diff --git a/Cargo.lock b/Cargo.lock
index 80649de..a206ea4 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -137,8 +137,10 @@ dependencies = [
 name = "attic"
 version = "0.1.0"
 dependencies = [
+ "async-stream",
  "base64 0.20.0",
  "bindgen",
+ "bytes",
  "cxx",
  "cxx-build",
  "digest",
@@ -197,6 +199,7 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-compression",
+ "async-stream",
  "async-trait",
  "attic",
  "attic-token",
@@ -213,6 +216,7 @@ dependencies = [
  "digest",
  "displaydoc",
  "enum-as-inner",
+ "fastcdc",
  "futures",
  "hex",
  "humantime",
@@ -229,6 +233,7 @@ dependencies = [
  "serde_with",
  "sha2",
  "tokio",
+ "tokio-test",
  "tokio-util",
  "toml",
  "tower-http",
@@ -1483,6 +1488,12 @@ version = "2.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0206175f82b8d6bf6652ff7d71a1e27fd2e4efde587fd368662814d6ec1d9ce0"
 
+[[package]]
+name = "fastcdc"
+version = "1.0.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f3f1596230ad22715a97a82deba0403ece9e4a458d008fd2511518d72a115bef"
+
 [[package]]
 name = "fastrand"
 version = "1.8.0"
diff --git a/attic/Cargo.toml b/attic/Cargo.toml
index 3ef5ee6..d120cba 100644
--- a/attic/Cargo.toml
+++ b/attic/Cargo.toml
@@ -5,7 +5,9 @@ edition = "2021"
 publish = false
 
 [dependencies]
+async-stream = { version = "0.3.3", optional = true }
 base64 = "0.20.0"
+bytes = "1.3.0"
 displaydoc = "0.2.3"
 digest = "0.10.6"
 ed25519-compact = "2.0.4"
@@ -56,4 +58,4 @@ nix_store = [ "dep:cxx", "dep:bindgen", "dep:cxx-build" ]
 # Tokio.
 #
 # When disabled, any part depending on tokio is unavailable.
-tokio = [ "dep:tokio" ]
+tokio = [ "dep:tokio", "dep:async-stream" ]
diff --git a/attic/src/stream.rs b/attic/src/stream.rs
index 8d71c43..21908fa 100644
--- a/attic/src/stream.rs
+++ b/attic/src/stream.rs
@@ -1,13 +1,19 @@
 //! Stream utilities.
 
+use std::collections::VecDeque;
+use std::future::Future;
 use std::marker::Unpin;
 use std::pin::Pin;
 use std::sync::Arc;
 use std::task::{Context, Poll};
 
+use async_stream::try_stream;
+use bytes::Bytes;
 use digest::{Digest, Output as DigestOutput};
+use futures::stream::{BoxStream, Stream, StreamExt};
 use tokio::io::{AsyncRead, AsyncReadExt, ReadBuf};
 use tokio::sync::OnceCell;
+use tokio::task::spawn;
 
 /// Stream filter that hashes the bytes that have been read.
 ///
@@ -19,6 +25,79 @@ pub struct StreamHasher<R: AsyncRead + Unpin, D: Digest + Unpin> {
     finalized: Arc<OnceCell<(DigestOutput<D>, usize)>>,
 }
 
+/// Merge chunks lazily into a continuous stream.
+///
+/// For each chunk, a function is called to transform it into a
+/// `Stream<Item = Result<Bytes>>`. This function does something like
+/// opening the local file or sending a request to S3.
+///
+/// We call this function some time before the start of the chunk
+/// is reached to eliminate delays between chunks so the merged
+/// stream is smooth. We don't want to start streaming all chunks
+/// at once as it's a waste of resources.
+///
+/// ```text
+/// | S3 GET | Chunk | S3 GET | ... | S3 GET | Chunk
+/// ```
+///
+/// ```text
+/// | S3 GET | Chunk | Chunk | Chunk | Chunk
+/// | S3 GET |-----------^       ^       ^
+///              | S3 GET |------|       |
+///              | S3 GET |--------------|
+///
+/// ```
+///
+/// TODO: Support range requests so we can have seekable NARs.
+pub fn merge_chunks<C, F, S, Fut, E>(
+    mut chunks: VecDeque<C>,
+    streamer: F,
+    streamer_arg: S,
+    num_prefetch: usize,
+) -> Pin<Box<impl Stream<Item = Result<Bytes, E>>>>
+where
+    F: Fn(C, S) -> Fut,
+    S: Clone,
+    Fut: Future<Output = Result<BoxStream<'static, Result<Bytes, E>>, E>> + Send + 'static,
+    E: Send + 'static,
+{
+    let s = try_stream! {
+        let mut streams = VecDeque::new(); // a queue of JoinHandles
+
+        // otherwise type inference gets confused :/
+        if false {
+            let chunk = chunks.pop_front().unwrap();
+            let stream = spawn(streamer(chunk, streamer_arg.clone()));
+            streams.push_back(stream);
+        }
+
+        loop {
+            if let Some(stream) = streams.pop_front() {
+                let mut stream = stream.await.unwrap()?;
+                while let Some(item) = stream.next().await {
+                    let item = item?;
+                    yield item;
+                }
+            }
+
+            while streams.len() < num_prefetch {
+                if let Some(chunk) = chunks.pop_front() {
+                    let stream = spawn(streamer(chunk, streamer_arg.clone()));
+                    streams.push_back(stream);
+                } else {
+                    break;
+                }
+            }
+
+            if chunks.is_empty() && streams.is_empty() {
+                // we are done!
+                break;
+            }
+        }
+    };
+    Box::pin(s)
+}
+
 impl<R: AsyncRead + Unpin, D: Digest + Unpin> StreamHasher<R, D> {
     pub fn new(inner: R, digest: D) -> (Self, Arc<OnceCell<(DigestOutput<D>, usize)>>) {
         let finalized = Arc::new(OnceCell::new());
@@ -105,6 +184,9 @@ pub async fn read_chunk_async<S: AsyncRead + Unpin + Send>(
 mod tests {
     use super::*;
 
+    use async_stream::stream;
+    use bytes::{BufMut, BytesMut};
+    use futures::future;
     use tokio::io::AsyncReadExt;
     use tokio_test::block_on;
 
@@ -135,4 +217,45 @@ mod tests {
         assert_eq!(expected.len(), *count);
         eprintln!("finalized = {:x?}", finalized);
     }
+
+    #[test]
+    fn test_merge_chunks() {
+        let chunk_a: BoxStream<Result<Bytes, ()>> = {
+            let s = stream! {
+                yield Ok(Bytes::from_static(b"Hello"));
+            };
+            Box::pin(s)
+        };
+
+        let chunk_b: BoxStream<Result<Bytes, ()>> = {
+            let s = stream! {
+                yield Ok(Bytes::from_static(b", "));
+                yield Ok(Bytes::from_static(b"world"));
+            };
+            Box::pin(s)
+        };
+
+        let chunk_c: BoxStream<Result<Bytes, ()>> = {
+            let s = stream! {
+                yield Ok(Bytes::from_static(b"!"));
+            };
+            Box::pin(s)
+        };
+
+        let chunks: VecDeque<BoxStream<'static, Result<Bytes, ()>>> =
+            [chunk_a, chunk_b, chunk_c].into_iter().collect();
+
+        let streamer = |c, _| future::ok(c);
+        let mut merged = merge_chunks(chunks, streamer, (), 2);
+
+        let bytes = block_on(async move {
+            let mut bytes = BytesMut::with_capacity(100);
+            while let Some(item) = merged.next().await {
+                bytes.put(item.unwrap());
+            }
+            bytes.freeze()
+        });
+
+        assert_eq!(&*bytes, b"Hello, world!");
+    }
 }
diff --git a/integration-tests/basic/default.nix b/integration-tests/basic/default.nix
index 7ab9476..8b86f48 100644
--- a/integration-tests/basic/default.nix
+++ b/integration-tests/basic/default.nix
@@ -126,6 +126,13 @@ in {
           credentialsFile = "/etc/atticd.env";
           settings = {
             listen = "[::]:8080";
+
+            chunking = {
+              nar-size-threshold = 1;
+              min-size = 64 * 1024;
+              avg-size = 128 * 1024;
+              max-size = 256 * 1024;
+            };
           };
         };
 
@@ -196,6 +203,13 @@ in {
           server.succeed("${cmd.atticd} --mode garbage-collector-once")
           client.fail(f"curl -sL --fail-with-body http://server:8080/test/{test_file_hash}.narinfo")
 
+      ${lib.optionalString (config.storage == "local") ''
+      with subtest("Check that all chunks are actually deleted after GC"):
+          files = server.succeed("find /var/lib/atticd/storage -type f")
+          print(f"Remaining files: {files}")
+          assert files.strip() == ""
+      ''}
+
       with subtest("Check that we can destroy the cache"):
           client.succeed("attic cache info test")
           client.succeed("attic cache destroy --no-confirm test")
diff --git a/server/Cargo.toml b/server/Cargo.toml
index 873cc39..8623091 100644
--- a/server/Cargo.toml
+++ b/server/Cargo.toml
@@ -23,6 +23,7 @@ attic = { path = "../attic", default-features = false, features = [ "tokio" ] }
 attic-token = { path = "../token" }
 
 anyhow = "1.0.68"
+async-stream = "0.3.3"
 async-trait = "0.1.60"
 aws-config = "0.52.0"
 aws-sdk-s3 = "0.22.0"
@@ -36,6 +37,7 @@ derivative = "2.2.0"
 digest = "0.10.6"
 displaydoc = "0.2.3"
 enum-as-inner = "0.5.1"
+fastcdc = "1.0.7"
 futures = "0.3.25"
 hex = "0.4.3"
 humantime = "2.1.0"
@@ -92,3 +94,6 @@ features = [
 	"rt-multi-thread",
 	"sync",
 ]
+
+[dev-dependencies]
+tokio-test = "0.4.2"
diff --git a/server/src/api/binary_cache.rs b/server/src/api/binary_cache.rs
index ec4bd0a..90b7afd 100644
--- a/server/src/api/binary_cache.rs
+++ b/server/src/api/binary_cache.rs
@@ -4,7 +4,10 @@
 //!
 //! The implementation is based on the specifications at <https://github.com/fzakaria/nix-http-binary-cache-api-spec>.
 
+use std::collections::VecDeque;
+use std::io::{Error as IoError, ErrorKind as IoErrorKind};
 use std::path::PathBuf;
+use std::sync::Arc;
 
 use axum::{
     body::StreamBody,
@@ -14,19 +17,22 @@ use axum::{
     routing::get,
     Router,
 };
+use futures::stream::BoxStream;
 use serde::Serialize;
 use tokio_util::io::ReaderStream;
 use tracing::instrument;
 
+use crate::database::entity::chunk::ChunkModel;
 use crate::database::AtticDatabase;
 use crate::error::{ErrorKind, ServerResult};
 use crate::narinfo::NarInfo;
 use crate::nix_manifest;
-use crate::storage::Download;
+use crate::storage::{Download, StorageBackend};
 use crate::{RequestState, State};
 use attic::cache::CacheName;
 use attic::mime;
 use attic::nix_store::StorePathHash;
+use attic::stream::merge_chunks;
 
 /// Nix cache information.
 ///
@@ -128,10 +134,10 @@ async fn get_store_path_info(
         cache_name
     );
 
-    let (object, cache, nar) = state
+    let (object, cache, nar, chunks) = state
         .database()
         .await?
-        .find_object_by_store_path_hash(&cache_name, &store_path_hash)
+        .find_object_and_chunks_by_store_path_hash(&cache_name, &store_path_hash)
         .await?;
 
     let permission = req_state
@@ -141,6 +147,11 @@ async fn get_store_path_info(
 
     req_state.set_public_cache(cache.is_public);
 
+    if chunks.iter().any(Option::is_none) {
+        // at least one of the chunks is missing :(
+        return Err(ErrorKind::IncompleteNar.into());
+    }
+
     let mut narinfo = object.to_nar_info(&nar)?;
 
     if narinfo.signature().is_none() {
@@ -184,8 +195,8 @@ async fn get_nar(
 
     let database = state.database().await?;
 
-    let (object, cache, nar) = database
-        .find_object_by_store_path_hash(&cache_name, &store_path_hash)
+    let (object, cache, _nar, chunks) = database
+        .find_object_and_chunks_by_store_path_hash(&cache_name, &store_path_hash)
         .await?;
 
     let permission = req_state
@@ -195,18 +206,62 @@ async fn get_nar(
 
     req_state.set_public_cache(cache.is_public);
 
+    if chunks.iter().any(Option::is_none) {
+        // at least one of the chunks is missing :(
+        return Err(ErrorKind::IncompleteNar.into());
+    }
+
     database.bump_object_last_accessed(object.id).await?;
 
