Medicaid Provider Spending

Why This Data

As a nurse, I've worked at the front line of healthcare, where care is actually delivered. This dataset was a chance to see one of the system's most important parts from the other side: how Medicaid pays for that care, and where the money goes.

Medicaid is one of the largest parts of the U.S. healthcare system. Together with CHIP, it covered about 73 million people in June 2026, nearly half of them children,² and in 2024 Medicaid spending reached $931.7 billion, about 18% of all U.S. health spending.³

The charts below show what the public data says, as plainly as possible: how much was paid, for what services, to whom, and how that changed over time. How the data was prepared, how the numbers were checked, and what the data can't tell us are all covered further down.

The Data

Dataset

Medicaid Provider Spending by HCPCS¹

Publisher

U.S. Department of Health and Human Services, HHS Open Data

Underlying data

CMS T-MSIS, the system states use to report Medicaid and CHIP claims

Version

Published February 9, 2026

Accessed

May 19, 2026

Coverage

January 2018 to December 2024; fee-for-service, managed care, and CHIP; all states and territories

One row equals

One billing provider × one servicing provider × one procedure code × one month

Size

10.76 GB CSV, 238,015,729 rows

What the Data Shows

Loading data…

SELECT CLAIM_MONTH, SUM(TOTAL_PAID), SUM(TOTAL_CLAIM_LINES), SUM(TOTAL_PATIENTS)
FROM prepared_data GROUP BY CLAIM_MONTH ORDER BY CLAIM_MONTH

SELECT HCPCS_CODE, SUM(TOTAL_PAID), SUM(TOTAL_CLAIM_LINES)
FROM prepared_data
WHERE YEAR(CLAIM_MONTH) BETWEEN :first_year AND :last_year
GROUP BY HCPCS_CODE ORDER BY SUM(TOTAL_PAID) DESC LIMIT 20

SELECT BILLING_PROVIDER_NPI_NUM, SUM(TOTAL_PAID)
FROM prepared_data
WHERE YEAR(CLAIM_MONTH) BETWEEN :first_year AND :last_year
  AND BILLING_PROVIDER_NPI_NUM <> 'UNKNOWN'
GROUP BY BILLING_PROVIDER_NPI_NUM ORDER BY SUM(TOTAL_PAID) DESC LIMIT 20

SELECT BILLING_PROVIDER_NPI_NUM, SUM(TOTAL_PAID) AS provider_total
FROM prepared_data
WHERE YEAR(CLAIM_MONTH) BETWEEN :first_year AND :last_year
  AND BILLING_PROVIDER_NPI_NUM <> 'UNKNOWN'
GROUP BY BILLING_PROVIDER_NPI_NUM
-- then each provider_total is counted into its range

A Closer Look: How Medicaid Billing Works

Every row in this dataset names two providers. The billing provider is the one Medicaid paid. The servicing provider is the one recorded as delivering the care. Often they are the same, such as a clinic billing for its own visits. They differ when an organization bills for care delivered by its staff or contractors, or when an intermediary bills on behalf of caregivers, as in the self-directed care programs mentioned above. Neither case suggests anything unusual; it is how much of Medicaid billing is structured.

The chart splits each year's dollars four ways, including money where one of the two providers isn't listed at all.

SELECT YEAR(CLAIM_MONTH),
  CASE WHEN BILLING_PROVIDER_NPI_NUM = 'UNKNOWN' THEN 'Billing provider not listed'
       WHEN SERVICING_PROVIDER_NPI_NUM = 'UNKNOWN' THEN 'Servicing provider not listed'
       WHEN BILLING_PROVIDER_NPI_NUM = SERVICING_PROVIDER_NPI_NUM THEN 'Same provider'
       ELSE 'Different provider' END,
  SUM(TOTAL_PAID)
FROM prepared_data GROUP BY 1, 2

What "not listed" means. The dataset's documentation keeps claim lines whose provider number is blank, and the blank shows up here as "not listed." The documentation doesn't say why a provider number is missing, so this page doesn't guess. It includes all of the state-specific code Z9525 from the procedure chart above.

