Shovel Watch
A source-aware operating view of the infrastructure contracts, bottlenecks, and overlooked suppliers turning AI demand into physical capacity.
6 seeded signals · facts / analysis split
Top new signals
Ranked by commercial momentum × bottleneck exposure
VERIFIED / ANALYSIS SPLIT
QumulusAI: $240.9M of new take-or-pay capacity
More than 2,000 NVIDIA Blackwell B300 GPUs across two three-year inference contracts; both customers are repeat buyers.
Alignment Engine: AMD clusters at Ohio campus
$1.17B in four-year take-or-pay orders with approximately $133M in contractual prepayments; customer identity and GPU count not disclosed.
Firmus: Meta regional AI-factory agreements
Contracted capacity plus expansion options in Southeast Asia; Firmus says portfolio contracted capacity exceeds 900MW.
Category heatmap
Current intensity, not a price target
Bottleneck map
Where time-to-token is most exposed
Parallel opportunity
Map the signal to the ecosystem
- BOSSTOX — rank repeat-customer and take-or-pay quality.
- Paxton SCOUT — map sites, power, suppliers, and contract timelines.
- ATLAS / Fenway Web — ship the operator intelligence layer.
- Kingswell / KEYSTONE — scout obscure component and regional operators.
Contracted Infrastructure
Signed demand beats speculative capacity. Every row retains its evidence state.
QumulusAI · $240.9M · 36 months
2,000+ B300 GPUs; take-or-pay; repeat customers; capacity targeted December 2026 / early 2027.
verified
Alignment Engine · $1.17B · 48 months
AMD MI355X / MI455X; Ohio; approximately $133M prepayment; customer and count undisclosed.
verified
Corning / AT&T · >$3B · multi-year
Fiber and cable supply agreement supporting network expansion as AI drives data demand.
verified
Optical Bottleneck Index
Physical connectivity is becoming a contracted input to AI scale.
88
Discovery targets
- Optical packaging and test equipment
- InP lasers, 800G / 1.6T transceivers
- Specialty glass, connectors, cable management
- Installers serving data-center corridors
AMD Infrastructure Supply Chain
A dedicated basket for non-NVIDIA accelerator deployments.
MI355X / MI455X
| Layer | Known signal | SCOUT / KEYSTONE angle |
|---|---|---|
| Accelerator | AMD MI355X; MI455X | Deployment timing, allocation, thermal envelope |
| Campus | Ohio; 154MW available, 480MW path | Power, substations, cooling, construction milestones |
| Commercial | $1.17B / 48 months / take-or-pay | Prepayment-backed infrastructure finance |
| Unresolved | Customer identity and GPU count not disclosed | Do not infer from contract value |
Quality score
High contract quality, high capacity expansion, high bottleneck exposure; obscurity remains high because the operating entity and customer are under-covered.
Regional AI Factories
Distributed capacity where power, land, and speed create an edge.
Watch for
- New anchor customers with expansion options
- Grid and submarine-cable constraints
- Factory-built power / cooling module suppliers
- Contracted capacity crossing 1GW
Grid-to-Chip
Power conversion, cooling, and onsite energy are converging into one delivery system.
92
Parallel opportunity
BOSSTOX: score suppliers by stack control. SCOUT: map the grid-to-rack chain. ATLAS: model alarms, commissioning, and maintenance. KEYSTONE: find pumps, heat exchangers, busbars, SSTs, and dielectric fluids.
AI for AI Infrastructure
Facility operations are becoming an intelligence layer on top of the physical build.
Phaidra Prism
Aligned Data Centers is deploying Phaidra Prism across its portfolio, connecting BMS, EPMS, mechanical, electrical, power, cooling, and IT data with NVIDIA DSX Exchange. The disclosed opportunity is not generic “AI software”; it is operator-grade reasoning against facility state.
Product mapping
- ATLAS — facility intelligence, evidence, and workflow memory.
- Fenway Web — customer-facing operator portal and dashboards.
- Paxton-Digital — source ingestion, entity graph, and commercial signals.
- Kingswell — investable vendor discovery and diligence trail.
System contract
Normalized records, deterministic scoring, and source-aware alerts.
SignalRecord {
id, company, entity_type, customer[], amount_usd, term_months,
capacity_mw, capacity_gw, accelerator[], geography[], source[],
confidence: verified | reported | inferred,
event_date, category, bottleneck_type[], commercial_momentum,
repeat_customer, prepayment_usd, take_or_pay, related_supplier[],
scores: { contract_quality, momentum, obscurity, bottleneck,
capacity_expansion, parallel_opportunity },
evidence: { fact[], analysis[], undisclosed[] }, dedupe_key
}
Scoring: ContractQuality = 30% disclosed terms + 25% take_or_pay +
20% prepayment + 15% repeat customer + 10% source reliability.
Alerts: new source event AND materiality threshold AND confidence != inferred.
Ingestion: source fetch → entity resolve → contract fingerprint → dedupe →
confidence label → score → publish to watchlist. Undisclosed ≠ zero.
