Scrutica
Every value traces to a primary source, carries an authority tier (T1–T4), and documents its full derivation chain. Parameters are adjustable in-browser; changing an assumption propagates through the entire estimation pipeline so you can read off the sensitivity of each output to each input.
This page documents how every number on Scrutica was produced. Each section covers one estimation methodology, its input data, the confidence we assign to the result, and the main sources of uncertainty. You do not need to read the formulas to follow the methodology; the plain-language explanations stand on their own.
The methodology applies across ~4,500 facilities in 128 countries, ~18,600 bilateral supply-chain edges (FactSet Revere, WRDS-FactSet, CSET ETO, WRDS Compustat, and chip-deployment chains), 34 sovereign AI programs, 3,349export-control designations, and 13 GPU/accelerator generations plus two announced placeholders (R100 Rubin, Maia 200). Ten scheduled cron routes feed the tables; staleness is computed per record under four TTL classes (regulatory 1–6h, market 6–24h, structural 24–72h, narrative 24–168h).
The methodology applies across roughly 4,500 facilities in 128 countries, the ~18,600 supplier–customer relationships that connect them (drawn from financial-research databases and SEC filings), 34 sovereign AI programs, and 3,349 export-control designations. The site refreshes itself on a fixed schedule — ten scheduled jobs across each week — and every record carries its own freshness window: regulatory updates age within hours; structural facility data within days.
| Coverage | |||
|---|---|---|---|
Primary source for US-listed company financials. Capex, revenue, and subsidiary disclosures from regulatory filings. | T1: Government / Filing | Quarterly (10-Q, 10-K) | US-listed companies: financials, capex, segment data, Exhibit 21 subsidiaries |
US Department of Commerce list of entities subject to export license requirements. Primary source for export control status. | T1: Government / Filing | Monthly (Federal Register) | 3,400+ entities under US export controls |
Government Accountability Office reports and SIA tracker for CHIPS Act funding recipients, amounts, and facility status. | T1: Government / Filing | Quarterly (SIA); annual (GAO) | 52 CHIPS Act funded projects across 35 companies |
Peer-reviewed dataset of GPU clusters and frontier data center facilities. City-level location data; no coordinates for individual clusters. | T2: Peer-Reviewed | Monthly | 1,065 GPU clusters worldwide; 200 frontier data centers tracked in the Epoch frontier-DC dataset |
Georgetown CSET analysis of semiconductor supply chain relationships and market share data. | T2: Peer-Reviewed | Annual | Semiconductor supply chain topology: equipment, foundry, packaging relationships |
Quarterly earnings call transcripts with CFO revenue attribution by sovereign AI customer country. | T2: Peer-Reviewed | Quarterly | Sovereign AI revenue, customer disclosures, GPU shipment context |
Private capital transaction database. Deal sizes undisclosed ~45% of the time; relationship data always available. | T3: Industry / Analyst | Daily | Private capital deals: DC investments, AI infrastructure funding rounds |
Capital IQ corporate event feed. NLP extraction yields ~534 high-confidence facility/compute events from 15,810 unique events. | T3: Industry / Analyst | Daily | Corporate announcements: facility expansions, capacity additions, GPU deployments |
Grid Queue Filings (PJM, NYISO) Regional transmission organization interconnection queue data. Early-stage intelligence on planned facilities; queued capacity is not the same as operational capacity. | T3: Industry / Analyst | Monthly | 6,093 PJM interconnection queue entries (MW, status, applicant); NYISO and CAISO ingest planned post-launch |
Sovereign AI Research Files Compiled from government announcements, NVIDIA earnings transcripts, CRS reports, and press coverage. All FX conversions documented. | T3: Industry / Analyst | As published | 34 country sovereign AI programs: announced investment, government vs. private split, deployed-vs-announced reconciliation |
Press Reports / Industry Analysis Secondary source. Used only when no primary source available. Always marked is_estimated: true. | T4: Press / Secondary | As published | Facility announcements, capacity rumors, deployment speculation |
Three independent estimation paths for facility compute capacity: hardware-based (GPU count + spec sheet), power-based (thermal envelope inversion), cost-based (capex disaggregation). Each produces a point estimate with propagated estimation bounds.