-    let remote_file = nar.remote_file.0;
-    let backend = state.storage().await?;
-    match backend.download_file_db(&remote_file).await? {
-        Download::Redirect(uri) => Ok(Redirect::temporary(&uri).into_response()),
-        Download::Stream(stream) => {
-            let stream = ReaderStream::new(stream);
-            let body = StreamBody::new(stream);
-
-            Ok(body.into_response())
+    if chunks.len() == 1 {
+        // single chunk
+        let chunk = chunks[0].as_ref().unwrap();
+        let remote_file = &chunk.remote_file.0;
+        let storage = state.storage().await?;
+        match storage.download_file_db(remote_file, false).await? {
+            Download::Url(url) => Ok(Redirect::temporary(&url).into_response()),
+            Download::Stream(stream) => {
+                let body = StreamBody::new(stream);
+                Ok(body.into_response())
+            }
+            Download::AsyncRead(stream) => {
+                let stream = ReaderStream::new(stream);
+                let body = StreamBody::new(stream);
+                Ok(body.into_response())
+            }
         }
+    } else {
+        // reassemble NAR
+        fn io_error<E: std::error::Error + Send + Sync + 'static>(e: E) -> IoError {
+            IoError::new(IoErrorKind::Other, e)
+        }
+
+        let streamer = |chunk: ChunkModel, storage: Arc<Box<dyn StorageBackend + 'static>>| async move {
+            match storage
+                .download_file_db(&chunk.remote_file.0, true)
+                .await
+                .map_err(io_error)?
+            {
+                Download::Url(_) => Err(IoError::new(
+                    IoErrorKind::Other,
+                    "URLs not supported for NAR reassembly",
+                )),
+                Download::Stream(stream) => Ok(stream),
+                Download::AsyncRead(stream) => {
+                    let stream: BoxStream<_> = Box::pin(ReaderStream::new(stream));
+                    Ok(stream)
+                }
+            }
+        };
+
+        let chunks: VecDeque<_> = chunks.into_iter().map(Option::unwrap).collect();
+        let storage = state.storage().await?.clone();
+
+        // TODO: Make num_prefetch configurable
+        // The ideal size depends on the average chunk size
+        let merged = merge_chunks(chunks, streamer, storage, 2);
+        let body = StreamBody::new(merged);
+        Ok(body.into_response())
     }
 }
 
diff --git a/server/src/api/v1/get_missing_paths.rs b/server/src/api/v1/get_missing_paths.rs
index ce8e427..23b4530 100644
--- a/server/src/api/v1/get_missing_paths.rs
+++ b/server/src/api/v1/get_missing_paths.rs
@@ -6,6 +6,7 @@ use sea_orm::{FromQueryResult, QuerySelect};
 use tracing::instrument;
 
 use crate::database::entity::cache;
+use crate::database::entity::nar;
 use crate::database::entity::object::{self, Entity as Object};
 use crate::error::{ServerError, ServerResult};
 use crate::{RequestState, State};
@@ -48,8 +49,10 @@ pub(crate) async fn get_missing_paths(
         .select_only()
         .column_as(object::Column::StorePathHash, "store_path_hash")
         .join(sea_orm::JoinType::InnerJoin, object::Relation::Cache.def())
+        .join(sea_orm::JoinType::InnerJoin, object::Relation::Nar.def())
         .filter(cache::Column::Name.eq(payload.cache.as_str()))
         .filter(object::Column::StorePathHash.is_in(query_in))
+        .filter(nar::Column::CompletenessHint.eq(true))
         .into_model::<StorePathHashOnly>()
         .all(database)
         .await
diff --git a/server/src/api/v1/upload_path.rs b/server/src/api/v1/upload_path.rs
index ac284bd..f3e1aa0 100644
--- a/server/src/api/v1/upload_path.rs
+++ b/server/src/api/v1/upload_path.rs
@@ -1,23 +1,29 @@
 use std::io;
 
+use std::io::Cursor;
 use std::marker::Unpin;
 use std::sync::Arc;
 
 use anyhow::anyhow;
 use async_compression::tokio::bufread::{BrotliEncoder, XzEncoder, ZstdEncoder};
+use async_compression::Level as CompressionLevel;
 use axum::{
     extract::{BodyStream, Extension, Json},
     http::HeaderMap,
 };
+use bytes::Bytes;
 use chrono::Utc;
 use digest::Output as DigestOutput;
+use futures::future::join_all;
 use futures::StreamExt;
 use sea_orm::entity::prelude::*;
+use sea_orm::sea_query::Expr;
 use sea_orm::ActiveValue::Set;
-use sea_orm::TransactionTrait;
+use sea_orm::{QuerySelect, TransactionTrait};
 use sha2::{Digest, Sha256};
-use tokio::io::{AsyncRead, BufReader};
-use tokio::sync::OnceCell;
+use tokio::io::{AsyncBufRead, AsyncRead, BufReader};
+use tokio::sync::{OnceCell, Semaphore};
+use tokio::task::spawn;
 use tokio_util::io::StreamReader;
 use tracing::instrument;
 use uuid::Uuid;
@@ -31,14 +37,31 @@ use attic::hash::Hash;
 use attic::stream::StreamHasher;
 use attic::util::Finally;
 
+use crate::chunking::chunk_stream;
 use crate::database::entity::cache;
+use crate::database::entity::chunk::{self, ChunkState, Entity as Chunk};
+use crate::database::entity::chunkref::{self, Entity as ChunkRef};
 use crate::database::entity::nar::{self, Entity as Nar, NarState};
 use crate::database::entity::object::{self, Entity as Object};
 use crate::database::entity::Json as DbJson;
-use crate::database::{AtticDatabase, NarGuard};
+use crate::database::{AtticDatabase, ChunkGuard, NarGuard};
+
+/// Number of chunks to upload to the storage backend at once.
+///
+/// TODO: Make this configurable
+const CONCURRENT_CHUNK_UPLOADS: usize = 10;
 
 type CompressorFn<C> = Box<dyn FnOnce(C) -> Box<dyn AsyncRead + Unpin + Send> + Send>;
 
+/// Data of a chunk.
+enum ChunkData {
+    /// Some bytes in memory.
+    Bytes(Bytes),
+
+    /// A stream with a user-claimed hash and size that are potentially incorrect.
+    Stream(Box<dyn AsyncRead + Send + Unpin + 'static>, Hash, usize),
+}
+
 /// Applies compression to a stream, computing hashes along the way.
 ///
 /// Our strategy is to stream directly onto a UUID-keyed file on the
@@ -112,18 +135,33 @@ pub(crate) async fn upload_path(
     // Try to acquire a lock on an existing NAR
     let existing_nar = database.find_and_lock_nar(&upload_info.nar_hash).await?;
     match existing_nar {
+        // FIXME: existing NAR may be missing chunks
         Some(existing_nar) => {
-            // Deduplicate
-            upload_path_dedup(
-                username,
-                cache,
-                upload_info,
-                stream,
-                database,
-                &state,
-                existing_nar,
-            )
-            .await
+            // Deduplicate?
+            let missing_chunk = ChunkRef::find()
+                .filter(chunkref::Column::NarId.eq(existing_nar.id))
+                .filter(chunkref::Column::ChunkId.is_null())
+                .limit(1)
+                .one(database)
+                .await
+                .map_err(ServerError::database_error)?;
+
+            if missing_chunk.is_some() {
+                // Need to repair
+                upload_path_new(username, cache, upload_info, stream, database, &state).await
+            } else {
+                // Can actually be deduplicated
+                upload_path_dedup(
+                    username,
+                    cache,
+                    upload_info,
+                    stream,
+                    database,
+                    &state,
+                    existing_nar,
+                )
+                .await
+            }
         }
         None => {
             // New NAR
@@ -161,10 +199,6 @@ async fn upload_path_dedup(
         }
     }
 
-    let file_size = existing_nar.file_size
-        .map(|dbs| dbs.try_into().map_err(ServerError::database_error))
-        .transpose()?;
-
     // Finally...
     let txn = database
         .begin()
@@ -197,7 +231,7 @@ async fn upload_path_dedup(
 
     Ok(Json(UploadPathResult {
         kind: UploadPathResultKind::Deduplicated,
-        file_size,
+        file_size: None, // TODO: Sum the chunks
     }))
 }
 
@@ -214,105 +248,232 @@ async fn upload_path_new(
     database: &DatabaseConnection,
     state: &State,
 ) -> ServerResult<Json<UploadPathResult>> {
+    let nar_size_threshold = state.config.chunking.nar_size_threshold;
+
+    if nar_size_threshold == 0 || upload_info.nar_size < nar_size_threshold {
+        upload_path_new_unchunked(username, cache, upload_info, stream, database, state).await
+    } else {
+        upload_path_new_chunked(username, cache, upload_info, stream, database, state).await
+    }
+}
+
+/// Uploads a path when there is no matching NAR in the global cache (chunked).
+async fn upload_path_new_chunked(
+    username: Option<String>,
+    cache: cache::Model,
+    upload_info: UploadPathNarInfo,
+    stream: impl AsyncRead + Send + Unpin + 'static,
+    database: &DatabaseConnection,
+    state: &State,
+) -> ServerResult<Json<UploadPathResult>> {
+    let chunking_config = &state.config.chunking;
     let compression_config = &state.config.compression;
-    let compression: Compression = compression_config.r#type.into();
-    let level = compression_config.level();
-    let compressor: CompressorFn<_> = match compression_config.r#type {
-        CompressionType::None => Box::new(|c| Box::new(c)),
-        CompressionType::Brotli => {
-            Box::new(move |s| Box::new(BrotliEncoder::with_quality(s, level)))
-        }
-        CompressionType::Zstd => Box::new(move |s| Box::new(ZstdEncoder::with_quality(s, level))),
-        CompressionType::Xz => Box::new(move |s| Box::new(XzEncoder::with_quality(s, level))),
-    };
+    let compression_type = compression_config.r#type;
+    let compression_level = compression_config.level();
+    let compression: Compression = compression_type.into();
 
-    let backend = state.storage().await?;
+    let nar_size_db = i64::try_from(upload_info.nar_size).map_err(ServerError::request_error)?;
 
-    let key = format!("{}.nar", Uuid::new_v4());
+    // FIXME: Maybe the client will send much more data than claimed
+    let (stream, nar_compute) = StreamHasher::new(stream, Sha256::new());
+    let mut chunks = chunk_stream(
+        stream,
+        chunking_config.min_size,
+        chunking_config.avg_size,
+        chunking_config.max_size,
+    );
 
-    let remote_file = backend.make_db_reference(key.clone()).await?;
-    let remote_file_id = remote_file.remote_file_id();
-    let nar_id = {
-        let nar_size_db =
-            i64::try_from(upload_info.nar_size).map_err(ServerError::request_error)?;
-        let model = nar::ActiveModel {
-            state: Set(NarState::PendingUpload),
-            compression: Set(compression.to_string()),
+    let upload_chunk_limit = Arc::new(Semaphore::new(CONCURRENT_CHUNK_UPLOADS));
+    let mut futures = Vec::new();
 
-            // Untrusted data - To be confirmed later
-            nar_hash: Set(upload_info.nar_hash.to_typed_base16()),
-            nar_size: Set(nar_size_db),
+    while let Some(bytes) = chunks.next().await {
+        let bytes = bytes.map_err(ServerError::request_error)?;
+        let data = ChunkData::Bytes(bytes);
 
-            remote_file: Set(DbJson(remote_file)),
-            remote_file_id: Set(remote_file_id),
+        // Wait for a permit before spawning
+        //
+        // We want to block the receive process as well, otherwise it stays ahead and
+        // consumes too much memory
+        let permit = upload_chunk_limit.clone().acquire_owned().await.unwrap();
+        futures.push({
+            let database = database.clone();
+            let state = state.clone();
+            let require_proof_of_possession = state.config.require_proof_of_possession;
 
-            created_at: Set(Utc::now()),
-            ..Default::default()
-        };
-
-        let insertion = Nar::insert(model)
-            .exec(database)
-            .await
-            .map_err(ServerError::database_error)?;
-
-        insertion.last_insert_id
-    };
-
-    let cleanup = Finally::new({
-        let database = database.clone();
-        let nar_model = nar::ActiveModel {
-            id: Set(nar_id),
-            ..Default::default()
-        };
-        let backend = backend.clone();
-        let key = key.clone();
-
-        async move {
-            tracing::warn!("Error occurred - Cleaning up uploaded file and NAR entry");
-
-            if let Err(e) = backend.delete_file(key).await {
-                tracing::warn!("Failed to clean up failed upload: {}", e);
-            }
-
-            if let Err(e) = Nar::delete(nar_model).exec(&database).await {
-                tracing::warn!("Failed to unregister failed NAR: {}", e);
-            }
-        }
-    });
-
-    let mut stream = CompressionStream::new(stream, compressor);
-
-    // Stream the object to the storage backend
-    backend
-        .upload_file(key, stream.stream())
-        .await
-        .map_err(ServerError::storage_error)?;
+            spawn(async move {
+                let chunk = upload_chunk(
+                    data,
+                    compression_type,
+                    compression_level,
+                    database,
+                    state,
+                    require_proof_of_possession,
+                )
+                .await?;
+                drop(permit);
+                Ok(chunk)
+            })
+        });
+    }
 
     // Confirm that the NAR Hash and Size are correct
     // FIXME: errors
-    let (nar_hash, nar_size) = stream.nar_hash_and_size().unwrap();
-    let (file_hash, file_size) = stream.file_hash_and_size().unwrap();
-
+    let (nar_hash, nar_size) = nar_compute.get().unwrap();
     let nar_hash = Hash::Sha256(nar_hash.as_slice().try_into().unwrap());
-    let file_hash = Hash::Sha256(file_hash.as_slice().try_into().unwrap());
 
     if nar_hash != upload_info.nar_hash || *nar_size != upload_info.nar_size {
         return Err(ErrorKind::RequestError(anyhow!("Bad NAR Hash or Size")).into());
     }
 
+    // Wait for all uploads to complete
+    let chunks: Vec<ChunkGuard> = join_all(futures)
+        .await
+        .into_iter()
+        .map(|join_result| join_result.unwrap())
+        .collect::<ServerResult<Vec<_>>>()?;
+
+    let file_size = chunks
+        .iter()
+        .fold(0, |acc, c| acc + c.file_size.unwrap() as usize);
+
     // Finally...
     let txn = database
         .begin()
         .await
         .map_err(ServerError::database_error)?;
 
