diff --git a/web/src/components/FleetRail.tsx b/web/src/components/FleetRail.tsx
index aa5e8af..54b2c42 100644
--- a/web/src/components/FleetRail.tsx
+++ b/web/src/components/FleetRail.tsx
@@ -11,9 +11,10 @@
  * here — it is the existing attach flow, reached from a fleet node.
  */
 
-import { Component, For, Show, createMemo, onCleanup, onMount } from "solid-js";
+import { Component, For, Show, createMemo, createSignal, onCleanup, onMount } from "solid-js";
 
-import { formatCostUsd } from "../lib/format";
+import { formatCostUsd, formatTokens } from "../lib/format";
+import { costShare, fleetEconomics, type FleetEconomics } from "../lib/fleet-economics";
 import {
   groupIntoLanes,
   needsYouCount,
@@ -77,6 +78,16 @@ const FleetRail: Component = () => {
   });
   const attention = createMemo(() => needsYouCount(lanes()));
 
+  // Per-backend spend across the fleet the board actually describes: the
+  // conductor plus the sessions it has dispatched to or spawned (§9's zoom
+  // ABOVE a single session). Collapsed by default — it answers a question you
+  // ask occasionally, and the lanes answer the one you ask constantly.
+  const [showSpend, setShowSpend] = createSignal(false);
+  const econ = createMemo<FleetEconomics>(() => {
+    const b = board();
+    return fleetEconomics(b.conductor ? [b.conductor, ...b.workers] : b.workers);
+  });
+
   return (
     <aside class="flex w-72 shrink-0 flex-col overflow-y-auto border-l border-border bg-bg-elev/40">
       <header class="sticky top-0 z-10 space-y-1 border-b border-border bg-bg-elev/95 px-3 py-2 backdrop-blur">
@@ -101,10 +112,19 @@ const FleetRail: Component = () => {
               {board().agg.blockedTasks} blocked
             </span>
           </Show>
-          <span class="ml-auto text-accent" title="Fleet cost across every backend">
-            {formatCostUsd(board().agg.totalCostUsd)}
-          </span>
+          <button
+            type="button"
+            onClick={() => setShowSpend((v) => !v)}
+            aria-expanded={showSpend()}
+            class="ml-auto rounded px-1 text-accent transition hover:bg-bg-hover"
+            title="Fleet cost across every backend — click for the per-backend split"
+          >
+            {formatCostUsd(board().agg.totalCostUsd)} {showSpend() ? "▾" : "▸"}
+          </button>
         </div>
+        <Show when={showSpend()}>
+          <BackendSpendTable econ={econ()} />
+        </Show>
       </header>
 
       <Show when={board().error}>
@@ -131,6 +151,51 @@ const FleetRail: Component = () => {
   );
 };
 
+/**
+ * Per-backend spend — the metaharness view no single-vendor tool needs (§7).
+ *
+ * The daemon's own `agg.totalCostUsd` stays the headline number rather than a
+ * sum of these rows: it counts tasks that may have aged off this capped board,
+ * so a locally derived total would drift low on a long-lived fleet. This table
+ * answers a different question — how the spend SPLITS — and says so by showing
+ * its own total separately when the two differ.