How much computing power does a facility have? Three different ways to answer that question, depending on what data is available. When we know the exact hardware, the estimate is tight; when we only know the investment amount, the range widens accordingly.
Hardware: narrow bounds. Power: medium bounds. Cost: widest bounds.
Four sub-indices tracking $/petaFLOP-day across procurement models: Cloud Spot, Cloud Reserved, On-Premises, and Sovereign. Derived from published hourly rates and Epoch AI hardware specs (BF16 TFLOP/s with 2:4 structured sparsity).
What does a standard unit of AI training compute cost, and how does that cost vary by provider, geography, and procurement model? The same computing power can cost 3-10x more depending on where and how you buy it; that spread determines who can afford frontier model training.
Cloud pricing: Tier 1 (provider APIs). On-prem TCO: Tier 3 (modeled).
BFS propagation model over a weighted supply chain graph. Configurable shock severity, propagation delay, and per-category substitutability decay rates. Edge weights from FactSet Revere bilateral relationships.
If a key supplier is disrupted, who loses access and how badly? The simulation traces disruption through the AI supply chain; the output identifies which dependencies are hardest to replace and which countries face the greatest concentration risk.
Topology: Tier 2 (CSET, SEC filings). Decay rates: Tier 3 (expert assessment).
Every record carries data_source, source_url, and is_estimated. No value exists without attribution.
Every number on Scrutica links back to where it came from. You can always trace a value to its original source document, and every estimated value is labeled as such.
Append-only compute_capacity_snapshots track when values change. Every update preserves the previous state and records when the new value was learned.
When a value changes, we keep the old one. This means you can see how a facility’s reported capacity has evolved over time and when each update was recorded, not just the latest figure.
Values are set to null rather than guessed. A missing value is honest; a fabricated one is worse than useless.
If we do not have reliable data for a value, we leave it blank rather than fill it with a guess. A gap in the data is more useful than a confident-sounding number with no basis.
Three independent estimation paths, each usable when different input data is available. Hardware path takes GPU count and spec sheet directly; power path inverts the thermal envelope; cost path disaggregates capex. Estimation bounds widen as the input data becomes more indirect.
A facility’s compute capacity determines what models can be trained there and whether regulatory thresholds (such as the EU AI Act’s 1025 FLOP trigger) are met. Three estimation approaches accommodate different levels of available data; when less is known, the uncertainty range widens accordingly.
When GPU count and model are known. Narrowest estimation bounds.When the number and type of GPUs in a facility are known. Produces the most precise estimate.
When only power capacity is known. Derives GPU count from thermal envelope.When only the facility's electricity capacity is known. The estimate works backward from power consumption to GPU count.
When only investment amount is known. Widest estimation bounds.When only the total investment amount is known. Produces the widest range of uncertainty because hardware costs vary.
When GPU count and model are known. Narrowest estimation bounds.When the number and type of GPUs in a facility are known. Produces the most precise estimate.
| Range | |||
|---|---|---|---|
Interconnect Efficiency Fraction of peak throughput achievable across multi-node interconnect. 0.85 for NVLink, 0.7 for InfiniBand/Ethernet backend. Ethernet surpassed InfiniBand in AI back-end network market share in 2025 (Dell'Oro Group); UEC 1.0 spec released June 2025. Meta validated Ethernet RoCE at 24K-GPU scale for LLaMA 3. | 0.85 | 0.60–0.95 | Epoch AI |
Model FLOP Utilization (MFU) Fraction of theoretical peak FLOPs achieved during training. Typical range 30–50%. Chinchilla reported 0.46–0.57 (depending on model size); PaLM achieved 0.46–0.57; Meta reported 38–43% for LLaMA 3 405B at 16K-GPU scale. | 0.40 | 0.20–0.65 | Epoch AI |
When only power capacity is known. Derives GPU count from thermal envelope.When only the facility's electricity capacity is known. The estimate works backward from power consumption to GPU count.