-    // Update the file hash and size, and set the nar to valid
-    let file_size_db = i64::try_from(*file_size).map_err(ServerError::request_error)?;
-    Nar::update(nar::ActiveModel {
-        id: Set(nar_id),
-        state: Set(NarState::Valid),
-        file_hash: Set(Some(file_hash.to_typed_base16())),
-        file_size: Set(Some(file_size_db)),
+    // Create a NAR entry
+    let nar_id = {
+        let model = nar::ActiveModel {
+            state: Set(NarState::Valid),
+            compression: Set(compression.to_string()),
+
+            nar_hash: Set(upload_info.nar_hash.to_typed_base16()),
+            nar_size: Set(nar_size_db),
+
+            num_chunks: Set(chunks.len() as i32),
+
+            created_at: Set(Utc::now()),
+            ..Default::default()
+        };
+
+        let insertion = Nar::insert(model)
+            .exec(&txn)
+            .await
+            .map_err(ServerError::database_error)?;
+
+        insertion.last_insert_id
+    };
+
+    // Create mappings from the NAR to the chunks
+    for (i, chunk) in chunks.iter().enumerate() {
+        ChunkRef::insert(chunkref::ActiveModel {
+            nar_id: Set(nar_id),
+            seq: Set(i as i32),
+            chunk_id: Set(Some(chunk.id)),
+            chunk_hash: Set(chunk.chunk_hash.clone()),
+            compression: Set(chunk.compression.clone()),
+            ..Default::default()
+        })
+        .exec(&txn)
+        .await
+        .map_err(ServerError::database_error)?;
+    }
+
+    // Create a mapping granting the local cache access to the NAR
+    Object::delete_many()
+        .filter(object::Column::CacheId.eq(cache.id))
+        .filter(object::Column::StorePathHash.eq(upload_info.store_path_hash.to_string()))
+        .exec(&txn)
+        .await
+        .map_err(ServerError::database_error)?;
+    Object::insert({
+        let mut new_object = upload_info.to_active_model();
+        new_object.cache_id = Set(cache.id);
+        new_object.nar_id = Set(nar_id);
+        new_object.created_at = Set(Utc::now());
+        new_object.created_by = Set(username);
+        new_object
+    })
+    .exec(&txn)
+    .await
+    .map_err(ServerError::database_error)?;
+
+    txn.commit().await.map_err(ServerError::database_error)?;
+
+    Ok(Json(UploadPathResult {
+        kind: UploadPathResultKind::Uploaded,
+        file_size: Some(file_size),
+    }))
+}
+
+/// Uploads a path when there is no matching NAR in the global cache (unchunked).
+///
+/// We upload the entire NAR as a single chunk.
+async fn upload_path_new_unchunked(
+    username: Option<String>,
+    cache: cache::Model,
+    upload_info: UploadPathNarInfo,
+    stream: impl AsyncRead + Send + Unpin + 'static,
+    database: &DatabaseConnection,
+    state: &State,
+) -> ServerResult<Json<UploadPathResult>> {
+    let compression_config = &state.config.compression;
+    let compression_type = compression_config.r#type;
+    let compression: Compression = compression_type.into();
+
+    // Upload the entire NAR as a single chunk
+    let data = ChunkData::Stream(
+        Box::new(stream),
+        upload_info.nar_hash.clone(),
+        upload_info.nar_size,
+    );
+    let chunk = upload_chunk(
+        data,
+        compression_type,
+        compression_config.level(),
+        database.clone(),
+        state.clone(),
+        state.config.require_proof_of_possession,
+    )
+    .await?;
+    let file_size = chunk.file_size.unwrap() as usize;
+
+    // Finally...
+    let txn = database
+        .begin()
+        .await
+        .map_err(ServerError::database_error)?;
+
+    // Create a NAR entry
+    let nar_id = {
+        let model = nar::ActiveModel {
+            state: Set(NarState::Valid),
+            compression: Set(compression.to_string()),
+
+            nar_hash: Set(upload_info.nar_hash.to_typed_base16()),
+            nar_size: Set(chunk.chunk_size),
+
+            num_chunks: Set(1),
+
+            created_at: Set(Utc::now()),
+            ..Default::default()
+        };
+
+        let insertion = Nar::insert(model)
+            .exec(&txn)
+            .await
+            .map_err(ServerError::database_error)?;
+
+        insertion.last_insert_id
+    };
+
+    // Create a mapping from the NAR to the chunk
+    ChunkRef::insert(chunkref::ActiveModel {
+        nar_id: Set(nar_id),
+        seq: Set(0),
+        chunk_id: Set(Some(chunk.id)),
+        chunk_hash: Set(upload_info.nar_hash.to_typed_base16()),
         ..Default::default()
     })
     .exec(&txn)
@@ -340,15 +501,225 @@ async fn upload_path_new(
 
     txn.commit().await.map_err(ServerError::database_error)?;
 
-    cleanup.cancel();
-
     Ok(Json(UploadPathResult {
         kind: UploadPathResultKind::Uploaded,
-        file_size: Some(*file_size),
+        file_size: Some(file_size),
     }))
 }
 
+/// Uploads a chunk with the desired compression.
+///
+/// This will automatically perform deduplication if the chunk exists.
+async fn upload_chunk(
+    data: ChunkData,
+    compression_type: CompressionType,
+    compression_level: CompressionLevel,
+    database: DatabaseConnection,
+    state: State,
+    require_proof_of_possession: bool,
+) -> ServerResult<ChunkGuard> {
+    let compression: Compression = compression_type.into();
+
+    let given_chunk_hash = data.hash();
+    let given_chunk_size = data.size();
+
+    if let Some(existing_chunk) = database
+        .find_and_lock_chunk(&given_chunk_hash, compression)
+        .await?
+    {
+        // There's an existing chunk matching the hash
+        if require_proof_of_possession && !data.is_hash_trusted() {
+            let stream = data.into_async_read();
+
+            let (mut stream, nar_compute) = StreamHasher::new(stream, Sha256::new());
+            tokio::io::copy(&mut stream, &mut tokio::io::sink())
+                .await
+                .map_err(ServerError::request_error)?;
+
+            // FIXME: errors
+            let (nar_hash, nar_size) = nar_compute.get().unwrap();
+            let nar_hash = Hash::Sha256(nar_hash.as_slice().try_into().unwrap());
+
+            // Confirm that the NAR Hash and Size are correct
+            if nar_hash.to_typed_base16() != existing_chunk.chunk_hash
+                || *nar_size != given_chunk_size
+                || *nar_size != existing_chunk.chunk_size as usize
+            {
+                return Err(ErrorKind::RequestError(anyhow!("Bad chunk hash or size")).into());
+            }
+        }
+
+        return Ok(existing_chunk);
+    }
+
+    let key = format!("{}.chunk", Uuid::new_v4());
+
+    let backend = state.storage().await?;
+    let remote_file = backend.make_db_reference(key.clone()).await?;
+    let remote_file_id = remote_file.remote_file_id();
+
+    let chunk_size_db = i64::try_from(given_chunk_size).map_err(ServerError::request_error)?;
+
+    let chunk_id = {
+        let model = chunk::ActiveModel {
+            state: Set(ChunkState::PendingUpload),
+            compression: Set(compression.to_string()),
+
+            // Untrusted data - To be confirmed later
+            chunk_hash: Set(given_chunk_hash.to_typed_base16()),
+            chunk_size: Set(chunk_size_db),
+
+            remote_file: Set(DbJson(remote_file)),
+            remote_file_id: Set(remote_file_id),
+
+            created_at: Set(Utc::now()),
+            ..Default::default()
+        };
+
+        let insertion = Chunk::insert(model)
+            .exec(&database)
+            .await
+            .map_err(ServerError::database_error)?;
+
+        insertion.last_insert_id
+    };
+
+    let cleanup = Finally::new({
+        let database = database.clone();
+        let chunk_model = chunk::ActiveModel {
+            id: Set(chunk_id),
+            ..Default::default()
+        };
+        let backend = backend.clone();
+        let key = key.clone();
+
+        async move {
+            tracing::warn!("Error occurred - Cleaning up uploaded file and chunk entry");
+
+            if let Err(e) = backend.delete_file(key).await {
+                tracing::warn!("Failed to clean up failed upload: {}", e);
+            }
+
+            if let Err(e) = Chunk::delete(chunk_model).exec(&database).await {
+                tracing::warn!("Failed to unregister failed chunk: {}", e);
+            }
+        }
+    });
+
+    // Compress and stream to the storage backend
+    let compressor = get_compressor_fn(compression_type, compression_level);
+    let mut stream = CompressionStream::new(data.into_async_read(), compressor);
+
+    backend
+        .upload_file(key, stream.stream())
+        .await
+        .map_err(ServerError::storage_error)?;
+
+    // Confirm that the chunk hash is correct
+    let (chunk_hash, chunk_size) = stream.nar_hash_and_size().unwrap();
+    let (file_hash, file_size) = stream.file_hash_and_size().unwrap();
+
+    let chunk_hash = Hash::Sha256(chunk_hash.as_slice().try_into().unwrap());
+    let file_hash = Hash::Sha256(file_hash.as_slice().try_into().unwrap());
+
+    if chunk_hash != given_chunk_hash || *chunk_size != given_chunk_size {
+        return Err(ErrorKind::RequestError(anyhow!("Bad chunk hash or size")).into());
+    }
+
+    // Finally...
+    let txn = database
+        .begin()
+        .await
+        .map_err(ServerError::database_error)?;
+
+    // Update the file hash and size, and set the chunk to valid
+    let file_size_db = i64::try_from(*file_size).map_err(ServerError::request_error)?;
+    let chunk = Chunk::update(chunk::ActiveModel {
+        id: Set(chunk_id),
+        state: Set(ChunkState::Valid),
+        file_hash: Set(Some(file_hash.to_typed_base16())),
+        file_size: Set(Some(file_size_db)),
+        holders_count: Set(1),
+        ..Default::default()
+    })
+    .exec(&txn)
+    .await
+    .map_err(ServerError::database_error)?;
+
+    // Also repair broken chunk references pointing at the same chunk
+    let repaired = ChunkRef::update_many()
+        .col_expr(chunkref::Column::ChunkId, Expr::value(chunk_id))
+        .filter(chunkref::Column::ChunkId.is_null())
+        .filter(chunkref::Column::ChunkHash.eq(chunk_hash.to_typed_base16()))
+        .filter(chunkref::Column::Compression.eq(compression.to_string()))
+        .exec(&txn)
+        .await
+        .map_err(ServerError::database_error)?;
+
+    txn.commit().await.map_err(ServerError::database_error)?;
+
+    cleanup.cancel();
+
+    tracing::debug!("Repaired {} chunkrefs", repaired.rows_affected);
+
+    let guard = ChunkGuard::from_locked(database.clone(), chunk);
+
+    Ok(guard)
+}
+
+/// Returns a compressor function that takes some stream as input.
+fn get_compressor_fn<C: AsyncBufRead + Unpin + Send + 'static>(
+    ctype: CompressionType,
+    level: CompressionLevel,
+) -> CompressorFn<C> {
+    match ctype {
+        CompressionType::None => Box::new(|c| Box::new(c)),
+        CompressionType::Brotli => {
+            Box::new(move |s| Box::new(BrotliEncoder::with_quality(s, level)))
+        }
+        CompressionType::Zstd => Box::new(move |s| Box::new(ZstdEncoder::with_quality(s, level))),
+        CompressionType::Xz => Box::new(move |s| Box::new(XzEncoder::with_quality(s, level))),
+    }
+}
+
+impl ChunkData {
+    /// Returns the potentially-incorrect hash of the chunk.
+    fn hash(&self) -> Hash {
+        match self {
+            Self::Bytes(bytes) => {
+                let mut hasher = Sha256::new();
+                hasher.update(bytes);
+                let hash = hasher.finalize();
+                Hash::Sha256(hash.as_slice().try_into().unwrap())
+            }
+            Self::Stream(_, hash, _) => hash.clone(),
+        }
+    }
+
+    /// Returns the potentially-incorrect size of the chunk.
+    fn size(&self) -> usize {
+        match self {
+            Self::Bytes(bytes) => bytes.len(),
+            Self::Stream(_, _, size) => *size,
+        }
+    }
+
+    /// Returns whether the hash is trusted.
+    fn is_hash_trusted(&self) -> bool {
+        matches!(self, ChunkData::Bytes(_))
+    }
+
+    /// Turns the data into a stream.
+    fn into_async_read(self) -> Box<dyn AsyncRead + Unpin + Send> {
+        match self {
+            Self::Bytes(bytes) => Box::new(Cursor::new(bytes)),
+            Self::Stream(stream, _, _) => stream,
+        }
+    }
+}
+
 impl CompressionStream {
+    /// Creates a new compression stream.
     fn new<R>(stream: R, compressor: CompressorFn<BufReader<StreamHasher<R, Sha256>>>) -> Self
     where
         R: AsyncRead + Unpin + Send + 'static,
@@ -369,6 +740,29 @@ impl CompressionStream {
         }
     }
 
