diff --git a/src/server/billing/autumn.ts b/src/server/billing/autumn.ts
index 3785f85..be9dbe6 100644
--- a/src/server/billing/autumn.ts
+++ b/src/server/billing/autumn.ts
@@ -6,13 +6,20 @@ export const autumn = new Autumn({
   // Retries 429/500/502/503/504 (per-operation retryCodes) plus connection
   // errors. Cloudflare 52x statuses are not in the SDK's retry list, so those
   // still surface immediately.
+  //
+  // These reads/gates (customers.getOrCreate, check) sit on the request hot
+  // path, so keep the total retry window short: when Autumn is rate-limiting or
+  // slow, an 8s backoff window held the isolate open for seconds on every
+  // request and cascaded into region-wide congestion (incident 2026-07-06).
+  // Fail fast instead — a caller that can't gate surfaces an error rather than
+  // hanging.
   retryConfig: {
     strategy: "backoff",
     backoff: {
       initialInterval: 250,
-      maxInterval: 2000,
+      maxInterval: 1000,
       exponent: 1.5,
-      maxElapsedTime: 8000,
+      maxElapsedTime: 2500,
     },
     retryConnectionErrors: true,
   },
diff --git a/src/server/billing/subscription.test.ts b/src/server/billing/subscription.test.ts
index e06059e..cad82a8 100644
--- a/src/server/billing/subscription.test.ts
+++ b/src/server/billing/subscription.test.ts
@@ -1,9 +1,15 @@
 import { beforeEach, describe, expect, it, vi } from "vitest";
 import { AUTUMN_PAID_PLAN_FEATURE_ID } from "@/shared/billing";
 
-const { checkMock, getOrCreateMock } = vi.hoisted(() => ({
+const { checkMock, getOrCreateMock, kvGetMock, kvPutMock } = vi.hoisted(() => ({
   checkMock: vi.fn(),
   getOrCreateMock: vi.fn(),
+  kvGetMock: vi.fn(),
+  kvPutMock: vi.fn(),
+}));
+
+vi.mock("cloudflare:workers", () => ({
+  env: { KV: { get: kvGetMock, put: kvPutMock } },
 }));
 
 vi.mock("@/server/billing/autumn", () => ({
@@ -33,6 +39,7 @@ import {
 describe("subscription billing", () => {
   beforeEach(() => {
     vi.clearAllMocks();
+    kvGetMock.mockResolvedValue(null);
   });
 
   it("checks the paid plan entitlement", async () => {
@@ -65,5 +72,19 @@ describe("subscription billing", () => {
       customerId: "org_123",
       email: "alice@example.com",
     });
+    expect(kvPutMock).toHaveBeenCalled();
+  });
+
+  it("skips the Autumn round trip when the customer was recently ensured", async () => {
+    kvGetMock.mockResolvedValue("1");
+
+    const result = await getOrCreateOrganizationCustomer({
+      organizationId: "org_123",
+      userId: "user_123",
+      userEmail: "alice@example.com",
+    });
+
+    expect(result).toEqual({ id: "org_123" });
+    expect(getOrCreateMock).not.toHaveBeenCalled();
   });
 });
diff --git a/src/server/billing/subscription.ts b/src/server/billing/subscription.ts
index 440e165..918de8b 100644
--- a/src/server/billing/subscription.ts
+++ b/src/server/billing/subscription.ts
@@ -1,3 +1,4 @@
+import { env } from "cloudflare:workers";
 import type { EnsuredUserContext } from "@/middleware/ensure-user/types";
 import {
   AUTUMN_MANAGED_ACCESS_FEATURE_ID,
@@ -20,9 +21,24 @@ export type BillingCustomerContext = Pick<
   projectId?: string;
 };
 
+// Existence is monotonic and the Autumn customer id is always the org id we
+// pass, so once we've confirmed a customer exists we can skip the round trip
+// and reuse the org id. Callers only need `.id` (they read balances via
+// `check`), and a degraded Autumn API otherwise added seconds to every hot-path
+// request that ensured the customer (incident 2026-07-06). Long TTL is safe:
+// we only ever cache confirmed existence, never absence.
+const CUSTOMER_ENSURED_TTL_SECONDS = 24 * 60 * 60;
+const customerEnsuredKey = (organizationId: string) =>
+  `autumn:customer-ensured:${organizationId}`;
+
 export async function getOrCreateOrganizationCustomer(
   context: BillingCustomerContext,
-) {
+): Promise<{ id: string }> {
+  const cacheKey = customerEnsuredKey(context.organizationId);
+  if (await env.KV.get(cacheKey)) {
+    return { id: context.organizationId };
+  }
+
   const customer = await autumn.customers.getOrCreate({
     customerId: context.organizationId,
     email: context.userEmail,
@@ -32,10 +48,11 @@ export async function getOrCreateOrganizationCustomer(
     throw new AppError("INTERNAL_ERROR", "Failed to resolve billing customer");
   }
 
-  return {
-    ...customer,
-    id: customer.id,
-  };
+  await env.KV.put(cacheKey, "1", {
+    expirationTtl: CUSTOMER_ENSURED_TTL_SECONDS,
+  });
+
+  return { id: customer.id };
 }
 
 export async function customerHasPaidPlan(customerId: string) {