+ */
+const BackendSpendTable: Component<{ econ: FleetEconomics }> = (props) => (
+  <Show
+    when={props.econ.byBackend.length > 0}
+    fallback={<p class="pt-1 font-mono text-[10px] text-fg-faint">No sessions on the board yet.</p>}
+  >
+    <ul class="flex flex-col gap-0.5 pt-1">
+      <For each={props.econ.byBackend}>
+        {(b) => (
+          <li class="flex items-center gap-2 font-mono text-[10px]">
+            <span class="w-20 shrink-0 truncate text-fg-muted" title={b.providerId}>
+              {b.providerId}
+            </span>
+            {/* A share bar, not a percentage: the question is "which backend is
+                most of the bill", and a bar answers it without arithmetic. */}
+            <span class="h-1 flex-1 overflow-hidden rounded bg-bg">
+              <span
+                class="block h-full bg-accent/60"
+                style={{ width: `${Math.round(costShare(props.econ, b) * 100)}%` }}
+              />
+            </span>
+            <span class="shrink-0 text-fg-faint" title={`${b.sessions} session(s), ${b.active} working`}>
+              {b.active}/{b.sessions}
+            </span>
+            <span
+              class="w-14 shrink-0 text-right text-accent"
+              title={`${formatTokens(b.inputTokens)} in · ${formatTokens(b.outputTokens)} out`}
+            >
+              {formatCostUsd(b.costUsd)}
+            </span>
+          </li>
+        )}
+      </For>
+    </ul>
+  </Show>
+);
+
 const Lane: Component<{ group: FleetLaneGroup }> = (props) => (
   <section class="flex flex-col">
     <h4
diff --git a/web/src/lib/fleet-economics.test.ts b/web/src/lib/fleet-economics.test.ts
new file mode 100644
index 0000000..695d2b1
--- /dev/null
+++ b/web/src/lib/fleet-economics.test.ts
@@ -0,0 +1,113 @@
+import { describe, it, expect } from "vitest";
+
+import { costShare, fleetEconomics, UNKNOWN_PROVIDER } from "./fleet-economics";
+import type { SessionInfo, SessionUsage } from "../protocol/types";
+
+const usage = (over: Partial<SessionUsage> = {}): SessionUsage =>
+  ({
+    inputTokens: 0,
+    outputTokens: 0,
+    cacheReadTokens: 0,
+    cacheCreationTokens: 0,
+    totalCostUsd: 0,
+    numTurns: 0,
+    durationMs: 0,
+    ...over,
+  }) as SessionUsage;
+
+const session = (id: string, over: Partial<SessionInfo> = {}): SessionInfo =>
+  ({ id, name: id, status: "idle", ...over }) as SessionInfo;
+
+describe("fleetEconomics", () => {
+  it("rolls spend up per backend and totals it", () => {
+    const econ = fleetEconomics([
+      session("a", { providerId: "claude", usage: usage({ totalCostUsd: 1, inputTokens: 10 }) }),
+      session("b", { providerId: "claude", usage: usage({ totalCostUsd: 2, outputTokens: 5 }) }),
+      session("c", { providerId: "qwen", usage: usage({ totalCostUsd: 0.5 }) }),
+    ]);
+    expect(econ.byBackend.map((b) => b.providerId)).toEqual(["claude", "qwen"]);
+    expect(econ.byBackend[0]!.costUsd).toBe(3);
+    expect(econ.byBackend[0]!.sessions).toBe(2);
+    expect(econ.byBackend[0]!.inputTokens).toBe(10);
+    expect(econ.byBackend[0]!.outputTokens).toBe(5);
+    expect(econ.totalCostUsd).toBe(3.5);
+    expect(econ.sessions).toBe(3);
+  });
+
+  it("orders by cost descending — the view answers 'what is burning money'", () => {
+    const econ = fleetEconomics([
+      session("cheap", { providerId: "qwen", usage: usage({ totalCostUsd: 0.1 }) }),
+      session("dear", { providerId: "claude", usage: usage({ totalCostUsd: 9 }) }),
+    ]);
+    expect(econ.byBackend[0]!.providerId).toBe("claude");
+  });
+
+  it("breaks cost ties by name so the list does not reshuffle between renders", () => {
+    const econ = fleetEconomics([
+      session("z", { providerId: "zeta", usage: usage({ totalCostUsd: 1 }) }),
+      session("a", { providerId: "alpha", usage: usage({ totalCostUsd: 1 }) }),
+    ]);
+    expect(econ.byBackend.map((b) => b.providerId)).toEqual(["alpha", "zeta"]);
+  });
+
+  it("counts a session with no usage without inventing spend for it", () => {
+    // Ten idle sessions must read as ten sessions at $0, not as an empty fleet.