+    /*
+    /// Creates a compression stream without compute the uncompressed hash/size.
+    ///
+    /// This is useful if you already know the hash. `nar_hash_and_size` will
+    /// always return `None`.
+    fn new_without_nar_hash<R>(stream: R, compressor: CompressorFn<BufReader<R>>) -> Self
+    where
+        R: AsyncRead + Unpin + Send + 'static,
+    {
+        // compress NAR
+        let stream = compressor(BufReader::new(stream));
+
+        // compute file hash and size
+        let (stream, file_compute) = StreamHasher::new(stream, Sha256::new());
+
+        Self {
+            stream: Box::new(stream),
+            nar_compute: Arc::new(OnceCell::new()),
+            file_compute,
+        }
+    }
+    */
+
     /// Returns the stream of the compressed object.
     fn stream(&mut self) -> &mut (impl AsyncRead + Unpin) {
         &mut self.stream
diff --git a/server/src/chunking/mod.rs b/server/src/chunking/mod.rs
new file mode 100644
index 0000000..4896046
--- /dev/null
+++ b/server/src/chunking/mod.rs
@@ -0,0 +1,123 @@
+//! Chunking.
+//!
+//! We perform chunking on uncompressed NARs using the FastCDC
+//! algorithm.
+
+use async_stream::try_stream;
+use bytes::{BufMut, Bytes, BytesMut};
+use fastcdc::FastCDC;
+use futures::stream::Stream;
+use tokio::io::{AsyncRead, AsyncReadExt};
+
+/// Splits a streams into content-defined chunks.
+///
+/// This is a wrapper over fastcdc-rs that takes an `AsyncRead` and
+/// returns a `Stream` of chunks as `Bytes`s.
+pub fn chunk_stream<R>(
+    mut stream: R,
+    min_size: usize,
+    avg_size: usize,
+    max_size: usize,
+) -> impl Stream<Item = std::io::Result<Bytes>>
+where
+    R: AsyncRead + Unpin + Send,
+{
+    let s = try_stream! {
+        let mut buf = BytesMut::with_capacity(max_size);
+
+        loop {
+            let read = read_chunk_async(&mut stream, buf).await?;
+
+            let mut eof = false;
+            if read.len() == 0 {
+                // Already EOF
+                break;
+            } else if read.len() < max_size {
+                // Last read
+                eof = true;
+            }
+
+            let chunks = FastCDC::with_eof(&read, min_size, avg_size, max_size, eof);
+            let mut consumed = 0;
+
+            for chunk in chunks {
+                consumed += chunk.length;
+
+                let slice = read.slice(chunk.offset..chunk.offset + chunk.length);
+                yield slice;
+            }
+
+            if eof {
+                break;
+            }
+
+            buf = BytesMut::with_capacity(max_size);
+
+            if consumed < read.len() {
+                // remaining bytes for the next read
+                buf.put_slice(&read[consumed..]);
+            }
+        }
+    };
+
+    Box::pin(s)
+}
+
+async fn read_chunk_async<S: AsyncRead + Unpin + Send>(
+    stream: &mut S,
+    mut chunk: BytesMut,
+) -> std::io::Result<Bytes> {
+    while chunk.len() < chunk.capacity() {
+        let read = stream.read_buf(&mut chunk).await?;
+
+        if read == 0 {
+            break;
+        }
+    }
+
+    Ok(chunk.freeze())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use std::io::Cursor;
+
+    use futures::StreamExt;
+    use tokio_test::block_on;
+
+    /// Chunks and reconstructs a file.
+    #[test]
+    fn test_chunking_basic() {
+        block_on(async move {
+            let test_file = get_data(32 * 1024 * 1024); // 32 MiB
+            let mut reconstructed_file = Vec::new();
+
+            let cursor = Cursor::new(&test_file);
+            let mut chunks = chunk_stream(cursor, 8 * 1024, 16 * 1024, 32 * 1024);
+
+            while let Some(chunk) = chunks.next().await {
+                let chunk = chunk.unwrap();
+                eprintln!("Got a {}-byte chunk", chunk.len());
+                reconstructed_file.extend(chunk);
+            }
+
+            assert_eq!(reconstructed_file, test_file);
+        });
+    }
+
+    /// Returns some fake data.
+    fn get_data(len: usize) -> Vec<u8> {
+        let mut state = 42u32;
+        let mut data = vec![0u8; len];
+
+        for i in 0..data.len() {
+            (state, _) = state.overflowing_mul(1664525u32);
+            (state, _) = state.overflowing_add(1013904223u32);
+            data[i] = ((state >> (i % 24)) & 0xff) as u8;
+        }
+
+        data
+    }
+}
diff --git a/server/src/config-template.toml b/server/src/config-template.toml
index 1add4c8..84127d2 100644
--- a/server/src/config-template.toml
+++ b/server/src/config-template.toml
@@ -86,6 +86,28 @@ path = "%storage_path%"
 #  access_key_id = ""
 #  secret_access_key = ""
  
+# Data chunking
+#
+# Warning: If you change any of the values here, it will be
+# difficult to reuse existing chunks for newly-uploaded NARs
+# since the cutpoints will be different. As a result, the
+# deduplication ratio will suffer for a while after the change.
+[chunking]
+# The minimum NAR size to trigger chunking
+#
+# If 0, chunking is disabled entirely for newly-uploaded NARs.
+# If 1, all NARs are chunked.
+nar-size-threshold = 65536 # chunk files that are 64 KiB or larger
+
+# The preferred minimum size of a chunk, in bytes
+min-size = 16384            # 16 KiB
+
+# The preferred average size of a chunk, in bytes
+avg-size = 65536            # 64 KiB
+
+# The preferred maximum size of a chunk, in bytes
+max-size = 262144           # 256 KiB
+
 # Compression
 [compression]
 # Compression type
diff --git a/server/src/config.rs b/server/src/config.rs
index d77e811..3de159f 100644
--- a/server/src/config.rs
+++ b/server/src/config.rs
@@ -92,6 +92,9 @@ pub struct Config {
     /// Storage.
     pub storage: StorageConfig,
 
+    /// Data chunking.
+    pub chunking: ChunkingConfig,
+
     /// Compression.
     #[serde(default = "Default::default")]
     pub compression: CompressionConfig,
@@ -137,6 +140,46 @@ pub enum StorageConfig {
     S3(S3StorageConfig),
 }
 
+/// Data chunking.
+///
+/// This must be set, but a default set of values is provided
+/// through the OOBE sequence. The reason is that this allows
+/// us to provide a new set of recommended "defaults" for newer
+/// deployments without affecting existing ones.
+///
+/// Warning: If you change any of the values here, it will be
+/// difficult to reuse existing chunks for newly-uploaded NARs
+/// since the cutpoints will be different. As a result, the
+/// deduplication ratio will suffer for a while after the change.
+///
+/// `atticadm test-chunking` provides a way to test chunking
+/// on a set of files so you can fine-tune the values.
+#[derive(Debug, Clone, Deserialize)]
+pub struct ChunkingConfig {
+    /// The minimum NAR size to trigger chunking.
+    ///
+    /// If 0, chunking is disabled entirely for newly-uploaded
+    /// NARs.
+    ///
+    /// If 1, all newly-uploaded NARs are chunked.
+    ///
+    /// By default, the threshold is 128KB.
+    #[serde(rename = "nar-size-threshold")]
+    pub nar_size_threshold: usize,
+
+    /// The preferred minimum size of a chunk, in bytes.
+    #[serde(rename = "min-size")]
+    pub min_size: usize,
+
+    /// The preferred average size of a chunk, in bytes.
+    #[serde(rename = "avg-size")]
+    pub avg_size: usize,
+
+    /// The preferred maximum size of a chunk, in bytes.
+    #[serde(rename = "max-size")]
+    pub max_size: usize,
+}
+
 /// Compression configuration.
 #[derive(Debug, Clone, Deserialize)]
 pub struct CompressionConfig {
@@ -294,7 +337,7 @@ fn load_config_from_str(s: &str) -> Result<Config> {
 /// Loads the configuration in the standard order.
 pub async fn load_config(config_path: Option<&Path>, allow_oobe: bool) -> Result<Config> {
     if let Some(config_path) = config_path {
-        load_config_from_path(&config_path)
+        load_config_from_path(config_path)
     } else if let Ok(config_env) = env::var(ENV_CONFIG_BASE64) {
         let decoded = String::from_utf8(base64::decode(config_env.as_bytes())?)?;
         load_config_from_str(&decoded)
diff --git a/server/src/database/entity/chunk.rs b/server/src/database/entity/chunk.rs
new file mode 100644
index 0000000..0c302ae
--- /dev/null
+++ b/server/src/database/entity/chunk.rs
@@ -0,0 +1,111 @@
+//! A content-addressed chunk in the global chunk store.
+
+use sea_orm::entity::prelude::*;
+
+use super::Json;
+use crate::storage::RemoteFile;
+
+pub type ChunkModel = Model;
+
+/// The state of a chunk.
+#[derive(EnumIter, DeriveActiveEnum, Debug, Clone, PartialEq, Eq)]
+#[sea_orm(rs_type = "String", db_type = "String(Some(1))")]
+pub enum ChunkState {
+    /// The chunk can be used.
+    ///
+    /// The raw and compressed hashes are available.
+    #[sea_orm(string_value = "V")]
+    Valid,
+
+    /// The chunk is a pending upload.
+    ///
+    /// The raw and compressed hashes may not be available.
+    #[sea_orm(string_value = "P")]
+    PendingUpload,
+
+    /// The chunk can be deleted because it already exists.
+    ///
+    /// This state can be transitioned into from `PendingUpload`
+    /// if some other client completes uploading the same chunk
+    /// faster.
+    #[sea_orm(string_value = "C")]
+    ConfirmedDeduplicated,
+
+    /// The chunk is being deleted.
+    ///
+    /// This row will be deleted shortly.
+    #[sea_orm(string_value = "D")]
+    Deleted,
+}
+
+/// A content-addressed chunk in the global cache.
+#[derive(Debug, Clone, PartialEq, Eq, DeriveEntityModel)]
+#[sea_orm(table_name = "chunk")]
+pub struct Model {
+    /// Unique numeric ID of the chunk.
+    #[sea_orm(primary_key)]
+    pub id: i64,
+
+    /// The state of the chunk.
+    state: ChunkState,
+
+    /// The hash of the uncompressed chunk.
+    ///
+    /// This always begins with "sha256:" with the hash in the
+    /// hexadecimal format.
+    ///
+    /// The global chunk store may have several chunks with the same
+    /// hash:
+    ///
+    /// - Racing uploads from different clients
+    /// - Different compression methods
+    #[sea_orm(indexed)]
+    pub chunk_hash: String,
+
+    /// The size of the uncompressed chunk.
+    pub chunk_size: i64,
+
+    /// The hash of the compressed chunk.
+    ///
+    /// This always begins with "sha256:" with the hash in the
+    /// hexadecimal format.
+    ///
+    /// This field may not be available if the file hashes aren't
+    /// confirmed.
+    pub file_hash: Option<String>,
+
+    /// The size of the compressed chunk.
+    ///
+    /// This field may not be available if the file hashes aren't
+    /// confirmed.
+    pub file_size: Option<i64>,
+
+    /// The type of compression in use.
+    #[sea_orm(column_type = "String(Some(10))")]
+    pub compression: String,
+
+    /// The remote file backing this chunk.
+    pub remote_file: Json<RemoteFile>,
+
+    /// Unique string identifying the remote file.
+    #[sea_orm(unique)]
+    pub remote_file_id: String,
+
+    /// Number of processes holding this chunk.
+    ///
+    /// This is for preventing garbage collection of chunks when
+    /// there is a pending upload that can be deduplicated and
+    /// there are no existing NAR references.
+    pub holders_count: i32,
+
+    /// Timestamp when the chunk is created.
+    pub created_at: ChronoDateTimeUtc,
+}
+
+#[derive(Copy, Clone, Debug, EnumIter, DeriveRelation)]
+pub enum Relation {
+    #[sea_orm(has_many = "super::chunkref::Entity")]
+    ChunkRef,
+}
+
+impl ActiveModelBehavior for ActiveModel {}
diff --git a/server/src/database/entity/chunkref.rs b/server/src/database/entity/chunkref.rs
new file mode 100644
index 0000000..e0c5a8c
--- /dev/null
+++ b/server/src/database/entity/chunkref.rs
@@ -0,0 +1,84 @@
+//! A reference binding a NAR and a chunk.
+//!
+//! A NAR is backed by a sequence of chunks.
+//!
+//! A chunk may become unavailable (e.g., disk corruption) and
+//! removed from the database, in which case all dependent NARs
+//! will become unavailable.
+//!
+//! Such scenario can be recovered from by reuploading any object
+//! that has the missing chunk. `atticadm` will have the functionality
+//! to kill/delete a corrupted chunk from the database and to find
+//! objects with missing chunks so they can be repaired.
+
+use sea_orm::entity::prelude::*;
+
+pub type ChunkRefModel = Model;
+
+/// A reference binding a NAR to a chunk.
+#[derive(Debug, Clone, PartialEq, Eq, DeriveEntityModel)]
+#[sea_orm(table_name = "chunkref")]
+pub struct Model {
+    /// Unique numeric ID of the link.
+    #[sea_orm(primary_key)]
+    pub id: i64,
+
+    /// ID of the NAR.
+    #[sea_orm(indexed)]
+    pub nar_id: i64,
+
+    /// The zero-indexed sequence number of the chunk.
+    pub seq: i32,
+
+    /// ID of the chunk.
+    ///
+    /// This may be NULL when the chunk is missing from the
+    /// database.
+    #[sea_orm(indexed)]
+    pub chunk_id: Option<i64>,
+
+    /// The hash of the uncompressed chunk.
+    ///
+    /// This always begins with "sha256:" with the hash in the
+    /// hexadecimal format.
+    ///
+    /// This is used for recovering from a missing chunk.
+    #[sea_orm(indexed)]
+    pub chunk_hash: String,
+
+    /// The compression of the compressed chunk.
+    ///
+    /// This is used for recovering from a missing chunk.
+    pub compression: String,
+}
+
+#[derive(Copy, Clone, Debug, EnumIter, DeriveRelation)]
+pub enum Relation {
+    #[sea_orm(
+        belongs_to = "super::chunk::Entity",
+        from = "Column::ChunkId",
+        to = "super::chunk::Column::Id"
+    )]
+    Chunk,
+
+    #[sea_orm(
+        belongs_to = "super::nar::Entity",
+        from = "Column::NarId",
+        to = "super::nar::Column::Id"
+    )]
+    Nar,
+}
+
+impl Related<super::chunk::Entity> for Entity {
+    fn to() -> RelationDef {
+        Relation::Chunk.def()
+    }
+}
+
+impl Related<super::nar::Entity> for Entity {
+    fn to() -> RelationDef {
+        Relation::Nar.def()
+    }
+}
+
+impl ActiveModelBehavior for ActiveModel {}
diff --git a/server/src/database/entity/mod.rs b/server/src/database/entity/mod.rs
index 6f6acb3..8a4957a 100644
--- a/server/src/database/entity/mod.rs
+++ b/server/src/database/entity/mod.rs
@@ -3,6 +3,8 @@
 //! We use SeaORM and target PostgreSQL (production) and SQLite (development).
 
