diff --git a/puppet/zulip/files/nginx/zulip-include-frontend/uploads-internal.conf b/puppet/zulip/files/nginx/zulip-include-frontend/uploads-internal.conf
index 8afaf3ac78..a8a2a70418 100644
--- a/puppet/zulip/files/nginx/zulip-include-frontend/uploads-internal.conf
+++ b/puppet/zulip/files/nginx/zulip-include-frontend/uploads-internal.conf
@@ -43,6 +43,16 @@ location ~ ^/internal/s3/(?<s3_hostname>[^/]+)/(?<s3_path>.*) {
     # the first response.  Django explicitly unsets the first, and
     # does not set the latter two.
 
+    # We slice the content into 5M chunks; this means that the client
+    # doesn't need to wait for nginx to download and cache the full
+    # content if the client just requested a small range (e.g. for
+    # showing a thumbnail of a large video).  5M is chosen to be
+    # enough for videos to be able to thumbnail in one slice, but not
+    # take overly long to retrieve from S3, or cause overwhelming
+    # numbers of cache entries for large files.
+    slice 5m;
+    proxy_set_header Range $slice_range;
+
     proxy_pass $download_url$is_args$args;
     proxy_cache uploads;
     # If the S3 response doesn't contain Cache-Control headers (which
@@ -50,12 +60,12 @@ location ~ ^/internal/s3/(?<s3_hostname>[^/]+)/(?<s3_path>.*) {
     # long time.  The size of the cache is controlled by
     # `s3_disk_cache_size` and read frequency, set via
     # `s3_cache_inactive_time`.
-    proxy_cache_valid 200 1y;
+    proxy_cache_valid 200 206 1y;
 
-    # We only include the requested content-disposition in the cache
-    # key, so that we cache "Content-Disposition: attachment"
-    # separately from the inline version.
-    proxy_cache_key $download_url$s3_disposition_cache_key;
+    # We only include the requested content-disposition (and range
+    # slice) in the cache key, so that we cache "Content-Disposition:
+    # attachment" separately from the inline version.
+    proxy_cache_key $download_url$s3_disposition_cache_key$slice_range;
 }
 
 # Internal file-serving