| Range | |||
|---|---|---|---|
GPU Fraction of IT Load Fraction of IT power draw attributable to GPU accelerators vs. CPUs, storage, networking. | 0.70 | 0.40–0.90 | Industry estimates |
Interconnect Efficiency Fraction of peak throughput achievable across multi-node interconnect. 0.85 for NVLink, 0.7 for InfiniBand/Ethernet backend. Ethernet surpassed InfiniBand in AI back-end network market share in 2025 (Dell'Oro Group); UEC 1.0 spec released June 2025. Meta validated Ethernet RoCE at 24K-GPU scale for LLaMA 3. | 0.85 | 0.60–0.95 | Epoch AI |
Model FLOP Utilization (MFU) Fraction of theoretical peak FLOPs achieved during training. Typical range 30–50%. Chinchilla reported 0.46–0.57 (depending on model size); PaLM achieved 0.46–0.57; Meta reported 38–43% for LLaMA 3 405B at 16K-GPU scale. | 0.40 | 0.20–0.65 | Epoch AI |
Power Usage Effectiveness (PUE) Ratio of total facility power to IT equipment power. Lower is more efficient. Google: 1.10, industry average ~1.20. | 1.20 | 1.05–1.60 | Industry average; Google reports 1.10 |
When only investment amount is known. Widest estimation bounds.When only the total investment amount is known. Produces the widest range of uncertainty because hardware costs vary.
| Range | |||
|---|---|---|---|
GPU Fraction of Capex GPUs as a fraction of total data center capital expenditure. SemiAnalysis estimates 40–50%. | 0.45 | 0.30–0.60 | SemiAnalysis |
Interconnect Efficiency Fraction of peak throughput achievable across multi-node interconnect. 0.85 for NVLink, 0.7 for InfiniBand/Ethernet backend. Ethernet surpassed InfiniBand in AI back-end network market share in 2025 (Dell'Oro Group); UEC 1.0 spec released June 2025. Meta validated Ethernet RoCE at 24K-GPU scale for LLaMA 3. | 0.85 | 0.60–0.95 | Epoch AI |
Model FLOP Utilization (MFU) Fraction of theoretical peak FLOPs achieved during training. Typical range 30–50%. Chinchilla reported 0.46–0.57 (depending on model size); PaLM achieved 0.46–0.57; Meta reported 38–43% for LLaMA 3 405B at 16K-GPU scale. | 0.40 | 0.20–0.65 | Epoch AI |
Models trained above this line must be reported under the EU AI Act as general-purpose AI with systemic risk, triggering red-teaming, incident reporting, and cybersecurity requirements.
Normalized pricing in $/petaFLOP-day derived from published hourly rates and Epoch AI hardware specs (BF16 TFLOP/s with 2:4 structured sparsity). Four sub-indices cover cloud spot, cloud reserved, on-premises TCO, and sovereign procurement. All prices are converted to a common unit so effective compute cost is directly comparable across providers.
What does a standard unit of AI training compute actually cost, and how does that cost vary by provider, geography, and procurement model? This index converts all pricing to a single comparable unit. The same computing power can cost 3-10x more depending on where and how it is procured; that gap directly shapes who can afford to train frontier models and who cannot.
One petaFLOP-day is 1015 FLOP/s sustained for 86,400 seconds. This standardized unit enables direct cost comparison across hardware generations, cloud providers, and procurement models. The Cost Index expresses all prices as $/SCU.
The SCU is a standardized measure of AI computing power sustained for one day. Converting all prices to this common unit makes it possible to compare the cost of training compute across different providers, chip generations, and purchasing arrangements. When a hyperscaler quotes $2.50/hr for a GPU instance and a neocloud quotes $1.80/hr for a different configuration, the $/SCU conversion reveals which actually delivers cheaper compute per unit of training work.