 pub mod cache;
+pub mod chunk;
+pub mod chunkref;
 pub mod nar;
 pub mod object;
 
diff --git a/server/src/database/entity/nar.rs b/server/src/database/entity/nar.rs
index fe404ef..01f29bc 100644
--- a/server/src/database/entity/nar.rs
+++ b/server/src/database/entity/nar.rs
@@ -2,9 +2,6 @@
 
 use sea_orm::entity::prelude::*;
 
-use super::Json;
-use crate::storage::RemoteFile;
-
 pub type NarModel = Model;
 
 /// The state of a NAR.
@@ -35,6 +32,7 @@ pub enum NarState {
     /// The NAR is being deleted.
     ///
     /// This row will be deleted shortly.
+    /// This variant is no longer used since the actual storage is managed as chunks.
     #[sea_orm(string_value = "D")]
     Deleted,
 }
@@ -44,6 +42,27 @@ pub enum NarState {
 /// A NAR without `nix-store --export` metadata is context-free,
 /// meaning that it's not associated with a store path and only
 /// depends on its contents.
+///
+/// ## NAR Repair
+///
+/// After a NAR is transitioned into the `Valid` state, its list
+/// of constituent chunks in `chunkref` is immutable. When a client
+/// uploads an existing NAR and the NAR has unavailable chunks,
+/// a new `nar` entry is created and all dependent `object` rows
+/// will have the `nar_id` updated. The old `nar` entry will
+/// be garbage-collected.
+///
+/// Why don't we just fill in the missing chunks in the existing
+/// `nar`? Because the NAR stream from the client _might_ be chunked
+/// differently. This is not supposed to happen since FastCDC
+/// has a deterministic lookup table for cut-point judgment, however
+/// we want the system to tolerate different chunking behavior because
+/// of table changes, for example.
+///
+/// However, when a chunk is added, all broken `chunkref`s with
+/// the same `chunk_hash` _are_ repaired. In other words, by
+/// re-uploading a broken NAR you are helping other NARs with
+/// the same broken chunk.
 #[derive(Debug, Clone, PartialEq, Eq, DeriveEntityModel)]
 #[sea_orm(table_name = "nar")]
 pub struct Model {
@@ -70,31 +89,24 @@ pub struct Model {
     /// The size of the NAR archive.
     pub nar_size: i64,
 
-    /// The hash of the compressed file.
-    ///
-    /// This always begins with "sha256:" with the hash in the
-    /// hexadecimal format.
-    ///
-    /// This field may not be available if the file hashes aren't
-    /// confirmed.
-    pub file_hash: Option<String>,
-
-    /// The size of the compressed file.
-    ///
-    /// This field may not be available if the file hashes aren't
-    /// confirmed.
-    pub file_size: Option<i64>,
-
     /// The type of compression in use.
     #[sea_orm(column_type = "String(Some(10))")]
     pub compression: String,
 
-    /// The remote file backing this NAR.
-    pub remote_file: Json<RemoteFile>,
+    /// Number of chunks that make up this NAR.
+    pub num_chunks: i32,
 
-    /// Unique string identifying the remote file.
-    #[sea_orm(unique)]
-    pub remote_file_id: String,
+    /// Hint indicating whether all chunks making up this NAR are available.
+    ///
+    /// This is used by the `get-missing-paths` endpoint to
+    /// also return store paths that are inaccessible due to
+    /// missing chunks in the associated NARs. They can then be
+    /// repaired by any client uploading.
+    ///
+    /// This flag may be outdated, but it's okay since when a client
+    /// tries to upload the same NAR, it will be immediately deduplicated
+    /// if all chunks are present and the flag will be updated.
+    pub completeness_hint: bool,
 
     /// Number of processes holding this NAR.
     ///
@@ -111,6 +123,9 @@ pub struct Model {
 pub enum Relation {
     #[sea_orm(has_many = "super::object::Entity")]
     Object,
+
+    #[sea_orm(has_many = "super::chunkref::Entity")]
+    ChunkRef,
 }
 
 impl ActiveModelBehavior for ActiveModel {}
diff --git a/server/src/database/entity/object.rs b/server/src/database/entity/object.rs
index da6233f..37ad893 100644
--- a/server/src/database/entity/object.rs
+++ b/server/src/database/entity/object.rs
@@ -90,12 +90,6 @@ pub enum Relation {
 impl Model {
     /// Converts this object to a NarInfo.
     pub fn to_nar_info(&self, nar: &NarModel) -> ServerResult<NarInfo> {
-        // FIXME: Return Err if file_hash and file_size don't exist
-        let file_size = nar
-            .file_size
-            .unwrap()
-            .try_into()
-            .map_err(ServerError::database_error)?;
         let nar_size = nar
             .nar_size
             .try_into()
@@ -106,8 +100,8 @@ impl Model {
             url: format!("nar/{}.nar", self.store_path_hash.as_str()),
 
