diff --git a/performance-indexes.sql b/performance-indexes.sql
new file mode 100644
index 0000000..482a485
--- /dev/null
+++ b/performance-indexes.sql
@@ -0,0 +1,26 @@
+-- Performance indexes for server-recommend
+-- Run these on the cloud-instances-db D1 database
+
+-- instance_types: Optimize filtering by provider and specs
+CREATE INDEX IF NOT EXISTS idx_instance_types_provider_specs
+ON instance_types(provider_id, vcpu, memory_mb);
+
+-- pricing: Optimize joins with instance_types and regions
+CREATE INDEX IF NOT EXISTS idx_pricing_instance_region
+ON pricing(instance_type_id, region_id);
+
+-- pricing: Optimize price-based sorting
+CREATE INDEX IF NOT EXISTS idx_pricing_price
+ON pricing(monthly_price_retail, monthly_price_krw);
+
+-- regions: Optimize region code lookups
+CREATE INDEX IF NOT EXISTS idx_regions_codes
+ON regions(region_code, country_code);
+
+-- vps_benchmarks: Optimize spec-based lookups
+CREATE INDEX IF NOT EXISTS idx_vps_benchmarks_specs
+ON vps_benchmarks(vcpu, memory_gb, provider_name);
+
+-- vps_benchmarks: Optimize performance queries
+CREATE INDEX IF NOT EXISTS idx_vps_benchmarks_performance
+ON vps_benchmarks(geekbench_single DESC, monthly_price_usd);
diff --git a/src/index.ts b/src/index.ts
index 9fbc627..50af0be 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -997,9 +997,23 @@ async function handleRecommend(
     const bandwidthEstimate = estimateBandwidth(body.expected_users, body.use_case, body.traffic_pattern);
     console.log(`[Recommend] Bandwidth estimate: ${bandwidthEstimate.monthly_tb >= 1 ? bandwidthEstimate.monthly_tb + ' TB' : bandwidthEstimate.monthly_gb + ' GB'}/month (${bandwidthEstimate.category})`);
 
-    // Phase 2: Query candidate servers (depends on minMemoryMb, minVcpu, bandwidth)
-    const candidates = await queryCandidateServers(env.DB, body, minMemoryMb, minVcpu, bandwidthEstimate, lang);
+    // Estimate specs for VPS benchmark query (doesn't need exact candidates)
+    const estimatedCores = minVcpu || 2;
+    const estimatedMemory = minMemoryMb ? Math.ceil(minMemoryMb / 1024) : 4;
+    const defaultProviders = bandwidthEstimate?.category === 'very_heavy' ? ['Linode'] : ['Linode', 'Vultr'];
+
+    // Phase 2: Query candidate servers and VPS benchmarks in parallel
+    const [candidates, vpsBenchmarks] = await Promise.all([
+      queryCandidateServers(env.DB, body, minMemoryMb, minVcpu, bandwidthEstimate, lang),
+      queryVPSBenchmarksBatch(env.DB, estimatedCores, estimatedMemory, defaultProviders).catch((err: unknown) => {
+        const message = err instanceof Error ? err.message : String(err);
+        console.warn('[Recommend] VPS benchmarks unavailable:', message);
+        return [] as VPSBenchmark[];
+      }),
+    ]);
+
     console.log('[Recommend] Candidate servers:', candidates.length);
+    console.log('[Recommend] VPS benchmark data points:', vpsBenchmarks.length);
 
     if (candidates.length === 0) {
       return jsonResponse(
@@ -1013,28 +1027,9 @@ async function handleRecommend(
       );
     }
 
-    // Calculate average specs from candidates for VPS benchmark queries
-    const avgCores = Math.round(
-      candidates.reduce((sum, s) => sum + s.vcpu, 0) / candidates.length
-    );
-    const avgMemory = Math.round(
-      candidates.reduce((sum, s) => sum + s.memory_gb, 0) / candidates.length
-    );
-
     // Use initially fetched benchmark data (already filtered by tech stack)
     const benchmarkData = benchmarkDataAll;
 
-    // Get unique providers from candidates
-    const providers = [...new Set(candidates.map((c) => c.provider_name))];
-    console.log('[Recommend] Providers:', providers);
-
-    // Query VPS benchmarks using consolidated single query (includes provider-specific and spec-based matching)
-    const vpsBenchmarks = await queryVPSBenchmarksBatch(env.DB, avgCores, avgMemory, providers).catch(err => {
-      console.warn('[Recommend] VPS benchmark data unavailable:', err.message);
-      return [] as VPSBenchmark[];
-    });
-    console.log('[Recommend] VPS benchmark data points:', vpsBenchmarks.length);
-
     // Use OpenAI GPT-4o-mini to analyze and recommend (techSpecs already queried above)
     const aiResult = await getAIRecommendations(
       env,