BF16 TFLOP/s dense, no 2:4 structured sparsity (source: Epoch AI ML Hardware dataset). On-prem defaults: MFU , PUE , $5K networking + $3K facility share per GPU (editorial estimates, tier 4).
Both the FLOP Capacity Engine and the Cost Index quote dense BF16 Tensor Core throughput without 2:4 structured sparsity (989.5 TFLOP/s per H100 SXM, 312 per A100, 2,250 per B200). This matches the Epoch AI ml_hardware reference and the throughput real training jobs achieve; most production workloads do not enable sparsity, and the MFU parameter is calibrated against this dense baseline. NVIDIA marketing and cloud-provider cut sheets sometimes quote the 2:4-sparsity figure (1,979 TFLOP/s for H100) — cost-per-petaFLOP-day values in this index will be roughly double a number computed against those sparsity-inclusive specs; that gap is the unit mismatch, not a pricing disagreement.
On-demand / spot pricing for GPU compute from major cloud providers (AWS, GCP, Azure, Lambda, CoreWeave).
1-year and 3-year reserved instance pricing. Committed-use discounts.
Total cost of ownership for self-operated GPU clusters including hardware, power, cooling, facility, and labor.
Effective cost for government-funded sovereign AI compute including subsidies, grants, and below-market financing.
| Data Source | Cadence | Method |
|---|---|---|
| Cloud Provider APIs | Weekly | Automated API polling |
| Enterprise Contract Disclosures | Quarterly | SEC filing extraction |
| Hardware Vendor ASPs | Quarterly | Earnings transcript parsing |
| Electricity Tariff Data | Annually | EIA / regional utility databases |
| Construction Cost Indices | Annually | Turner / RSMeans indices |
BFS propagation over a weighted directed graph of supply chain relationships. Severity attenuates per hop via category-specific substitutability decay rates; edges weighted by FactSet Revere relationship values where available. When market correlation weighting is enabled, edge criticality blends editorial scores with FactSet 3-month stock price correlation to reflect market-revealed economic coupling. Configurable shock severity, propagation delay, and threshold cutoff.
If a critical supplier is disrupted, who loses access to what, and for how long? This simulation traces disruptions through the AI supply chain. Some components (EUV lithography equipment) have no substitutes; others (power delivery, construction) can be sourced from multiple vendors. The decay rate at each hop reflects that asymmetry.
The model starts at the disrupted company and traces outward through the supply chain, one step at a time. At each step, the severity of the disruption decreases based on how easy it is to find an alternative supplier for that category of goods. Equipment with no substitute (ASML’s EUV lithography machines) passes nearly full disruption downstream; categories with many vendors (construction, power delivery) absorb most of the shock. The process continues until the remaining severity falls below a 5% threshold or reaches six hops from the origin.
Severity attenuates at each hop based on the supply chain category's substitutability decay rate. Categories with near-zero substitutability (EUV lithography) propagate disruptions with minimal loss; categories with many alternatives (construction, power delivery) attenuate rapidly.
| Parameter | Default | Range | Description |
|---|---|---|---|
| propagation_delay | 30 days | 1–365 | Time for a disruption to propagate from one supply chain node to the next. |
| initial_shock_severity | 1 | 0.1–1 | Severity of the initial disruption at the source node (1.0 = complete disruption). |
Complete loss of TSMC Taiwan fab capacity (natural disaster, blockade).
Results from BFS propagation over the live supply chain graph (5 scenarios, 14 decay points each). For the full animated simulation, see the Cascade Simulation page. Higher decay = more severe propagation (each hop retains a larger fraction of the disruption).
Each supply chain category has a decay rate representing how much severity is absorbed per hop. Lower decay = harder to substitute = more severe cascade propagation.