             compression: Compression::from_str(&nar.compression)?,
-            file_hash: Hash::from_typed(nar.file_hash.as_ref().unwrap())?,
-            file_size,
+            file_hash: None, // FIXME
+            file_size: None, // FIXME
             nar_hash: Hash::from_typed(&nar.nar_hash)?,
             nar_size,
             system: self.system.to_owned(),
diff --git a/server/src/database/migration/m20221227_000002_create_nar_table.rs b/server/src/database/migration/m20221227_000002_create_nar_table.rs
index bcc2dd8..875c451 100644
--- a/server/src/database/migration/m20221227_000002_create_nar_table.rs
+++ b/server/src/database/migration/m20221227_000002_create_nar_table.rs
@@ -33,12 +33,16 @@ impl MigrationTrait for Migration {
                     )
                     .col(ColumnDef::new(Column::NarHash).string().not_null())
                     .col(ColumnDef::new(Column::NarSize).big_integer().not_null())
-                    .col(ColumnDef::new(Column::FileHash).string().null())
-                    .col(ColumnDef::new(Column::FileSize).big_integer().null())
+                    .col(ColumnDef::new(Alias::new("file_hash")).string().null())
+                    .col(ColumnDef::new(Alias::new("file_size")).big_integer().null())
                     .col(ColumnDef::new(Column::Compression).string().not_null())
-                    .col(ColumnDef::new(Column::RemoteFile).string().not_null())
                     .col(
-                        ColumnDef::new(Column::RemoteFileId)
+                        ColumnDef::new(Alias::new("remote_file"))
+                            .string()
+                            .not_null(),
+                    )
+                    .col(
+                        ColumnDef::new(Alias::new("remote_file_id"))
                             .string()
                             .not_null()
                             .unique_key(),
diff --git a/server/src/database/migration/m20230112_000001_add_chunk_table.rs b/server/src/database/migration/m20230112_000001_add_chunk_table.rs
new file mode 100644
index 0000000..3aeef94
--- /dev/null
+++ b/server/src/database/migration/m20230112_000001_add_chunk_table.rs
@@ -0,0 +1,70 @@
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::chunk::*;
+
+pub struct Migration;
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000001_add_chunk_table"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        manager
+            .create_table(
+                Table::create()
+                    .table(Entity)
+                    .col(
+                        ColumnDef::new(Column::Id)
+                            .big_integer()
+                            .not_null()
+                            .auto_increment()
+                            .primary_key(),
+                    )
+                    .col(
+                        ColumnDef::new(Column::State)
+                            .r#char()
+                            .char_len(1)
+                            .not_null(),
+                    )
+                    .col(ColumnDef::new(Column::ChunkHash).string().not_null())
+                    .col(ColumnDef::new(Column::ChunkSize).big_integer().not_null())
+                    .col(ColumnDef::new(Alias::new("file_hash")).string().null())
+                    .col(ColumnDef::new(Alias::new("file_size")).big_integer().null())
+                    .col(ColumnDef::new(Column::Compression).string().not_null())
+                    .col(ColumnDef::new(Column::RemoteFile).string().not_null())
+                    .col(
+                        ColumnDef::new(Column::RemoteFileId)
+                            .string()
+                            .not_null()
+                            .unique_key(),
+                    )
+                    .col(
+                        ColumnDef::new(Column::HoldersCount)
+                            .integer()
+                            .not_null()
+                            .default(0),
+                    )
+                    .col(
+                        ColumnDef::new(Column::CreatedAt)
+                            .timestamp_with_time_zone()
+                            .not_null(),
+                    )
+                    .to_owned(),
+            )
+            .await?;
+
+        manager
+            .create_index(
+                Index::create()
+                    .name("idx-chunk-chunk-hash")
+                    .table(Entity)
+                    .col(Column::ChunkHash)
+                    .to_owned(),
+            )
+            .await
+    }
+}
diff --git a/server/src/database/migration/m20230112_000002_add_chunkref_table.rs b/server/src/database/migration/m20230112_000002_add_chunkref_table.rs
new file mode 100644
index 0000000..2153841
--- /dev/null
+++ b/server/src/database/migration/m20230112_000002_add_chunkref_table.rs
@@ -0,0 +1,76 @@
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::chunk;
+use crate::database::entity::chunkref::*;
+use crate::database::entity::nar;
+
+pub struct Migration;
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000002_add_chunkref_table"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        manager
+            .create_table(
+                Table::create()
+                    .table(Entity)
+                    .col(
+                        ColumnDef::new(Column::Id)
+                            .big_integer()
+                            .not_null()
+                            .auto_increment()
+                            .primary_key(),
+                    )
+                    .col(ColumnDef::new(Column::NarId).big_integer().not_null())
+                    .col(ColumnDef::new(Column::Seq).integer().not_null())
+                    .col(ColumnDef::new(Column::ChunkId).big_integer().null())
+                    .col(ColumnDef::new(Column::ChunkHash).string().not_null())
+                    .col(ColumnDef::new(Column::Compression).string().not_null())
+                    .foreign_key(
+                        ForeignKeyCreateStatement::new()
+                            .name("fk_chunkref_chunk")
+                            .from_tbl(Entity)
+                            .from_col(Column::ChunkId)
+                            .to_tbl(chunk::Entity)
+                            .to_col(chunk::Column::Id)
+                            .on_delete(ForeignKeyAction::SetNull),
+                    )
+                    .foreign_key(
+                        ForeignKeyCreateStatement::new()
+                            .name("fk_chunkref_nar")
+                            .from_tbl(Entity)
+                            .from_col(Column::NarId)
+                            .to_tbl(nar::Entity)
+                            .to_col(nar::Column::Id)
+                            .on_delete(ForeignKeyAction::Cascade),
+                    )
+                    .to_owned(),
+            )
+            .await?;
+
+        manager
+            .create_index(
+                Index::create()
+                    .name("idx-chunk-nar-id")
+                    .table(Entity)
+                    .col(Column::NarId)
+                    .to_owned(),
+            )
+            .await?;
+
+        manager
+            .create_index(
+                Index::create()
+                    .name("idx-chunk-chunk-id")
+                    .table(Entity)
+                    .col(Column::ChunkId)
+                    .to_owned(),
+            )
+            .await
+    }
+}
diff --git a/server/src/database/migration/m20230112_000003_add_nar_num_chunks.rs b/server/src/database/migration/m20230112_000003_add_nar_num_chunks.rs
new file mode 100644
index 0000000..bbf9fb7
--- /dev/null
+++ b/server/src/database/migration/m20230112_000003_add_nar_num_chunks.rs
@@ -0,0 +1,32 @@
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::nar::*;
+
+pub struct Migration;
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000003_add_nar_num_chunks"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        manager
+            .alter_table(
+                Table::alter()
+                    .table(Entity)
+                    .add_column(
+                        ColumnDef::new(Column::NumChunks)
+                            .integer()
+                            .not_null()
+                            .default(1),
+                    )
+                    .to_owned(),
+            )
+            .await?;
+
+        Ok(())
+    }
+}
diff --git a/server/src/database/migration/m20230112_000004_migrate_nar_remote_files_to_chunks.rs b/server/src/database/migration/m20230112_000004_migrate_nar_remote_files_to_chunks.rs
new file mode 100644
index 0000000..42d70a6
--- /dev/null
+++ b/server/src/database/migration/m20230112_000004_migrate_nar_remote_files_to_chunks.rs
@@ -0,0 +1,143 @@
+use sea_orm::{ConnectionTrait, TransactionTrait};
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::chunk;
+use crate::database::entity::chunkref;
+use crate::database::entity::nar;
+
+pub struct Migration;
+
+pub enum TempChunkCols {
+    /// The ID of the NAR.
+    NarId,
+}
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000004_migrate_nar_remote_files_to_chunks"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        // When this migration is run, we assume that there are no
+        // preexisting chunks.
+
+        eprintln!("* Migrating NARs to chunks...");
+
+        // Add a temporary column into `chunk` to store the related `nar_id`.
+        manager
+            .alter_table(
+                Table::alter()
+                    .table(chunk::Entity)
+                    .add_column_if_not_exists(
+                        ColumnDef::new(TempChunkCols::NarId).integer().not_null(),
+                    )
+                    .to_owned(),
+            )
+            .await?;
+
+        // Get the original values from NARs
+        let select_remote_file = Query::select()
+            .from(nar::Entity)
+            .columns([
+                nar::Column::Id.into_iden(),
+                Alias::new("remote_file").into_iden(),
+                Alias::new("remote_file_id").into_iden(),
+                nar::Column::NarHash.into_iden(),
+                nar::Column::NarSize.into_iden(),
+                Alias::new("file_hash").into_iden(),
+                Alias::new("file_size").into_iden(),
+                nar::Column::Compression.into_iden(),
+                nar::Column::CreatedAt.into_iden(),
+            ])
+            .expr_as(chunk::ChunkState::Valid, chunk::Column::State.into_iden())
+            .to_owned();
+
+        // ... insert them into the `chunk` table
+        let insert_chunk = Query::insert()
+            .into_table(chunk::Entity)
+            .columns([
+                TempChunkCols::NarId.into_iden(),
+                chunk::Column::RemoteFile.into_iden(),
+                chunk::Column::RemoteFileId.into_iden(),
+                chunk::Column::ChunkHash.into_iden(),
+                chunk::Column::ChunkSize.into_iden(),
+                chunk::Column::FileHash.into_iden(),
+                chunk::Column::FileSize.into_iden(),
+                chunk::Column::Compression.into_iden(),
+                chunk::Column::CreatedAt.into_iden(),
+                chunk::Column::State.into_iden(),
+            ])
+            .select_from(select_remote_file)
+            .unwrap()
+            .returning(Query::returning().columns([
+                chunk::Column::Id.into_column_ref(),
+                TempChunkCols::NarId.into_column_ref(),
+            ]))
+            .to_owned();
+
+        let insert_chunk_stmt = manager.get_database_backend().build(&insert_chunk);
+
+        // ... then create chunkrefs binding the chunks and original NARs
+        let select_chunk = Query::select()
+            .from(chunk::Entity)
+            .columns([
+                chunk::Column::Id.into_iden(),
+                TempChunkCols::NarId.into_iden(),
+                chunk::Column::ChunkHash.into_iden(),
+                chunk::Column::Compression.into_iden(),
+            ])
+            .expr_as(0, chunkref::Column::Seq.into_iden())
+            .to_owned();
+
+        let insert_chunkref = Query::insert()
+            .into_table(chunkref::Entity)
+            .columns([
+                chunkref::Column::ChunkId.into_iden(),
+                chunkref::Column::NarId.into_iden(),
+                chunkref::Column::ChunkHash.into_iden(),
+                chunkref::Column::Compression.into_iden(),
+                chunkref::Column::Seq.into_iden(),
+            ])
+            .select_from(select_chunk)
+            .unwrap()
+            .returning(Query::returning().columns([chunkref::Column::Id.into_column_ref()]))
+            .to_owned();
+
+        let insert_chunkref_stmt = manager.get_database_backend().build(&insert_chunkref);
+
+        // Actually run the migration
+        let txn = manager.get_connection().begin().await?;
+        txn.execute(insert_chunk_stmt).await?;
+        txn.execute(insert_chunkref_stmt).await?;
+        txn.commit().await?;
+
+        // Finally, drop the temporary column
+        manager
+            .alter_table(
+                Table::alter()
+                    .table(chunk::Entity)
+                    .drop_column(TempChunkCols::NarId)
+                    .to_owned(),
+            )
+            .await?;
+
+        // We will drop the unused columns in `nar` in the next migration
+        Ok(())
+    }
+}
+
+impl Iden for TempChunkCols {
+    fn unquoted(&self, s: &mut dyn std::fmt::Write) {
+        write!(
+            s,
+            "{}",
+            match self {
+                Self::NarId => "temp_nar_id",
+            }
+        )
+        .unwrap();
+    }
+}
diff --git a/server/src/database/migration/m20230112_000005_drop_old_nar_columns.rs b/server/src/database/migration/m20230112_000005_drop_old_nar_columns.rs
new file mode 100644
index 0000000..9d29b66
--- /dev/null
+++ b/server/src/database/migration/m20230112_000005_drop_old_nar_columns.rs
@@ -0,0 +1,159 @@
+use sea_orm::{ConnectionTrait, DatabaseBackend, Statement};
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::nar::{self, *};
+
+pub struct Migration;
+
+const TEMP_NAR_TABLE: &str = "nar_new";
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000005_drop_old_nar_columns"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        eprintln!("* Migrating NAR schema...");
+
+        if manager.get_database_backend() == DatabaseBackend::Sqlite {
+            // Just copy all data to a new table
+            manager
+                .get_connection()
+                .execute(Statement::from_string(
+                    manager.get_database_backend(),
+                    "PRAGMA foreign_keys = OFF".to_owned(),
+                ))
+                .await?;
+
+            manager
+                .create_table(
+                    Table::create()
+                        .table(Alias::new(TEMP_NAR_TABLE))
+                        .if_not_exists()
+                        .col(
+                            ColumnDef::new(Column::Id)
+                                .big_integer()
+                                .not_null()
+                                .auto_increment()
+                                .primary_key(),
+                        )
+                        .col(
+                            ColumnDef::new(Column::State)
+                                .r#char()
+                                .char_len(1)
+                                .not_null(),
+                        )
+                        .col(ColumnDef::new(Column::NarHash).string().not_null())
+                        .col(ColumnDef::new(Column::NarSize).big_integer().not_null())
+                        .col(ColumnDef::new(Column::Compression).string().not_null())
+                        .col(
+                            ColumnDef::new(Column::NumChunks)
+                                .integer()
+                                .not_null()
+                                .default(1),
+                        )
+                        .col(
+                            ColumnDef::new(Column::HoldersCount)
+                                .integer()
+                                .not_null()
+                                .default(0),
+                        )
+                        .col(
+                            ColumnDef::new(Column::CreatedAt)
+                                .timestamp_with_time_zone()
+                                .not_null(),
+                        )
+                        .to_owned(),
+                )
+                .await?;
+
+            let columns = [
+                nar::Column::Id.into_iden(),
+                nar::Column::State.into_iden(),
+                nar::Column::NarHash.into_iden(),
+                nar::Column::NarSize.into_iden(),
+                nar::Column::Compression.into_iden(),
+                nar::Column::NumChunks.into_iden(),
+                nar::Column::HoldersCount.into_iden(),
+                nar::Column::CreatedAt.into_iden(),
+            ];
+
+            let select_nar = Query::select()
+                .from(nar::Entity)
+                .columns(columns.clone())
+                .to_owned();
+
+            let insertion = Query::insert()
+                .into_table(Alias::new(TEMP_NAR_TABLE))
+                .columns(columns.clone())
+                .select_from(select_nar)
+                .unwrap()
+                .to_owned();
+
+            let insertion_stmt = manager.get_database_backend().build(&insertion);
+            manager.get_connection().execute(insertion_stmt).await?;
+
+            manager
+                .drop_table(Table::drop().table(nar::Entity).to_owned())
+                .await?;
+
+            manager
+                .rename_table(
+                    Table::rename()
+                        .table(Alias::new(TEMP_NAR_TABLE), nar::Entity)
+                        .to_owned(),
+                )
+                .await?;
+
+            manager
+                .get_connection()
+                .execute(Statement::from_string(
+                    manager.get_database_backend(),
+                    "PRAGMA foreign_keys = ON".to_owned(),
+                ))
+                .await?;
+        } else {
+            // Just drop the columns
+            manager
+                .alter_table(
+                    Table::alter()
+                        .table(nar::Entity)
+                        .drop_column(Alias::new("file_hash"))
+                        .to_owned(),
+                )
+                .await?;
+
+            manager
+                .alter_table(
+                    Table::alter()
+                        .table(nar::Entity)
+                        .drop_column(Alias::new("file_size"))
+                        .to_owned(),
+                )
+                .await?;
+
+            manager
+                .alter_table(
+                    Table::alter()
+                        .table(nar::Entity)
+                        .drop_column(Alias::new("remote_file"))
+                        .to_owned(),
+                )
+                .await?;
+
+            manager
+                .alter_table(
+                    Table::alter()
+                        .table(nar::Entity)
+                        .drop_column(Alias::new("remote_file_id"))
+                        .to_owned(),
+                )
+                .await?;
+        }
+
+        Ok(())
+    }
+}
diff --git a/server/src/database/migration/m20230112_000006_add_nar_completeness_hint.rs b/server/src/database/migration/m20230112_000006_add_nar_completeness_hint.rs
new file mode 100644
index 0000000..d3f3a15
--- /dev/null
+++ b/server/src/database/migration/m20230112_000006_add_nar_completeness_hint.rs
@@ -0,0 +1,32 @@
+use sea_orm_migration::prelude::*;
+
+use crate::database::entity::nar::*;
+
+pub struct Migration;
+
+impl MigrationName for Migration {
+    fn name(&self) -> &str {
+        "m20230112_000006_add_nar_completeness_hint"
+    }
+}
+
+#[async_trait::async_trait]
+impl MigrationTrait for Migration {
+    async fn up(&self, manager: &SchemaManager) -> Result<(), DbErr> {
+        manager
+            .alter_table(
+                Table::alter()
+                    .table(Entity)
+                    .add_column(
+                        ColumnDef::new(Column::CompletenessHint)
+                            .boolean()
+                            .not_null()
+                            .default(true),
+                    )
+                    .to_owned(),
+            )
+            .await?;
+
+        Ok(())
+    }
+}
diff --git a/server/src/database/migration/mod.rs b/server/src/database/migration/mod.rs
index b42033f..406e709 100644
--- a/server/src/database/migration/mod.rs
+++ b/server/src/database/migration/mod.rs
@@ -8,6 +8,12 @@ mod m20221227_000003_create_object_table;
 mod m20221227_000004_add_object_last_accessed;
 mod m20221227_000005_add_cache_retention_period;
 mod m20230103_000001_add_object_created_by;
+mod m20230112_000001_add_chunk_table;
+mod m20230112_000002_add_chunkref_table;
+mod m20230112_000003_add_nar_num_chunks;
+mod m20230112_000004_migrate_nar_remote_files_to_chunks;
+mod m20230112_000005_drop_old_nar_columns;
+mod m20230112_000006_add_nar_completeness_hint;
 
 pub struct Migrator;
 
@@ -21,6 +27,12 @@ impl MigratorTrait for Migrator {
             Box::new(m20221227_000004_add_object_last_accessed::Migration),
             Box::new(m20221227_000005_add_cache_retention_period::Migration),
             Box::new(m20230103_000001_add_object_created_by::Migration),
+            Box::new(m20230112_000001_add_chunk_table::Migration),
+            Box::new(m20230112_000002_add_chunkref_table::Migration),
+            Box::new(m20230112_000003_add_nar_num_chunks::Migration),
+            Box::new(m20230112_000004_migrate_nar_remote_files_to_chunks::Migration),
+            Box::new(m20230112_000005_drop_old_nar_columns::Migration),
+            Box::new(m20230112_000006_add_nar_completeness_hint::Migration),
         ]
     }
 }
diff --git a/server/src/database/mod.rs b/server/src/database/mod.rs
index 1f2ab06..b7b144c 100644
--- a/server/src/database/mod.rs
+++ b/server/src/database/mod.rs
@@ -3,35 +3,43 @@ pub mod migration;
 
 use std::ops::Deref;
 
+use anyhow::anyhow;
 use async_trait::async_trait;
 use chrono::Utc;
 use sea_orm::entity::prelude::*;
 use sea_orm::entity::Iterable as EnumIterable;
-use sea_orm::query::{JoinType, QuerySelect, QueryTrait};
+use sea_orm::query::{JoinType, QueryOrder, QuerySelect, QueryTrait};
 use sea_orm::sea_query::{Expr, LockBehavior, LockType, Query, Value};
 use sea_orm::{ActiveValue::Set, ConnectionTrait, DatabaseConnection, FromQueryResult};
 use tokio::task;
 
 use crate::error::{ErrorKind, ServerError, ServerResult};
+use crate::narinfo::Compression;
 use attic::cache::CacheName;
 use attic::hash::Hash;
 use attic::nix_store::StorePathHash;
 use entity::cache::{self, CacheModel, Entity as Cache};
+use entity::chunk::{self, ChunkModel, ChunkState, Entity as Chunk};
+use entity::chunkref;
 use entity::nar::{self, Entity as Nar, NarModel, NarState};
 use entity::object::{self, Entity as Object, ObjectModel};
 
+// quintuple join time
 const SELECT_OBJECT: &str = "O_";
 const SELECT_CACHE: &str = "C_";
 const SELECT_NAR: &str = "N_";
+const SELECT_CHUNK: &str = "CH_";
+const SELECT_CHUNKREF: &str = "CHR_";
 
 #[async_trait]
 pub trait AtticDatabase: Send + Sync {
-    /// Retrieves an object in a binary cache by its store path hash.
-    async fn find_object_by_store_path_hash(
+    /// Retrieves an object in a binary cache by its store path hash, returning all its
+    /// chunks.
+    async fn find_object_and_chunks_by_store_path_hash(
         &self,
         cache: &CacheName,
         store_path_hash: &StorePathHash,
-    ) -> ServerResult<(ObjectModel, CacheModel, NarModel)>;
+    ) -> ServerResult<(ObjectModel, CacheModel, NarModel, Vec<Option<ChunkModel>>)>;
 
     /// Retrieves a binary cache.
     async fn find_cache(&self, cache: &CacheName) -> ServerResult<CacheModel>;
@@ -39,6 +47,13 @@ pub trait AtticDatabase: Send + Sync {
     /// Retrieves and locks a valid NAR matching a NAR Hash.
     async fn find_and_lock_nar(&self, nar_hash: &Hash) -> ServerResult<Option<NarGuard>>;
 