+    const econ = fleetEconomics([
+      session("a", { providerId: "claude" }),
+      session("b", { providerId: "claude" }),
+    ]);
+    expect(econ.sessions).toBe(2);
+    expect(econ.byBackend[0]!.sessions).toBe(2);
+    expect(econ.totalCostUsd).toBe(0);
+  });
+
+  it("reports a missing providerId as `unknown` rather than folding it into the default", () => {
+    // Attributing spend to a provider that may not have incurred it is worse
+    // than admitting the gap — the whole point is comparing backends.
+    const econ = fleetEconomics([session("a", { usage: usage({ totalCostUsd: 5 }) })]);
+    expect(econ.byBackend[0]!.providerId).toBe(UNKNOWN_PROVIDER);
+    expect(econ.byBackend[0]!.costUsd).toBe(5);
+  });
+
+  it("counts only genuinely in-flight sessions as active", () => {
+    const econ = fleetEconomics([
+      session("a", { providerId: "claude", status: "thinking" }),
+      session("b", { providerId: "claude", status: "tool_running" }),
+      session("c", { providerId: "claude", status: "waiting_approval" }),
+      session("d", { providerId: "claude", status: "idle" }),
+      session("e", { providerId: "claude", status: "error" }),
+    ]);
+    // waiting_approval is stopped, not working — it belongs in the attention
+    // queue, not the concurrency count.
+    expect(econ.active).toBe(2);
+    expect(econ.sessions).toBe(5);
+  });
+
+  it("returns an empty rollup for an empty fleet", () => {
+    const econ = fleetEconomics([]);
+    expect(econ.byBackend).toEqual([]);
+    expect(econ.totalCostUsd).toBe(0);
+    expect(econ.active).toBe(0);
+  });
+});
+
+describe("costShare", () => {
+  it("reports each backend's share of total spend", () => {
+    const econ = fleetEconomics([
+      session("a", { providerId: "claude", usage: usage({ totalCostUsd: 3 }) }),
+      session("b", { providerId: "qwen", usage: usage({ totalCostUsd: 1 }) }),
+    ]);
+    expect(costShare(econ, econ.byBackend[0]!)).toBeCloseTo(0.75);
+    expect(costShare(econ, econ.byBackend[1]!)).toBeCloseTo(0.25);
+  });
+
+  it("is zero when nothing has been spent — never an even split", () => {
+    // Dividing by a zero total to show four backends at 25% would invent a
+    // fact from an absence.
+    const econ = fleetEconomics([
+      session("a", { providerId: "claude" }),
+      session("b", { providerId: "qwen" }),
+    ]);
+    expect(costShare(econ, econ.byBackend[0]!)).toBe(0);
+  });
+});
diff --git a/web/src/lib/fleet-economics.ts b/web/src/lib/fleet-economics.ts
new file mode 100644
index 0000000..67e3e0b
--- /dev/null
+++ b/web/src/lib/fleet-economics.ts
@@ -0,0 +1,122 @@
+/**
+ * Fleet economics — cost and tokens rolled up per backend
+ * (conductor-frontends-design §9, §7).
+ *
+ * §9 asks for the zoom ABOVE a single session: total spend across the fleet,
+ * broken down per backend, so a Claude row and a Gemini row are comparable at a
+ * glance. §7 calls that out as a metaharness differentiator rather than a
+ * nicety — no single-vendor tool needs a per-backend breakdown, because it only
+ * ever has one backend.
+ *
+ * Pure functions: the rollup is the decision worth testing, and it needs no
+ * reactive root.
