H28publishedNPD release 2026-08-20

Endpoint URL validity + machine-readable share

Of the 1,128,169 Endpoint resources in the NDH bulk export, only 9.8% are FHIR REST URLs an integrator can GET. The other 90.2% are Direct Trust HISP messaging addresses, which solve a different problem entirely. The right denominator for any "find FHIR endpoint by NPI" feature is the FHIR REST subset, not the full Endpoint resource count.

Headline

Of 1,128,169 Endpoint resources in the 2026-08-20 NDH bulk export, only 110,973 (9.8%) are machine-readable FHIR REST URLs an integrator can GET. The remaining 1,017,196 (90.2%) are Direct Trust HISP addresses (clinical messaging, not a queryable API), and 0 use other connection types. The 8.4% machine-readable share is the right denominator for any "find the FHIR endpoint for this provider" feature built on top of NDH.

111.0K / 1.1M = 9.84%

FHIR REST (machine-readable)111.0K
Direct Trust messaging1.0M
Other / unspecified0

unit: count

What this means

Startups + integrators

When sizing a "find this provider's FHIR endpoint" product, the right population is 110,973 hl7-fhir-rest endpoints, not 1,128,169 total. Direct Trust HISP addresses (provider@hisp.example.com) are clinical-messaging routes, not API URLs; treating them as substitutable will fail at integration time.

Payer ops teams

CMS-9115-F adequacy reads against the FHIR REST subset only. The 9.8% machine-readable share is your effective denominator for compliance reporting, not the headline 1,128,169 total Endpoint count.

CMS publishing the data

Two values share one Endpoint resource shape today. Splitting the published count by `connectionType` (or publishing a top-level "machine-readable endpoints" counter) would make the resource census far less misleading at a glance.

Researchers

Reports that compare "endpoint counts across releases" should split by connection type. The 73% drop in raw Endpoint count between April 2026-04-09 and May 2026-05-08 is dominated by Direct Trust dedup; the FHIR REST count is much steadier.

Null hypothesis

100% of NDH `Endpoint.address` values are well-formed for their declared `connectionType`. There is no surprising shape mismatch between connection-type-by-count and machine-readable share.

Denominator

All `Endpoint` resources in the pinned NDH bulk export.

Data source

BigQuery scan of `cms_npd.endpoint`, partitioned by `_connection_type` and validated against `_address` regex by type.

Notes

NDH ships two distinct connectionType.code values that share an Endpoint shape: hl7-fhir-rest (FHIR REST API URLs) and direct-project (Direct Trust HISP messaging addresses, of the form provider@hisp.example.com). Both are valuable but solve different problems. Of the 110,973 hl7-fhir-rest endpoints, 100% have a valid https:// URL — there is no http:// or malformed URL noise in this slice. 110,973 are flagged status=active. Crawler liveness (/findings/endpoint-liveness, H1-H5) probes the active subset. For consumers wiring 'find FHIR endpoint by NPI' — the right denominator is the 110,973 hl7-fhir-rest endpoints, NOT the 1,128,169 total resource count. This was a non-obvious cliff in the data quality dashboard until H28 surfaced it.

Get the next finding in your inbox. One email per release, no filler.