+    /// Retrieves and locks a valid chunk matching a chunk Hash.
+    async fn find_and_lock_chunk(
+        &self,
+        chunk_hash: &Hash,
+        compression: Compression,
+    ) -> ServerResult<Option<ChunkGuard>>;
+
     /// Bumps the last accessed timestamp of an object.
     async fn bump_object_last_accessed(&self, object_id: i64) -> ServerResult<()>;
 }
@@ -48,6 +63,11 @@ pub struct NarGuard {
     nar: NarModel,
 }
 
+pub struct ChunkGuard {
+    database: DatabaseConnection,
+    chunk: ChunkModel,
+}
+
 fn prefix_column<E: EntityTrait, S: QuerySelect>(mut select: S, prefix: &str) -> S {
     for col in <E::Column as EnumIterable>::iter() {
         let alias = format!("{}{}", prefix, Iden::to_string(&col));
@@ -57,47 +77,139 @@ fn prefix_column<E: EntityTrait, S: QuerySelect>(mut select: S, prefix: &str) ->
 }
 
 pub fn build_cache_object_nar_query() -> Select<Object> {
+    /*
+        Build something like:
+
+        -- chunkrefs must exist but chunks may not exist
+
+        select * from object
+        inner join cache
+             on object.cache_id = cache.id
+        inner join nar
+             on object.nar_id = nar.id
+        inner join chunkref
+            on chunkref.nar_id = nar.id
+        left join chunk
+            on chunkref.chunk_id = chunk.id
+        where
+            object.store_path_hash = 'fiwsv60kgwrfvib2nf9dkq9q8bk1h7qh' and
+            nar.state = 'V' and
+            cache.name = 'zhaofeng' and
+            cache.deleted_at is null
+
+        Returns (CacheModel, ObjectModel, NarModel, Vec<Option<ChunkModel>>)
+        where the number of elements in the Vec must be equal to `nar.num_chunks`.
+
+        If any element in the chunk `Vec` is `None`, it means the chunk is missing
+        for some reason (e.g., corrupted) and the full NAR cannot be reconstructed.
+        In such cases, .narinfo/.nar requests will return HTTP 503 and the affected
+        store paths will be treated as non-existent in `get-missing-paths` so they
+        can be repaired automatically when any client upload a path containing the
+        missing chunk.
+
+        It's a quintuple join and the query plans look reasonable on SQLite
+        and Postgres. For each .narinfo/.nar request, we only submit a single query.
+    */
     let mut query = Object::find()
         .select_only()
-        .join(JoinType::LeftJoin, object::Relation::Cache.def())
-        .join(JoinType::LeftJoin, object::Relation::Nar.def());
+        .join(JoinType::InnerJoin, object::Relation::Cache.def())
+        .join(JoinType::InnerJoin, object::Relation::Nar.def())
+        .join(JoinType::InnerJoin, nar::Relation::ChunkRef.def())
+        .join(JoinType::LeftJoin, chunkref::Relation::Chunk.def())
+        .order_by_asc(chunkref::Column::Seq);
 
     query = prefix_column::<object::Entity, _>(query, SELECT_OBJECT);
     query = prefix_column::<cache::Entity, _>(query, SELECT_CACHE);
     query = prefix_column::<nar::Entity, _>(query, SELECT_NAR);
+    query = prefix_column::<chunk::Entity, _>(query, SELECT_CHUNK);
+    query = prefix_column::<chunkref::Entity, _>(query, SELECT_CHUNKREF);
 
     query
 }
 
 #[async_trait]
 impl AtticDatabase for DatabaseConnection {
-    async fn find_object_by_store_path_hash(
+    async fn find_object_and_chunks_by_store_path_hash(
         &self,
         cache: &CacheName,
         store_path_hash: &StorePathHash,
-    ) -> ServerResult<(ObjectModel, CacheModel, NarModel)> {
+    ) -> ServerResult<(ObjectModel, CacheModel, NarModel, Vec<Option<ChunkModel>>)> {
         let stmt = build_cache_object_nar_query()
             .filter(cache::Column::Name.eq(cache.as_str()))
             .filter(cache::Column::DeletedAt.is_null())
             .filter(object::Column::StorePathHash.eq(store_path_hash.as_str()))
             .filter(nar::Column::State.eq(NarState::Valid))
-            .limit(1)
+            .filter(
+                chunk::Column::State
+                    .eq(ChunkState::Valid)
+                    .or(chunk::Column::State.is_null()),
+            )
             .build(self.get_database_backend());
 
-        let result = self
-            .query_one(stmt)
+        let results = self
+            .query_all(stmt)
             .await
-            .map_err(ServerError::database_error)?
-            .ok_or(ErrorKind::NoSuchObject)?;
-
-        let object = object::Model::from_query_result(&result, SELECT_OBJECT)
-            .map_err(ServerError::database_error)?;
-        let cache = cache::Model::from_query_result(&result, SELECT_CACHE)
-            .map_err(ServerError::database_error)?;
-        let nar = nar::Model::from_query_result(&result, SELECT_NAR)
             .map_err(ServerError::database_error)?;
 
-        Ok((object, cache, nar))
+        if results.is_empty() {
+            return Err(ErrorKind::NoSuchObject.into());
+        }
+
+        let mut it = results.iter();
+        let first = it.next().unwrap();
+
+        let mut chunks = Vec::new();
+
+        let object = object::Model::from_query_result(first, SELECT_OBJECT)
+            .map_err(ServerError::database_error)?;
+        let cache = cache::Model::from_query_result(first, SELECT_CACHE)
+            .map_err(ServerError::database_error)?;
+        let nar = nar::Model::from_query_result(first, SELECT_NAR)
+            .map_err(ServerError::database_error)?;
+
+        if results.len() != nar.num_chunks as usize {
+            // Something went terribly wrong. This means there are a wrong number of `chunkref` rows.
+            return Err(ErrorKind::DatabaseError(anyhow!(
+                "Database returned the wrong number of chunks: Expected {}, got {}",
+                nar.num_chunks,
+                results.len()
+            ))
+            .into());
+        }
+
+        chunks.push({
+            let chunk_id: Option<i64> = first
+                .try_get(SELECT_CHUNK, chunk::Column::Id.as_str())
+                .map_err(ServerError::database_error)?;
+
+            if chunk_id.is_some() {
+                Some(
+                    chunk::Model::from_query_result(first, SELECT_CHUNK)
+                        .map_err(ServerError::database_error)?,
+                )
+            } else {
+                None
+            }
+        });
+
+        for chunk in it {
+            chunks.push({
+                let chunk_id: Option<i64> = chunk
+                    .try_get(SELECT_CHUNK, chunk::Column::Id.as_str())
+                    .map_err(ServerError::database_error)?;
+
+                if chunk_id.is_some() {
+                    Some(
+                        chunk::Model::from_query_result(chunk, SELECT_CHUNK)
+                            .map_err(ServerError::database_error)?,
+                    )
+                } else {
+                    None
+                }
+            });
+        }
+
+        Ok((object, cache, nar, chunks))
     }
 
     async fn find_cache(&self, cache: &CacheName) -> ServerResult<CacheModel> {
@@ -143,6 +255,45 @@ impl AtticDatabase for DatabaseConnection {
         Ok(guard)
     }
 
+    // FIXME: Repetition
+    async fn find_and_lock_chunk(
+        &self,
+        chunk_hash: &Hash,
+        compression: Compression,
+    ) -> ServerResult<Option<ChunkGuard>> {
+        let one = Value::Unsigned(Some(1));
+        let matched_ids = Query::select()
+            .from(Chunk)
+            .and_where(chunk::Column::ChunkHash.eq(chunk_hash.to_typed_base16()))
+            .and_where(chunk::Column::State.eq(ChunkState::Valid))
+            .and_where(chunk::Column::Compression.eq(compression.as_str()))
+            .expr(Expr::col(chunk::Column::Id))
+            .lock_with_behavior(LockType::Update, LockBehavior::SkipLocked)
+            .limit(1)
+            .to_owned();
+        let incr_holders = Query::update()
+            .table(Chunk)
+            .values([(
+                chunk::Column::HoldersCount,
+                Expr::col(chunk::Column::HoldersCount).add(one),
+            )])
+            .and_where(chunk::Column::Id.in_subquery(matched_ids))
+            .returning_all()
+            .to_owned();
+        let stmt = self.get_database_backend().build(&incr_holders);
+
+        let guard = chunk::Model::find_by_statement(stmt)
+            .one(self)
+            .await
+            .map_err(ServerError::database_error)?
+            .map(|chunk| ChunkGuard {
+                database: self.clone(),
+                chunk,
+            });
+
+        Ok(guard)
+    }
+
     async fn bump_object_last_accessed(&self, object_id: i64) -> ServerResult<()> {
         let now = Utc::now();
 
@@ -192,3 +343,43 @@ impl Drop for NarGuard {
         });
     }
 }
+
+impl ChunkGuard {
+    pub fn from_locked(database: DatabaseConnection, chunk: ChunkModel) -> Self {
+        Self { database, chunk }
+    }
+}
+
+impl Deref for ChunkGuard {
+    type Target = ChunkModel;
+
+    fn deref(&self) -> &Self::Target {
+        &self.chunk
+    }
+}
+
+impl Drop for ChunkGuard {
+    fn drop(&mut self) {
+        let database = self.database.clone();
+        let chunk_id = self.chunk.id;
+
+        task::spawn(async move {
+            tracing::debug!("Unlocking chunk");
+
+            let one = Value::Unsigned(Some(1));
+            let decr_holders = Query::update()
+                .table(Chunk)
+                .values([(
+                    chunk::Column::HoldersCount,
+                    Expr::col(chunk::Column::HoldersCount).sub(one),
+                )])
+                .and_where(chunk::Column::Id.eq(chunk_id))
+                .to_owned();
+            let stmt = database.get_database_backend().build(&decr_holders);
+
+            if let Err(e) = database.execute(stmt).await {
+                tracing::warn!("Failed to decrement holders count: {}", e);
+            }
+        });
+    }
+}
diff --git a/server/src/error.rs b/server/src/error.rs
index dcd559d..f5d5d41 100644
--- a/server/src/error.rs
+++ b/server/src/error.rs
@@ -54,6 +54,9 @@ pub enum ErrorKind {
     /// Invalid compression type "{name}".
     InvalidCompressionType { name: String },
 
+    /// The requested NAR has missing chunks and needs to be repaired.
+    IncompleteNar,
+
     /// Database error: {0}
     DatabaseError(AnyError),
 
@@ -174,6 +177,7 @@ impl ErrorKind {
             Self::NoSuchCache => "NoSuchCache",
             Self::CacheAlreadyExists => "CacheAlreadyExists",
             Self::InvalidCompressionType { .. } => "InvalidCompressionType",
+            Self::IncompleteNar => "IncompleteNar",
             Self::AtticError(e) => e.name(),
             Self::DatabaseError(_) => "DatabaseError",
             Self::StorageError(_) => "StorageError",
@@ -218,6 +222,7 @@ impl ErrorKind {
             Self::NoSuchCache => StatusCode::NOT_FOUND,
             Self::NoSuchObject => StatusCode::NOT_FOUND,
             Self::CacheAlreadyExists => StatusCode::BAD_REQUEST,
+            Self::IncompleteNar => StatusCode::SERVICE_UNAVAILABLE,
             Self::ManifestSerializationError(_) => StatusCode::BAD_REQUEST,
             Self::RequestError(_) => StatusCode::BAD_REQUEST,
             Self::InvalidCompressionType { .. } => StatusCode::BAD_REQUEST,
diff --git a/server/src/gc.rs b/server/src/gc.rs
index 1378673..7b92014 100644
--- a/server/src/gc.rs
+++ b/server/src/gc.rs
@@ -17,6 +17,8 @@ use tracing::instrument;
 use super::{State, StateInner};
 use crate::config::Config;
 use crate::database::entity::cache::{self, Entity as Cache};
+use crate::database::entity::chunk::{self, ChunkState, Entity as Chunk};
+use crate::database::entity::chunkref::{self, Entity as ChunkRef};
 use crate::database::entity::nar::{self, Entity as Nar, NarState};
 use crate::database::entity::object::{self, Entity as Object};
 
@@ -54,6 +56,7 @@ pub async fn run_garbage_collection_once(config: Config) -> Result<()> {
     let state = StateInner::new(config).await;
     run_time_based_garbage_collection(&state).await?;
     run_reap_orphan_nars(&state).await?;
+    run_reap_orphan_chunks(&state).await?;
 
     Ok(())
 }
@@ -122,7 +125,6 @@ async fn run_time_based_garbage_collection(state: &State) -> Result<()> {
 #[instrument(skip_all)]
 async fn run_reap_orphan_nars(state: &State) -> Result<()> {
     let db = state.database().await?;
-    let storage = state.storage().await?;
 
     // find all orphan NARs...
     let orphan_nar_ids = Query::select()
@@ -140,46 +142,78 @@ async fn run_reap_orphan_nars(state: &State) -> Result<()> {
         .lock_with_tables_behavior(LockType::Update, [Nar], LockBehavior::SkipLocked)
         .to_owned();
 