Some parts of the supply chain can absorb disruptions because alternatives exist. Construction firms, power suppliers, and data center operators have multiple vendors. Semiconductor equipment is the opposite: ASML has no competitor for EUV lithography, so disruptions pass through with almost no loss. The decay rate for each category captures this difference.
Each supply chain layer holds a different amount of inventory that delays disruption propagation. Buffer data sourced from 10-K/annual report inventory line items (Tier 1), TrendForce/SemiAnalysis (Tier 2), and analyst estimates (Tier 3).
Real supply chains have stockpiles: chip makers hold weeks of inventory, memory suppliers pre-sell their output. These buffers determine how long downstream companies can operate before a disruption reaches them.
| Layer | Buffer | Range | Shape | Tier |
|---|---|---|---|---|
| EUV Lithography | None | 0–0 wk | cliff | T1 |
| Advanced Wafers (N3/N5) | 10 wk | 8–16 wk | linear | T1 |
| Packaged Chips (GPU) | 3 wk | 2–5 wk | cliff | T1 |
| Server Assembly | 14 wk | 8–26 wk | exponential | T3 |
| HBM Memory | 3 wk | 0–8 wk | cliff | T2 |
| ABF Substrates | 12 wk | 4–30 wk | linear | T2 |
When a supplier is disrupted, how fast can alternatives absorb demand? Qualification time is the lag before any substitute output appears; capacity ceiling is the maximum fraction absorbable. Ramp follows an S-curve from qualification to ceiling.
Not all supply chain links are equally fragile. Some disrupted suppliers can be replaced in months (Micron HBM); others have no substitute at all (ASML EUV lithography). This table shows how long replacement takes and how much demand alternative suppliers can absorb.
TSMC Advanced Node
ASML EUVNo viable substitute
SK Hynix HBM
NVIDIA Training GPUs
TSMC CoWoS Packaging
Temporal parameters calibrated against three observed disruption events. Observed propagation delays and buffer absorption patterns validate the model's depletion and substitution functions.
We tested the model against real supply chain disruptions to verify that the predicted timelines match what actually happened. Three events span distinct disruption archetypes: sudden shutdown, commodity shock absorbed by buffers, and sustained capacity bottleneck.
Source: TrendForce March 2021; Samsung Q1 2021 earnings ($268-357M loss) (T1)
Source: USITC Executive Briefing (DeCarlo & Goodman, Apr 2022); CSIS March 2022; Reuters (T1)
Source: TSMC Chairman Mark Liu (Sep 2023); TrendForce Aug 2024 (SPIL order); UBS analyst estimates (T1)
Edge weights draw on three signals. Supply share (% of customer input from this supplier) is available on ~3.4% of edges; the remaining use a default of 30%. Criticality (1–10 replaceability) is editorial for all edges but can be blended with FactSet 3-month stock price correlation (Pearson r, available on ~78% of FactSet edges) when market correlation weighting is enabled. The blend formula is adjustedCriticality = w × 5(1+r) + (1−w) × editorial, where w defaults to 0.5. For edges without correlation data (WRDS, unlisted companies), editorial scores stand alone. Sole-source edges retain criticality ≥ 9 regardless of correlation. Price correlations are from FactSet Workstation downloads (vintage: April 2026). The structural centrality proxy — the product of flow share and criticality — follows Li et al. 2020 (PMC7546950).
The “total affected compute” metric blends two weighting approaches: for nodes with facility-level FLOP/day estimates, impact is FLOP-weighted; for nodes without FLOP data, impact is weighted by network degree centrality (number of supply chain connections). Users should interpret this as an approximation — the topology-weighted component is a structural proxy, not a measurement.
The supply chain graph is assembled from multiple data sources; each edge carries its own provenance and authority tier.
The supplier/customer relationships that feed this simulation come from four sources of varying reliability. Relationships from SEC regulatory filings (Tier 1) are the most reliable; market-share-derived links from CSET (Tier 2) capture industry structure without bilateral specificity; press and analyst reports (Tier 3) fill gaps.
