diff --git a/eth/fetcher/blob_fetcher.go b/eth/fetcher/blob_fetcher.go
index 6226b89820..1ba0c1c232 100644
--- a/eth/fetcher/blob_fetcher.go
+++ b/eth/fetcher/blob_fetcher.go
@@ -50,6 +50,14 @@ const (
 	maxPayloadAnnounces           = 4096
 	fetchProbability              = 15
 	MAX_CELLS_PER_PARTIAL_REQUEST = 8
+
+	// maxCellRequests caps the burst of cell requests we can issue at once
+	// to a single peer. Worst case 256 * 6 = 1536 cells (~3 MB)
+	maxCellRequests = 256
+	// refillInterval is the gap between token refill. Combined with
+	// maxCellRequests and 2-minute buffer timeout, a peer's worst case
+	// buffered cells is about 16 MB
+	refillInterval = time.Second / 9
 )
 
 type blobTxAnnounce struct {
@@ -137,12 +145,21 @@ type BlobFetcher struct {
 
 	fn BlobFetcherFunctions // callbacks
 
+	// peerTokens tracks each peer's remaining cell request token.
+	peerTokens map[string]*token
+
 	step     chan struct{}    // Notification channel when the fetcher loop iterates
 	clock    mclock.Clock     // Monotonic clock or simulated clock for tests
 	realTime func() time.Time // Real system time or simulated time for tests
 	rand     random           // Randomizer
 }
 
+// token is a per peer token bucket for outgoing cell requests.
+type token struct {
+	amount int64
+	last   mclock.AbsTime
+}
+
 func NewBlobFetcher(fn BlobFetcherFunctions, custody *types.CustodyBitmap, rand random) *BlobFetcher {
 	return &BlobFetcher{
 		notify:     make(chan *blobTxAnnounce),
@@ -158,6 +175,7 @@ func NewBlobFetcher(fn BlobFetcherFunctions, custody *types.CustodyBitmap, rand
 		fetches:    make(map[common.Hash]*fetchStatus),
 		requests:   make(map[string][]*cellRequest),
 		alternates: make(map[common.Hash]map[string]*types.CustodyBitmap),
+		peerTokens: make(map[string]*token),
 		fn:         fn,
 		custody:    custody,
 		clock:      mclock.System{},
@@ -574,6 +592,7 @@ func (f *BlobFetcher) loop() {
 			f.scheduleFetches(timeoutTimer, timeoutTrigger, nil) // Partial delivery may enable others to deliver too
 		case drop := <-f.drop:
 			// A peer was dropped, remove all traces of it
+			delete(f.peerTokens, drop.peer)
 			if _, ok := f.waitslots[drop.peer]; ok {
 				for hash := range f.waitslots[drop.peer] {
 					delete(f.waitlist[hash], drop.peer)
@@ -682,6 +701,34 @@ func (f *BlobFetcher) rescheduleTimeout(timer *mclock.Timer, trigger chan struct
 		trigger <- struct{}{}
 	})
 }
+
+// consumeToken consumes n tokens from peer's cell-request budget.
+// It returns false if the remaining tokens cannot cover n.
+func (f *BlobFetcher) consumeToken(peer string, n int) bool {
+	now := f.clock.Now()
+	b, ok := f.peerTokens[peer]
+	if !ok {
+		b = &token{amount: maxCellRequests, last: now}
+		f.peerTokens[peer] = b
+	} else {
+		// Here, fractional remaining elapsed time is left in b.last
+		// so that it can be carried over to the next call
+		elapsed := time.Duration(now - b.last)
+		if add := int64(elapsed / refillInterval); add > 0 {
+			b.amount += add
+			if b.amount > maxCellRequests {
+				b.amount = maxCellRequests
+			}
+			b.last += mclock.AbsTime(time.Duration(add) * refillInterval)
+		}
+	}
+	if b.amount < int64(n) {
+		return false
+	}
+	b.amount -= int64(n)
+	return true
+}
+
 func (f *BlobFetcher) scheduleFetches(timer *mclock.Timer, timeout chan struct{}, whitelist map[string]struct{}) {
 	// Gather the set of peers we want to retrieve from (default to all)
 	actives := whitelist
@@ -715,8 +762,10 @@ func (f *BlobFetcher) scheduleFetches(timer *mclock.Timer, timeout chan struct{}
 				unfetched = cells.Difference(f.fetches[hash].fetching)
 			}
 
-			// Mark fetching for unfetched cells
-			if unfetched.OneCount() > 0 {
+			// Mark fetching for unfetched cells if the peer has enough token.
+			// Otherwise, the next peer who announced the hash and has token will be selected
+			// in the next loop
+			if unfetched.OneCount() > 0 && f.consumeToken(peer, int(unfetched.OneCount())) {
 				if f.fetches[hash] == nil {
 					f.fetches[hash] = &fetchStatus{
 						fetching: unfetched,