diff --git a/proto-cuda/windows-miner/start-mining.ps1 b/proto-cuda/windows-miner/start-mining.ps1
index 877071a91..405b4f8d1 100644
--- a/proto-cuda/windows-miner/start-mining.ps1
+++ b/proto-cuda/windows-miner/start-mining.ps1
@@ -116,9 +116,9 @@ if ($haveCl) {
     $libFirst = if ($env:LIB) { ($env:LIB -split ';')[0] } else { 'EMPTY' }
     $incFirst = if ($env:INCLUDE) { ($env:INCLUDE -split ';')[0] } else { 'EMPTY' }
     Log "MSVC environment: cl at $((Get-Command cl.exe).Source); LIB starts with $libFirst; INCLUDE starts with $incFirst"
-    # Always add the newest MSVC lib dir and the Windows Kits dirs in front of LIB: a vcvars import can leave LIB without
-    # the MSVC lib directory (seen on Josh's PC: LIB started with ucrt and nvcc's link failed on LIBCMT.lib).
-    [void](Set-LibFallback); Log ('LIB now starts with ' + (($env:LIB -split ';')[0]))
+    # The fallback only when the import left nothing: it would otherwise put the newest toolset's headers in front of
+    # INCLUDE and CUDA 12.8's cudafe++ crashes on those. The library path for nvcc's link step comes from build.bat itself.
+    if (-not $env:LIB) { Log 'LIB is empty: trying the fallback'; [void](Set-LibFallback) }
 }
 $haveNvcc = [bool](Get-Command nvcc.exe -ErrorAction SilentlyContinue)
 $openclSdk = $null