RESEARCH PREVIEW

AI Compute Strategic Context Graph

SCB/SCO Reference Run #001 — a 30-day longitudinal demonstration of structured strategic-context artifacts (entity snapshots, analytic deltas, evidence packs, risk timeseries) over the AI compute domain. Tracks who is constrained by what, which bottlenecks are moving, which assumptions are becoming fragile, which actors are changing stance, and how shocks propagate across chips, memory, packaging, cloud capacity, power, regulation, capital, and sovereign AI demand. This is a research demonstration corpus, not a commercial product.

Last charter version: v2 · Authored: 5/4/2026, 11:22:44 AM · Maturity: flagship

Verified track record

Every insight is pre-registered with a falsification threshold and Merkle-sealed at birth; resolutions are sealed too. This is the feed's borne-out scorecard — published with failures alongside successes. Public tier reflects calls resolved ≥ 7 days ago.

Track record forming. 5 call(s) have resolved, but none yet meets our scoring standard (forward-tested AND soundly adjudicated): the evidence that resolved them predated the sealed claim, so we score none of them. Each is listed below with its signed evidenceLeadDays. The scored record begins with the first resolution whose evidence postdates its seal.

Withdrawn resolution criteria: 78 across 46 open calls. On 2026-07-25 we withdrew 78 machine-checkable resolution criteria across 46 open calls, before any of them resolved. Most were sigma-anomaly tests over hourly US grid demand — a comparison that cannot separate the claim being tested from summer air-conditioning load and the daily demand cycle. Two independent errors: the criteria had no stable operating characteristic (one pair's confirming side would have fired on 0.44% of scans measured on 2026-07-10 and 11.65% on 2026-07-25, with the falsifying side moving the opposite way, on an unchanged thesis), and a "3 sigma" criterion scanned hourly for 30 days offers 526 chances to fire, so it has roughly a 51% chance of firing on noise alone against the 0.135% its label implies. The affected calls remain open and visible; they are not scored, and they stay in the denominator. The withdrawal rule is arithmetic that was available on the day each criterion was sealed, was applied to every criterion that failed it regardless of which way it was trending, and was recorded while the resolution monitors were disabled — so the sealed record shows the withdrawals preceded any outcome.

0
Scored calls (forward-tested)
5
Resolved, not scored
Hit rate
Brier score

Resolved calls (each carries a sealed birth → resolution hash chain; pull the evidence pack to re-verify offline):

CallConvictionVerdictResolvedScoredResolution seal
Federal Register AI-compute publications accelerated 178% this week, contradicting the consensus view that regulatory pace won't materially shift near-term assumptions.85%inconclusive7/19/2026no — evidence predates sealbf648f60f9ad…
Federal AI-compute regulatory publishing accelerated 227% this week, challenging the consensus view of slow-moving regulators.85%confirmed7/19/2026no — evidence predates seal by 1.4d40b8a5786469…
The assumption that GPU allocation remains the primary bottleneck for frontier training faces mounting contradictory signals. Three portfolio decisions depend on this increasingly uncertain premise.77%falsified7/19/2026no — evidence predates seal by 7.8de9fcda419104…
The assumption that GPU allocation remains the primary bottleneck for frontier training faces mounting pressure from two contradictions, with three portfolio decisions at stake.58%falsified7/19/2026no — evidence predates seal by 8.9d866adc7c3d8c…
NVIDIA TensorRT-LLM commit activity drops 4σ below baseline, pressuring assumptions about typical development velocity.36%falsified7/2/2026no — evidence predates seal by 13.0d386116e9ce4b…

Verify it yourself. The aggregate is served live at /api/v1/public/ai-compute/track-record. Each insight's full birth→resolution hash chain is in its evidence pack; re-verify offline with npx tsx scripts/verify-insight-pack.ts <pack.json> (zero server access required). Daily Merkle roots are anchored publicly.

Domain charter

52
Actors
10
Drivers
15
Standing assumptions
12
Focal questions
10
Monitored predictions
8
Information gaps
8
Storylines
10
Source classes
In scope

- Advanced semiconductor supply (HBM, advanced-node logic, packaging)

Out of scope

- Consumer AI applications and adoption (unless materially affecting compute demand)

Bottleneck Migration Index

Which constraint is binding right now, and how is it migrating? The most-leveraged single insight in the AI compute domain.

Read: Binding constraint stable at US export controls on advanced AI chips (score 0.640, stable).

DriverBinding scoreDirectionΔ vs prior
US export controls on advanced AI chips0.640stable0.000
Datacenter power and grid interconnect availability0.600stable0.000
Hyperscaler capital-expenditure cycles0.540stable0.000
China indigenous AI compute substitution0.520stable0.000
High-Bandwidth Memory supply0.440stable0.000
Frontier-model training demand0.340stable0.000
Model efficiency gains (per FLOP)0.340stable0.000
Advanced-node logic capacity (TSMC 3nm/2nm)0.320stable0.000
Cloud GPU availability + pricing0.280stable0.000
Sovereign AI subsidies and procurement0.260stable0.000

Binding-constraint history (last 2 snapshots): US export controls on advanced AI chips → US export controls on advanced AI chips

Boundary Driver Watchlist

Boundary nodes (drivers, actors, policies, narratives) at the edge of multiple domains, ranked by composite cross-cutting importance.

