Health Samurai pushed out an open-source FHIR server performance benchmark on 2026-06-29 that puts Aidbox, HAPI FHIR, Medplum, and the Microsoft FHIR Server on the same hardware, with the same Synthea dataset, and reruns the suite daily. The full dashboard has a lot of numbers in it. Five of them are worth keeping in your head if you are deciding which servers to even short-list. For more on the surrounding picks, the FHIR primer index on the site collects the rest.
The 5 Numbers Worth Knowing
1. Aidbox CRUD Throughput: About 5,212 RPS
Across the create, read, update, and delete mix the benchmark drives, Aidbox handled around 5,212 requests per second. That is the highest CRUD number in the snapshot. The 8 vCPU and 24 GB of memory allocation is the same for every server, so the throughput gap is the engines themselves, not the hardware.
2. Microsoft FHIR Server CRUD Throughput: About 440 RPS
The Microsoft FHIR Server posted around 440 RPS on the same CRUD workload, the lowest of the four. The split is wider than the casual reader might expect, and it shows up clearly when the dashboard puts the bars next to each other. HAPI lands at 3,058 RPS and Medplum at 1,420 RPS in the middle of the field.
3. HAPI Storage Footprint: 22.6 GB
After loading the same 1,000-patient Synthea dataset, HAPI uses 22.6 GB of storage. That is the largest of the four. The reason is straightforward: HAPI pre-builds search indexes on write, which is the safer default for a general-purpose server. The storage cost is the trade. Medplum sits at 11.8 GB and the Microsoft server at 4.24 GB, with their own indexing strategies driving the spread.
4. Aidbox Storage Footprint: 6.83 GB
Aidbox stores the same data in 6.83 GB. The benchmark notes that Aidbox ships without default search indexes, which is faster on import and smaller on disk, but turns indexing into a deliberate operator choice rather than a free default. Some teams will read that as flexibility. Others will read it as a step they have to remember. Both readings are honest.
5. Top Search Throughput: About 3,404 RPS
On the search workload that exercises string, date, reference, token, quantity, and composite searches, Aidbox handled around 3,404 RPS. Medplum followed at 1,796 RPS, HAPI at 1,005 RPS, and the Microsoft server at 261 RPS. The note in the dashboard adds two useful details: Medplum does not support composite search, and the Microsoft server is slow on quantity and composite searches in particular.
How to Read the Spread
The spread between the top and bottom number on every metric is wider than most procurement teams expect when they walk into a comparison. That is partly because the four servers come from different architectural traditions. Aidbox is a FHIR-native database; HAPI is a Java reference implementation maintained by Smile Digital Health; Medplum is a TypeScript stack; the Microsoft server runs on SQL Server. Each one made a different bet on indexing, storage, and concurrency, and those bets show up as throughput and storage numbers when the same workload hits all four at once.
For the hospital-IT side of this picture, Top 6 FHIR terminology servers for hospital IT in 2026 covers the terminology layer that lives next to the FHIR server in most stacks. For the open-source side specifically, best open-source terminology servers for FHIR teams in 2026 walks through the adjacent picks.
The five numbers above are a starting point, not a verdict. The benchmark runs daily, the repo is public, and the snapshot can shift between this week and next. The short version is that the spread is wide enough to matter and narrow enough to argue about.