How the Data Was Prepared

The raw file is too large for a spreadsheet, so I converted it to Parquet, a compressed columnar format, and queried it with DuckDB. Each step and its effect on the row count and total:

Why procedure code 20 was removed. Its 34 rows add up to , about times all U.S. health spending from every source in 2024 ($5.3 trillion).³ Payments that large can't be real, so they were treated as a data error. Every other row was kept, including rows with negative payments ( in total), which are most likely payment corrections; the dataset includes finalized adjusted claims.

How the numbers were checked

• Matches HHS to the cent. On HHS's own window (federal fiscal years 2019 to 2024), the top 20 procedures, top 20 billing providers, and all 72 monthly totals match the example charts HHS publishes for this dataset,⁶ in the same order.
• Matches my original exploration. Every year's total equals the yearly totals from the notebook where I first explored the data.
• Everything adds up. Monthly totals, each year's billing split, and the provider ranges all sum back to the overall total, and raw rows minus code 20 equal the final set.

The checks run automatically every time the numbers are rebuilt, and the build fails if any don't pass. The scripts and summary data are on GitHub. My full exploration notebook is available on request at [email protected].

What the Data Can't Tell Us

• It isn't all Medicaid spending. It covers outpatient and professional claim lines that carry a procedure code, so most hospital inpatient stays and pharmacy are not included. That is why seven years here total , while Medicaid spends about $930 billion in a single year.
• Small groups are hidden for privacy. CMS removed any row with fewer than 12 patients or 12 claim lines before releasing the data.
• 2024 is preliminary, and its last two months are incomplete. CMS notes that states are still finalizing their 2024 submissions. November () and December () 2024 are far below October (), most likely because their claims hadn't all arrived. They're included in totals but marked as incomplete, and year-over-year comparisons for 2024 use January through October.
• Patient counts can't be added up. The same person appears in many rows, across months, services, and providers, so the totals here are patient-service counts, not people.
• $0 payments are common. rows () show $0 paid. The documentation sets missing payment amounts to $0, and some managed care encounters carry no payment amount.
• Dollars aren't adjusted for inflation.
• Quality varies by state. The data is only as accurate as each state's submissions, and CMS cautions against comparing states directly.

Glossary

NPI

National Provider Identifier: a unique 10-digit number every U.S. healthcare provider, person or organization, uses on claims.

Billing provider

The provider that submitted the claim and was paid by Medicaid.

Servicing provider

The provider recorded as actually delivering the service. It can differ from the billing provider.

HCPCS / procedure code

The standard code on a claim that says what service was provided, such as an office visit or an hour of personal care.

Claim line

One service on a claim. A single claim with three services has three claim lines.

Patient-service count

Unique patients within one row (one provider pair, code, and month). Added across rows, the same person is counted more than once.

CHIP

Children's Health Insurance Program: coverage for children in families that earn too much for Medicaid but can't afford private insurance.

T-MSIS

Transformed Medicaid Statistical Information System: the system all states use to send Medicaid and CHIP data to CMS each month.

Sources

1. U.S. Department of Health and Human Services. Medicaid Provider Spending by HCPCS. HHS Open Data, published February 9, 2026. Accessed May 19, 2026.
2. Centers for Medicare & Medicaid Services. Medicaid & CHIP Enrollment Data Highlights, June 2026.
3. Centers for Medicare & Medicaid Services. National Health Expenditure Fact Sheet, 2024 data.
4. Centers for Medicare & Medicaid Services. NPPES NPI Registry.
5. New York State Department of Health. Consumer Directed Personal Assistance Program (CDPAP).
6. U.S. Department of Health and Human Services. Example charts for Medicaid Provider Spending by HCPCS (Examples tab), FY2019 to FY2024.