+ */
+
+import type { SessionInfo } from "../protocol/types";
+
+export interface BackendSpend {
+  /** Provider id as the daemon reports it (`claude`, `gemini-cli`, `qwen`, …). */
+  providerId: string;
+  sessions: number;
+  /** Sessions currently running a turn or a tool. */
+  active: number;
+  inputTokens: number;
+  outputTokens: number;
+  costUsd: number;
+}
+
+export interface FleetEconomics {
+  byBackend: BackendSpend[];
+  totalCostUsd: number;
+  totalInputTokens: number;
+  totalOutputTokens: number;
+  /** Sessions counted in this rollup. */
+  sessions: number;
+  /** How many are working right now — §9's concurrent-agent count. */
+  active: number;
+}
+
+/** Session statuses that mean a turn is actually in flight. */
+const BUSY = new Set(["thinking", "tool_running"]);
+
+/**
+ * Bucket for a session whose backend the daemon did not report.
+ *
+ * Exported so consumers compare against this rather than re-typing the string:
+ * it is a sentinel that reaches the UI, and a caller that wants to style or
+ * filter it should not have to know the literal.
+ */
+export const UNKNOWN_PROVIDER = "unknown" as const;
+
+/**
+ * Backend label for a session.
+ *
+ * A session with no `providerId` is reported as `unknown` rather than silently
+ * folded into the default backend. Attributing spend to a provider that may not
+ * have incurred it is worse than admitting the gap — the entire point of this
+ * view is comparing backends against each other.
+ */
+function backendOf(s: SessionInfo): string {
+  return s.providerId ?? UNKNOWN_PROVIDER;
+}
+
+/**
+ * Roll `sessions` up per backend.
+ *
+ * Sessions with no usage still COUNT (they exist and occupy a slot) but
+ * contribute zero spend, so the session count and the cost stay independently
+ * true. A fleet of ten idle sessions reads as ten sessions at $0, not as an
+ * empty fleet.
+ *
+ * Ordered by cost descending — the question this view answers is "what is
+ * burning money", so the answer is on the first row. Backends that tie fall
+ * back to name order so the list does not reshuffle between renders.
+ */
+export function fleetEconomics(sessions: readonly SessionInfo[]): FleetEconomics {
+  const byId = new Map<string, BackendSpend>();
+
+  for (const s of sessions) {
+    const id = backendOf(s);
+    const row = byId.get(id) ?? {
+      providerId: id,
+      sessions: 0,
+      active: 0,
+      inputTokens: 0,
+      outputTokens: 0,
+      costUsd: 0,
+    };
+    row.sessions += 1;
+    if (BUSY.has(s.status)) row.active += 1;
+    const u = s.usage;
+    if (u) {
+      row.inputTokens += u.inputTokens;
+      row.outputTokens += u.outputTokens;
+      row.costUsd += u.totalCostUsd;
+    }
+    byId.set(id, row);
+  }
+
+  const byBackend = [...byId.values()].sort(
+    (a, b) => b.costUsd - a.costUsd || (a.providerId < b.providerId ? -1 : 1),
+  );
+
+  return {
+    byBackend,
+    totalCostUsd: byBackend.reduce((n, b) => n + b.costUsd, 0),
+    totalInputTokens: byBackend.reduce((n, b) => n + b.inputTokens, 0),
+    totalOutputTokens: byBackend.reduce((n, b) => n + b.outputTokens, 0),
+    sessions: byBackend.reduce((n, b) => n + b.sessions, 0),
+    active: byBackend.reduce((n, b) => n + b.active, 0),
+  };
+}
+
+/**
+ * Each backend's share of total spend, 0–1.
+ *
+ * Zero when nothing has been spent yet — NOT an even split. A fresh fleet has
+ * no shares to report, and dividing by a zero total to show four backends at
+ * 25% would be inventing a fact from an absence.
+ */
+export function costShare(econ: FleetEconomics, backend: BackendSpend): number {
+  return econ.totalCostUsd > 0 ? backend.costUsd / econ.totalCostUsd : 0;
+}