+    // ... and simply delete them
+    let deletion = Nar::delete_many()
+        .filter(nar::Column::Id.in_subquery(orphan_nar_ids))
+        .exec(db)
+        .await?;
+
+    tracing::info!("Deleted {} orphan NARs", deletion.rows_affected,);
+
+    Ok(())
+}
+
+#[instrument(skip_all)]
+async fn run_reap_orphan_chunks(state: &State) -> Result<()> {
+    let db = state.database().await?;
+    let storage = state.storage().await?;
+
+    // find all orphan chunks...
+    let orphan_chunk_ids = Query::select()
+        .from(Chunk)
+        .expr(chunk::Column::Id.into_expr())
+        .left_join(
+            ChunkRef,
+            chunkref::Column::ChunkId
+                .into_expr()
+                .eq(chunk::Column::Id.into_expr()),
+        )
+        .and_where(chunkref::Column::Id.is_null())
+        .and_where(chunk::Column::State.eq(ChunkState::Valid))
+        .and_where(chunk::Column::HoldersCount.eq(0))
+        .lock_with_tables_behavior(LockType::Update, [Chunk], LockBehavior::SkipLocked)
+        .to_owned();
+
     // ... and transition their state to Deleted
     //
-    // Deleted NARs are essentially invisible from our normal queries
+    // Deleted chunks are essentially invisible from our normal queries
     let change_state = Query::update()
-        .table(Nar)
-        .value(nar::Column::State, NarState::Deleted)
-        .and_where(nar::Column::Id.in_subquery(orphan_nar_ids))
+        .table(Chunk)
+        .value(chunk::Column::State, ChunkState::Deleted)
+        .and_where(chunk::Column::Id.in_subquery(orphan_chunk_ids))
         .returning_all()
         .to_owned();
 
     let stmt = db.get_database_backend().build(&change_state);
 
-    let orphan_nars = nar::Model::find_by_statement(stmt).all(db).await?;
+    let orphan_chunks = chunk::Model::find_by_statement(stmt).all(db).await?;
 
-    if orphan_nars.is_empty() {
+    if orphan_chunks.is_empty() {
         return Ok(());
     }
 
-    // Delete the NARs from remote storage
+    // Delete the chunks from remote storage
     let delete_limit = Arc::new(Semaphore::new(20)); // TODO: Make this configurable
-    let futures: Vec<_> = orphan_nars
+    let futures: Vec<_> = orphan_chunks
         .into_iter()
-        .map(|nar| {
+        .map(|chunk| {
             let delete_limit = delete_limit.clone();
             async move {
                 let permit = delete_limit.acquire().await?;
-                storage.delete_file_db(&nar.remote_file.0).await?;
+                storage.delete_file_db(&chunk.remote_file.0).await?;
                 drop(permit);
-                Result::<_, anyhow::Error>::Ok(nar.id)
+                Result::<_, anyhow::Error>::Ok(chunk.id)
             }
         })
         .collect();
 
     // Deletions can result in spurious failures, tolerate them
     //
-    // NARs that failed to be deleted from the remote storage will
+    // Chunks that failed to be deleted from the remote storage will
     // just be stuck in Deleted state.
     //
     // TODO: Maybe have an interactive command to retry deletions?
-    let deleted_nar_ids: Vec<_> = join_all(futures)
+    let deleted_chunk_ids: Vec<_> = join_all(futures)
         .await
         .into_iter()
         .filter(|r| {
@@ -193,12 +227,12 @@ async fn run_reap_orphan_nars(state: &State) -> Result<()> {
         .collect();
 
     // Finally, delete them from the database
-    let deletion = Nar::delete_many()
-        .filter(nar::Column::Id.is_in(deleted_nar_ids))
+    let deletion = Chunk::delete_many()
+        .filter(chunk::Column::Id.is_in(deleted_chunk_ids))
         .exec(db)
         .await?;
 
-    tracing::info!("Deleted {} NARs", deletion.rows_affected);
+    tracing::info!("Deleted {} orphan chunks", deletion.rows_affected);
 
     Ok(())
 }
diff --git a/server/src/lib.rs b/server/src/lib.rs
index 7a22b0d..38d3b1a 100644
--- a/server/src/lib.rs
+++ b/server/src/lib.rs
@@ -15,9 +15,10 @@
 
 pub mod access;
 mod api;
+mod chunking;
 pub mod config;
 pub mod database;
-mod error;
+pub mod error;
 pub mod gc;
 mod middleware;
 mod narinfo;
diff --git a/server/src/narinfo/mod.rs b/server/src/narinfo/mod.rs
index dd26210..ba104f2 100644
--- a/server/src/narinfo/mod.rs
+++ b/server/src/narinfo/mod.rs
@@ -86,12 +86,18 @@ pub struct NarInfo {
     pub compression: Compression,
 
     /// The hash of the compressed file.
+    ///
+    /// We don't know the file hash if it's chunked.
     #[serde(rename = "FileHash")]
-    pub file_hash: Hash,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub file_hash: Option<Hash>,
 
     /// The size of the compressed file.
+    ///
+    /// We may not know the file size if it's chunked.
     #[serde(rename = "FileSize")]
-    pub file_size: usize,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub file_size: Option<usize>,
 
     /// The hash of the NAR archive.
     ///
@@ -242,6 +248,18 @@ impl IntoResponse for NarInfo {
     }
 }
 
+impl Compression {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            Self::None => "none",
+            Self::Xz => "xz",
+            Self::Bzip2 => "bzip2",
+            Self::Brotli => "br",
+            Self::Zstd => "zstd",
+        }
+    }
+}
+
 impl FromStr for Compression {
     type Err = ServerError;
 
@@ -262,13 +280,7 @@ impl FromStr for Compression {
 
 impl ToString for Compression {
     fn to_string(&self) -> String {
-        String::from(match self {
-            Self::None => "none",
-            Self::Xz => "xz",
-            Self::Bzip2 => "bzip2",
-            Self::Brotli => "br",
-            Self::Zstd => "zstd",
-        })
+        String::from(self.as_str())
     }
 }
 
diff --git a/server/src/narinfo/tests.rs b/server/src/narinfo/tests.rs
index de43366..f06942a 100644
--- a/server/src/narinfo/tests.rs
+++ b/server/src/narinfo/tests.rs
@@ -34,9 +34,9 @@ Sig: cache.nixos.org-1:lo9EfNIL4eGRuNh7DTbAAffWPpI2SlYC/8uP7JnhgmfRIUNGhSbFe8qEa
         assert_eq!(Compression::Xz, narinfo.compression);
         assert_eq!(
             "sha256:0nqgf15qfiacfxrgm2wkw0gwwncjqqzzalj8rs14w9srkydkjsk9",
-            narinfo.file_hash.to_typed_base32()
+            narinfo.file_hash.as_ref().unwrap().to_typed_base32()
         );
-        assert_eq!(41104, narinfo.file_size);
+        assert_eq!(Some(41104), narinfo.file_size);
         assert_eq!(
             "sha256:16mvl7v0ylzcg2n3xzjn41qhzbmgcn5iyarx16nn5l2r36n2kqci",
             narinfo.nar_hash.to_typed_base32()
diff --git a/server/src/storage/local.rs b/server/src/storage/local.rs
index 383ed4f..93ff737 100644
--- a/server/src/storage/local.rs
+++ b/server/src/storage/local.rs
@@ -87,15 +87,19 @@ impl StorageBackend for LocalBackend {
         Ok(())
     }
 
-    async fn download_file(&self, name: String) -> ServerResult<Download> {
+    async fn download_file(&self, name: String, _prefer_stream: bool) -> ServerResult<Download> {
         let file = File::open(self.get_path(&name))
             .await
             .map_err(ServerError::storage_error)?;
 
-        Ok(Download::Stream(Box::new(file)))
+        Ok(Download::AsyncRead(Box::new(file)))
     }
 
-    async fn download_file_db(&self, file: &RemoteFile) -> ServerResult<Download> {
+    async fn download_file_db(
+        &self,
+        file: &RemoteFile,
+        _prefer_stream: bool,
+    ) -> ServerResult<Download> {
         let file = if let RemoteFile::Local(file) = file {
             file
         } else {
@@ -109,7 +113,7 @@ impl StorageBackend for LocalBackend {
             .await
             .map_err(ServerError::storage_error)?;
 
-        Ok(Download::Stream(Box::new(file)))
+        Ok(Download::AsyncRead(Box::new(file)))
     }
 
     async fn make_db_reference(&self, name: String) -> ServerResult<RemoteFile> {
diff --git a/server/src/storage/mod.rs b/server/src/storage/mod.rs
index fad2b1a..37cd0c3 100644
--- a/server/src/storage/mod.rs
+++ b/server/src/storage/mod.rs
@@ -3,6 +3,8 @@
 mod local;
 mod s3;
 
+use bytes::Bytes;
+use futures::stream::BoxStream;
 use serde::{Deserialize, Serialize};
 use tokio::io::AsyncRead;
 
@@ -33,11 +35,14 @@ pub enum RemoteFile {
 
 /// Way to download a file.
 pub enum Download {
-    /// A redirect to a (possibly ephemeral) URL.
-    Redirect(String),
+    /// A possibly ephemeral URL.
+    Url(String),
 
     /// A stream.
-    Stream(Box<dyn AsyncRead + Unpin + Send>),
+    Stream(BoxStream<'static, std::io::Result<Bytes>>),
+
+    /// An AsyncRead.
+    AsyncRead(Box<dyn AsyncRead + Unpin + Send>),
 }
 
 // TODO: Maybe make RemoteFile the one true reference instead of having two sets of APIs?
@@ -58,10 +63,14 @@ pub trait StorageBackend: Send + Sync + std::fmt::Debug {
     async fn delete_file_db(&self, file: &RemoteFile) -> ServerResult<()>;
 
     /// Downloads a file using the current configuration.
-    async fn download_file(&self, name: String) -> ServerResult<Download>;
+    async fn download_file(&self, name: String, prefer_stream: bool) -> ServerResult<Download>;
 
     /// Downloads a file using a database reference.
-    async fn download_file_db(&self, file: &RemoteFile) -> ServerResult<Download>;
+    async fn download_file_db(
+        &self,
+        file: &RemoteFile,
+        prefer_stream: bool,
+    ) -> ServerResult<Download>;
 
     /// Creates a database reference for a file.
     async fn make_db_reference(&self, name: String) -> ServerResult<RemoteFile>;
diff --git a/server/src/storage/s3.rs b/server/src/storage/s3.rs
index 494a840..bec687b 100644
--- a/server/src/storage/s3.rs
+++ b/server/src/storage/s3.rs
@@ -1,13 +1,18 @@
 //! S3 remote files.
 
+use std::io::{Error as IoError, ErrorKind as IoErrorKind};
 use std::time::Duration;
 
 use async_trait::async_trait;
 use aws_sdk_s3::{
-    config::Builder as S3ConfigBuilder, model::CompletedMultipartUpload, model::CompletedPart,
-    presigning::config::PresigningConfig, Client, Credentials, Endpoint, Region,
+    client::fluent_builders::GetObject,
+    config::Builder as S3ConfigBuilder,
+    model::{CompletedMultipartUpload, CompletedPart},
+    presigning::config::PresigningConfig,
+    Client, Credentials, Endpoint, Region,
 };
 use futures::future::join_all;
+use futures::stream::StreamExt;
 use serde::{Deserialize, Serialize};
 use tokio::io::AsyncRead;
 
@@ -134,6 +139,29 @@ impl S3Backend {
 
         Ok((client, file))
     }
+
+    async fn get_download(&self, req: GetObject, prefer_stream: bool) -> ServerResult<Download> {
+        if prefer_stream {
+            let output = req.send().await.map_err(ServerError::storage_error)?;
+
+            let stream = StreamExt::map(output.body, |item| {
+                item.map_err(|e| IoError::new(IoErrorKind::Other, e))
+            });
+
+            Ok(Download::Stream(Box::pin(stream)))
+        } else {
+            // FIXME: Configurable expiration
+            let presign_config = PresigningConfig::expires_in(Duration::from_secs(600))
+                .map_err(ServerError::storage_error)?;
+
+            let presigned = req
+                .presigned(presign_config)
+                .await
+                .map_err(ServerError::storage_error)?;
+
+            Ok(Download::Url(presigned.uri().to_string()))
+        }
+    }
 }
 
 #[async_trait]
@@ -313,38 +341,26 @@ impl StorageBackend for S3Backend {
         Ok(())
     }
 
-    async fn download_file(&self, name: String) -> ServerResult<Download> {
-        // FIXME: Configurable expiration
-        let presign_config = PresigningConfig::expires_in(Duration::from_secs(10))
-            .map_err(ServerError::storage_error)?;
-
-        let presigned = self
+    async fn download_file(&self, name: String, prefer_stream: bool) -> ServerResult<Download> {
+        let req = self
             .client
             .get_object()
             .bucket(&self.config.bucket)
-            .key(&name)
-            .presigned(presign_config)
-            .await
-            .map_err(ServerError::storage_error)?;
+            .key(&name);
 
-        Ok(Download::Redirect(presigned.uri().to_string()))
+        self.get_download(req, prefer_stream).await
     }
 
-    async fn download_file_db(&self, file: &RemoteFile) -> ServerResult<Download> {
+    async fn download_file_db(
+        &self,
+        file: &RemoteFile,
+        prefer_stream: bool,
+    ) -> ServerResult<Download> {
         let (client, file) = self.get_client_from_db_ref(file).await?;
 
-        let presign_config = PresigningConfig::expires_in(Duration::from_secs(600))
-            .map_err(ServerError::storage_error)?;
+        let req = client.get_object().bucket(&file.bucket).key(&file.key);
 
-        let presigned = client
-            .get_object()
-            .bucket(&file.bucket)
-            .key(&file.key)
-            .presigned(presign_config)
-            .await
-            .map_err(ServerError::storage_error)?;
-
-        Ok(Download::Redirect(presigned.uri().to_string()))
+        self.get_download(req, prefer_stream).await
     }
 
     async fn make_db_reference(&self, name: String) -> ServerResult<RemoteFile> {