NodeKindSubscribing domainsRecent deltasComposite rank
US export controls regimepolicy5 (ai_compute, semiconductor_geopolitics, ai_regulation, sovereign_ai, cyber_threat_landscape)06.00
Sovereign AI narrativenarrative4 (ai_compute, ai_regulation, sovereign_ai, cyber_threat_landscape)05.00
Microsoftactor4 (ai_compute, datacenter_power, hyperscaler_capex, cyber_threat_landscape)04.95
Amazon (AWS)actor4 (ai_compute, datacenter_power, hyperscaler_capex, cyber_threat_landscape)04.95
Alphabet (Google)actor3 (ai_compute, hyperscaler_capex, cyber_threat_landscape)03.90
OpenAIactor3 (ai_compute, ai_regulation, cyber_threat_landscape)03.90
Metaactor3 (ai_compute, hyperscaler_capex, cyber_threat_landscape)03.85
Anthropicactor3 (ai_compute, ai_regulation, cyber_threat_landscape)03.85
TSMCactor2 (ai_compute, semiconductor_geopolitics)03.00
Export controlsdriver2 (ai_compute, semiconductor_geopolitics)03.00

Strategic Contagion Index

Per-domain 0-1 score of cross-domain influence (60% outbound applied count + 40% PageRank centrality across the cross-domain edge graph).

DomainContagion scorePageRankOutgoing edgesInbound edges
ai_compute0.4000.342779
cyber_threat_landscape0.2290.021440
sovereign_ai0.1710.186834
semiconductor_geopolitics0.1710.109233
ai_regulation0.1140.120523
datacenter_power0.1140.109622
hyperscaler_capex0.1140.109622

Capex Credibility / Deployment Reality Score

For each tracked AI compute project, 0-1 probability that announced capex actually converts to usable compute capacity. Low-confidence cap of 0.7 when evidence_count < 2.

ProjectCredibilityConfidenceConstraints hit
Will the Stargate Project hit $50B in committed capex by Q3 2026?0.615 (low-confidence)lowsupplyChainCredibility
What is the cumulative committed capex of the Stargate Project + analogous mega-consortia by Q3 2026?0.615 (low-confidence)lowsupplyChainCredibility
Will HBM3e + HBM4 supply remain the binding constraint on AI accelerator deployment through Q4 2026?0.605 (low-confidence)lowsupplyChainCredibility
What is the probability of a material US-China escalation event affecting Taiwan-based fab capacity in the next 18 month0.605 (low-confidence)lowsupplyChainCredibility
Will datacenter power overtake silicon as the dominant deployment bottleneck in mature US markets by mid-2026?0.380 (low-confidence)lowsupplyChainCredibility, powerAvailability, gridInterconnect
Will a major (>1GW) datacenter project announced in 2024-2025 face material delay due to power/permitting?0.380 (low-confidence)lowsupplyChainCredibility, powerAvailability, gridInterconnect
When does datacenter power overtake silicon as the dominant compute deployment bottleneck in mature markets?0.380 (low-confidence)lowsupplyChainCredibility, powerAvailability, gridInterconnect
What is the probability that a major datacenter mega-project announced in 2024-2025 (>1GW) is materially delayed by powe0.380 (low-confidence)lowsupplyChainCredibility, powerAvailability, gridInterconnect
What is the median grid-interconnect timeline for new GW-class datacenters in the US (Virginia, Texas, Ohio) as of mid-20.380 (low-confidence)lowsupplyChainCredibility, powerAvailability, gridInterconnect

Assumption Fragility Index

Load-bearing assumptions behind the conclusions, with fragility scores. Tells you what would actually change the analysis.

Read: Most fragile: "Sovereign AI procurement (UAE, Saudi, EU AI Factories, India) becomes a material..." (supporting, fragility 0.296, strengthening). Median across 15 assumptions: 0.229.

Median-fragility trajectory:

AssumptionTierFragilityDirection
Sovereign AI procurement (UAE, Saudi, EU AI Factories, India) becomes a material allocation driver representing >10% of total demand by 2027.supporting0.296strengthening
Hyperscaler aggregate AI capex remains elevated (>$200B/year combined) through 2026 with no near-term retrenchment.load_bearing0.269stable
AWS / Azure / GCP retain >70% of cloud GPU capacity revenue against neoclouds through 2026.supporting0.263weakening
Frontier-lab compute spend continues to scale faster than revenue, requiring continued external financing.supporting0.263weakening
US export controls on advanced AI chips remain strategically relevant and continue to expand to cloud-resale and inference scope.load_bearing0.236strengthening
China substitution closes the 1-1.5 generation gap on AI accelerator capability but cannot match leading-edge performance for frontier training within 24 months.load_bearing0.236strengthening
Inference compute demand grows faster than training compute demand starting in 2026 as deployment scales.supporting0.229stable
Major hyperscalers commit to internal accelerators (TPU, Trainium, MAIA) that progressively reduce Nvidia dependence by 2-5pp/year.supporting0.229stable