diff --git a/_includes/quick-start-module.js b/_includes/quick-start-module.js
index 345c1d0434bb..41a155bbd232 100644
--- a/_includes/quick-start-module.js
+++ b/_includes/quick-start-module.js
@@ -142,17 +142,43 @@ function changeVersion(ptbuild) {
     var elems = document.querySelectorAll('[id^="'+arch_key+'"]');
     for (var i=0; i < elems.length;i++) {
       if(archMap[elems[i].id]) {
+        elems[i].style.display = "";
         elems[i].style.textDecoration = "";
         elems[i].children[0].textContent = info.title + " " + archMap[elems[i].id][1]
+      } else if (arch_key == "cuda") {
+        // The channel ships fewer CUDA versions than there are boxes.
+        elems[i].style.display = "none";
       } else {
         elems[i].style.textDecoration = "line-through";
       }
     }
   }
+  layoutCudaOptions();
   var stable_element = document.getElementById("stable");
   stable_element.children[0].textContent = stable_version;
 }
 
+// The CUDA boxes share 6 of the row's 12 columns; widen them when some are
+// hidden, and move the selection off a hidden box.
+function layoutCudaOptions() {
+  var elems = Array.from(document.querySelectorAll('[id^="cuda"]'));
+  var visible = elems.filter(function(e) { return e.style.display != "none"; });
+  if (visible.length == 0) {
+    return;
+  }
+  var width = "col-md-" + Math.floor(6 / visible.length);
+  for (var i = 0; i < elems.length; i++) {
+    elems[i].classList.remove("col-md-2", "col-md-3", "col-md-6");
+    elems[i].classList.add(width);
+  }
+  var selected = document.getElementById(opts.cuda);
+  if (selected && selected.style.display == "none") {
+    $(cuda).removeClass("selected");
+    $(visible[0]).addClass("selected");
+    opts.cuda = visible[0].id;
+  }
+}
+
 
 
 // Change accnone name depending on OS type
diff --git a/scripts/gen_quick_start_module.py b/scripts/gen_quick_start_module.py
index 7654c59ecdf4..1f074da6031c 100755
--- a/scripts/gen_quick_start_module.py
+++ b/scripts/gen_quick_start_module.py
@@ -107,6 +107,8 @@ def get_package_type(pkg_key: str, os_key: OperatingSystem) -> str:
 
 
 def get_gpu_info(acc_key, instr, acc_arch_map):
+    if acc_key not in acc_arch_map:
+        return (None, None)
     gpu_arch_type, gpu_arch_version = acc_arch_map[acc_key]
     if DEFAULT in instr:
         gpu_arch_type, gpu_arch_version = acc_arch_map["accnone"]
@@ -146,6 +148,23 @@ def update_versions(versions, release_matrix, release_version):
                     == (package_type, gpu_arch_type, gpu_arch_version)
                 ]
 
+                if gpu_arch_type is None:
+                    continue
+
+                # A CUDA version the channel ships but this OS does not (e.g. no
+                # Windows build): say so instead of keeping a stale command.
+                if gpu_arch_type == "cuda":
+                    if pkg_arch_matrix:
+                        instr["note"] = None
+                        if package_type == "libtorch" and instr.get("versions") is None:
+                            instr["versions"] = {}
+                    else:
+                        instr["note"] = (
+                            f"<b>NOTE:</b> CUDA {gpu_arch_version} is not available "
+                            f"on {os_key.capitalize()}"
+                        )
+                        instr["command" if package_type != "libtorch" else "versions"] = None
+
                 if pkg_arch_matrix:
                     if package_type != "libtorch":
                         instr["command"] = pkg_arch_matrix[0]["installation"]
@@ -236,8 +255,13 @@ def gen_ver_list(chan, gpu_arch_type):
         rocm_ver_list = gen_ver_list(chan, "rocm")
         cuda_list = sorted(cuda_ver_list.values())
         acc_arch_ver_map[chan]["rocm5.x"] = ("rocm", max(rocm_ver_list.values()))
-        for cuda_ver, label in zip(cuda_list, ["cuda.x", "cuda.y", "cuda.z"]):
+        cuda_labels = ["cuda.x", "cuda.y", "cuda.z"]
+        for cuda_ver, label in zip(cuda_list, cuda_labels):
             acc_arch_ver_map[chan][label] = ("cuda", cuda_ver)
+        # Drop labels the channel has no CUDA version for, so the page strikes
+        # them out instead of showing the hard-coded defaults.
+        for label in cuda_labels[len(cuda_list):]:
+            acc_arch_ver_map[chan].pop(label, None)
 
 
 def main():