights the largest players) makes it useful for detecting workload concentration. A single provider-region combination at 80% of your compute spend contributes 0.64 to the index. The remaining 20% spread across five combinations contributes at most 0.04 total. High concentration registers unmistakably in a way that averaged metrics never would.
Formula: plain language:
For each cloud provider-region combination where you run active compute workloads, calculate that combination's share of your total compute spend. Square each share. Sum the squares. Higher values mean more concentration.
Formula: notation:
ACR = Σ s_k²Where:
- s_k = share of total cloud compute spend in provider-region combination k (expressed as a decimal, 0 to 1)
- Summation over all provider-region combinations with active workloads- ΔACR = ACR(current quarter) − ACR(prior quarter)
Worked Example
An organization runs compute across three environments. AWS us-east-1 accounts for 60% of compute spend. AWS us-west-2 accounts for 25%. GCP us-central1 accounts for 15%.ACR = (0.60)² + (0.25)² + (0.15)²ACR = 0.36 + 0.0625 + 0.0225
ACR = 0.445
Compare this to six months ago, when the distribution was 45% / 35% / 20%:ACR = (0.45)² + (0.35)² + (0.20)² = 0.2025 + 0.1225 + 0.04 = 0.365.ΔACR = +0.080 over six months. The cost optimizer has concentrated workloads measurably in us-east-1, and the trend is upward. That quantified trend is the signal ACR is designed to surface, not the absolute level, but the direction and rate of change driven by optimization decisions.
Sign convention:
Lower is better. 0 represents perfect distribution (theoretical, unachievable). 1.0 represents total concentration in one provider-region combination (possible and dangerous). The most informative use of ACR is the trend, a rising ACR over consecutive quarters, driven by optimization decisions, is the pattern to detect and respond to.
Edge Cases
Single-cloud organizations with deliberate single-cloud strategy: calculate ACR at the region level. Regional concentration within a single provider is a real and measurable resilience risk even if cross-cloud diversification isn't on the roadmap. Data-residency-constrained workloads that cannot move should be excluded from multi-cloud ACR calculations (the constraint explains the concentration) but included in regional ACR calculations: legal constraint doesn't reduce the blast radius if the region goes down. Non-compute spend (object storage, databases, networking) carries different concentration risk characteristics and should be calculated separately rather than mixed with compute ACR.
Cross-references:
Upstream: Commitment Portfolio Flexibility Index (CPFI, GMI-IT-014). Low CPFI combined with high ACR creates a doubly constrained situation. Workloads are concentrated in a single provider-region and the commitment portfolio can't follow if workloads need to move. This combination is the highest-risk configuration in the FinOps portfolio. Upstream: Right-Sizing Churn Cost Ratio (RSCR, GMI-IT-016). Aggressive right-sizing that consistently identifies the cheapest instance family in the cheapest region drives workloads toward concentrated configurations over time, as a cumulative effect of individually rational sizing decisions. Downstream: Cost Anomaly Investigation Rate (CAIR, GMI-IT-018), a provider-region outage affecting concentrated workloads creates simultaneous cost anomalies (spending suddenly drops or spikes as emergency failover occurs or workloads go dark) that require immediate investigation. High ACR makes CAIR events more consequential when they occur.
Proposed formula, pending validation pilots.