Docs → API reference

AI Infrastructure Construction (data centers + fabs) API v1

Generated from core.contract.describe_ai_infrastructure_construction_v1() and checked fixture-backed examples. Do not hand-edit the example JSON files.

Capability

• Capability: ai_infrastructure.construction
• Primitive: query_ai_infrastructure_construction_v1
• Status: available
• Current source: construction.census.c30_vip
• Source publisher: U.S. Census Bureau
• Snapshot behavior: as_of = latest resolves to an exact snapshot date in every response.

What It Can Answer

• Monthly private data-center construction spending ($ millions) — Census's Value of Construction Put in Place — seasonally-adjusted annual rate or not-adjusted monthly level, cited to the source cell.
• Monthly private computer/electronic/electrical (semiconductor-fab) construction spending ($ millions) on the same two bases.
• The data-center vs fab construction-spending trend over time (each as its own per-month series; the data-center series begins 2014-01).
• Whether a given month's value is preliminary, revised, or final — Census's own revision marking, carried verbatim.
• Row-level detail records when include_records is true.
• Raw Census workbook-cell evidence for any returned row-level citation.

Represented Facts

• Census C30 (Value of Construction Put in Place) reports the monthly value of private DATA-CENTER construction put in place ($ millions), served both seasonally-adjusted (annual rate) and not-seasonally-adjusted (monthly level), verbatim and cited
• Census C30 reports the monthly value of private COMPUTER/ELECTRONIC/ELECTRICAL (semiconductor-fab) construction put in place ($ millions) on the same two bases

Data Point Contract

• Data point: ai_infrastructure.construction
• Product spec: blocks/ai_infrastructure_construction/card.md
• Grain: category_monthly
• Source basis: construction.census.c30_vip
• Represented fact: Census C30 / Value of Construction Put in Place — the monthly value of PRIVATE construction put in place ($ millions) for data centers (a named subcategory under Office) and computer/electronic/ electrical manufacturing (the semiconductor-fab line), served both seasonally-adjusted (an annual rate) and not-seasonally-adjusted (a monthly level) exactly as Census publishes. Construction spending only — never total investment (servers/chips), which Census does not publish.

Does not answer:

• total data-center INVESTMENT (servers, chips, cooling, equipment) — this is construction put-in-place only; an investment total would require imputation Census does not publish
• data-center MW capacity, count, square footage, or location (this is $ of construction, not physical capacity)
• public or government data-center construction (the series served is PRIVATE construction only)
• which company, operator, or project is building (Census C30 carries no operator/project breakdown)
• summing the seasonally-adjusted series across months (it is an annual rate, not a monthly level) or summing across bases (refused with a note, never silently computed)
• state, county, plant, or lat/lon attribution (structural — this is a national aggregate with no geographic anchor)
• construction outside the two AI-infrastructure-relevant categories (the source has ~70 categories; only data center + the fab line are served)

REST Surface

• GET /v1/health
• GET /v1/capabilities
• GET /v1/ai-infrastructure/construction/schema
• POST /v1/ai-infrastructure/construction/query
• POST /v1/evidence/source-row

MCP Surface

• list_capabilities_v1
• describe_ai_infrastructure_construction_v1
• query_ai_infrastructure_construction_v1
• get_source_evidence_v1

Local MCP Setup

Some MCP clients launch servers from the user's home directory or ignore a configured cwd. Use uv run --directory so the server always starts from the repository project.

{
  "command": "uv",
  "args": [
    "run",
    "--directory",
    "/absolute/path/to/OSINT",
    "python",
    "-m",
    "core.mcp_server"
  ]
}

Leave EXASCALE_PARQUET_BASE / EXASCALE_RAW_BASE unset: the one server hosts every data point's tools, and with no override it resolves each block's promoted snapshots and raw archive from the repository layout. Setting either env var points ALL tools at one directory — a per-block path breaks every other block's tools. They exist only for single-source sandboxes and tests.

Request Schema

Filters:

• as_of
• category
• basis
• data_month
• data_month_from
• data_month_to
• year
• revision_status

Group by:

• category
• basis
• data_month
• year
• revision_status

Date range parameters:

• data_month_from
• data_month_to

Controls:

• include_records
• include_evidence
• limit
• order_by
• top_n
• order
• rollup_other

Ranking (how order_by / top_n / order join — order_by ranks groups by a metric, never a group_by dimension; top_n needs both a group_by and an order_by):

{
  "no_ranking": "Omit order_by and top_n to return all groups in group-key order.",
  "order": {
    "default": "desc",
    "valid_values": [
      "desc",
      "asc"
    ]
  },
  "order_by": {
    "accepts": "one of output.metrics",
    "note": "Ranks the groups by a metric (a measure). Not a group_by dimension \u2014 rows already come back grouped by each group_by field.",
    "requires": [
      "group_by"
    ],
    "valid_values": [
      "construction_spending_musd",
      "source_record_count"
    ]
  },
  "top_n": {
    "note": "Keeps the top N groups by order_by; the rest fold into one (other) remainder (additive metrics sum into it, non-additive ones are nulled) so the result still reconciles to summary.totals.",
    "requires": [
      "group_by",
      "order_by"
    ],
    "type": "positive integer"
  }
}

Output Schema

Aggregate metrics:

• construction_spending_musd
• source_record_count

Metric groups:

{
  "construction_spending_musd": [
    "construction_spending_musd"
  ],
  "records": [
    "source_record_count"
  ]
}

Response summary fields:

• group_count
• totals

Accepted fact policy:

• Query responses contain accepted, gate-passed facts only.
• Gate, monitor, and audit quality signals are internal controls, not agent-facing answer caveats.
• If a source snapshot is not fit to serve, the source must fail closed before it reaches this API.

Metric metadata:

Rollup rules:

• summary.totals.<metric> is the authoritative total for the full matched query.
• Grouped row metrics may be summed only when additive_across_groups is true.
• Do not sum grouped values for these non-additive metrics: construction_spending_musd.

Detail record fields returned when include_records is true:

• source_id
• sheet_name
• source_record_key
• report_period
• category
• basis
• data_month
• year
• revision_status
• source_date_label
• construction_spending_musd
• source_row_number
• source_col_number
• as_of
• raw_file_sha256
• citation

Row-level citation fields:

• source_id
• source_url
• source_file
• sheet
• source_row
• source_col
• raw_file_sha256
• as_of

Aggregate citation fields:

• source_id
• publisher
• source_url
• source_file
• raw_file_sha256
• as_of
• source_rows_count
• source_rows_sample
• verify
• lineage_filter

Codebooks

The complete machine-readable codebooks are included in capability-schema.json.

Checked Examples

The checked schema output is capability-schema.json.

Agent Workflow

1. Call list_capabilities_v1 and select ai_infrastructure.construction.
2. Call describe_ai_infrastructure_construction_v1 to inspect valid filters, groupings, metrics, and citation fields.
3. Call query_ai_infrastructure_construction_v1 with bounded JSON params.
4. If the answer needs proof, pass a returned row-level citation object to get_source_evidence_v1.
5. Answer with the resolved as_of and relevant citations. Present returned metrics as authoritative for their declared source, snapshot, grain, and aggregation.