Four authority tiers (T1: primary measurement, T2: research database, T3: press/analyst, T4: estimated/inferred) assigned by original source authority, not intermediary. A CRS report citing a Goldman Sachs estimate remains T3. Aggregate confidence for derived values takes the weakest input tier.
A number from a company’s SEC filing is more reliable than one from a press report, which in turn is more reliable than an analyst estimate. Every value on Scrutica carries a confidence label (T1 through T4) so you can judge how much weight to place on it. When a derived value combines sources of different quality, the overall confidence reflects the weakest input.
Value from a primary authoritative source with direct measurement or legal disclosure obligation.
Value from academic publication or research organization with documented methodology.
Value derived from proxy methods, industry analyst estimates, or press reports. Methodology documented.
Value from secondary inference, expert opinion, or interpolation. Widest estimation bounds.
When a derived value combines inputs from multiple tiers, the aggregate confidence is the lowesttier among all inputs. A GPU count estimate (Tier 3) combined with a verified power capacity (Tier 1) yields a Tier 3 aggregate, because the weakest link determines the chain's strength.
When Scrutica calculates a number from multiple inputs, the overall confidence reflects the least reliable source. If one input comes from a manufacturer's SEC filing (highly reliable) but another comes from a press estimate (less reliable), the combined result inherits the lower confidence. This prevents derived values from appearing more certain than their weakest ingredient.
When sources disagree on a factual value, Scrutica shows both values with source attribution rather than silently picking a winner. The protocol:
is_estimated FlagThis boolean means "this value was derived via a model or proxy method." It does notmean "this value might be null" or "this value has uncertainty."
Scrutica estimates are compared against independently published figures from Epoch AI, CSET, TrendForce, and MLPerf. Where estimates diverge, the discrepancy traces to different assumptions about GPU utilization rates, hardware configuration, or facility-level power draw.
| Validation Source | Data Type | Coverage | Methodology Difference |
|---|---|---|---|
| Epoch AI GPU Clusters | Facility PFLOP/s estimates | ~786 clusters, ~26 with PFLOP estimates | Epoch uses disclosed GPU counts; Scrutica uses three estimation paths (hardware, power, cost) and reports estimation bounds for each |
| CSET AI Indicators | Company-level AI investment | ~691 companies | CSET aggregates financial disclosures; Scrutica adds FactSet Revere supply chain edges for interdependence analysis |
| TrendForce / Counterpoint | Semiconductor market share | Quarterly updates | Analyst estimates vs. Scrutica concentration scoring derived from supply chain topology |
| MLPerf Inference | Hardware specifications (accelerator type, memory, count) | 92 datacenter system submissions (v4.1) | MLPerf publishes verified hardware configurations per submission; Scrutica cross-checks accelerator types and memory specs against its chip catalog |
MLPerf Hardware Cross-Check
20 unique system configurations from MLPerf Inference v4.1 submissions provide independent verification of accelerator specifications. Accelerator families represented: AMD Instinct MI300X, NVIDIA H100, NVIDIA L40S, NVIDIA Jetson AGX Orin 64G, NVIDIA H200, AMD MI300X, NVIDIA L4, NVIDIA B200, NVIDIA A100, NVIDIA GH200, TPU v5e, TPU v6, NVIDIA GH200 Grace Hopper Superchip 144GB, NVIDIA GH200 Grace Hopper Superchip 96GB, UntetherAI speedAI240 Preview, UntetherAI speedAI240 Slim.
Facility-level PFLOP/s cross-validation requires overlapping estimates from independent sources. As Scrutica and Epoch coverage expands, matched facility pairs will appear in the table above. Users who want to reproduce our estimates can adjust all parameters in the Interactive Methodology Explorer above.
When citing specific estimates, include the methodology version and estimation path (e.g., "Hardware Path, v1.2.0") so readers can reproduce the calculation.