The Hidden Pockets of Packet Loss: Where the Internet Quietly Drops Your Data
Almost every "internet quality" map you've seen is really a speed map wearing a disguise. It shades the world by download throughput, hands South Korea and the Gulf states a victory lap of green, and moves on. For figuring out where your connection will actually let you down, those maps are close to worthless.
Speed is a best-case number. It tells you how fast one connection can move data once it's up and behaving itself. What it doesn't tell you is whether the connection completes at all, whether it survives a forty-minute video call, or whether it's quietly losing one packet in twenty every evening when the neighborhood gets home. That second bucket is packet loss, and it's the thing that ruins games, calls, and remote work. It also happens to be spread across the planet in a pattern that barely resembles the speed map.
So I went looking for the places where data goes to die. Some of them you could probably guess. A few genuinely surprised me. The thread running through all of them is that where a connection breaks usually tells you why it broke, which turns out to be the most useful idea in the whole exercise. Let me hand you that idea first, because everything after it makes more sense once you have it.
The Trick: Failures Have Fingerprints
A TCP connection has a shape. There's a handshake to open it, data in the middle, and a clean close at the end. When a connection dies, the moment in that sequence when it died is a fingerprint, and Cloudflare happens to publish this data through its Radar platform. A 2026 analysis of 83 countries by TechnologyChecker turned those numbers into a tidy diagnostic, and the headline figure is worth sitting with: globally, only about 80% of TCP connections complete cleanly. One in five fails before it finishes, and your speed test will cheerfully hide every single one of them.
There are three failure signatures that matter, and they point at three completely different culprits.
When a connection dies right after the handshake, you're almost always looking at a firewall or some other middlebox throwing a reset packet to kill the thing on purpose. That's a policy decision, not a hardware limitation. Censorship and filtering gear doing exactly what someone configured it to do.
When a connection dies after the handshake succeeds, after data has started flowing, the usual suspect is stateful equipment buckling under its own load. Carrier-grade NAT, deep-packet-inspection boxes, session trackers. They open the connection fine, then drop it partway through because their memory of who's connected to what fills up and spills over.
And when a connection dies deep in an already-established stream, that's the boring, honest kind of packet loss: the network just couldn't carry the traffic. Congestion, buffer bloat, a last-mile link that's been oversubscribed for years.
Hold onto those three, because each pocket below is dominated by one of them, and a fix for one does nothing for the others. People waste a lot of money learning that the hard way.
Cable Chokepoints
The internet isn't in the cloud. It's at the bottom of the ocean, in a few hundred fiber cables, and a frankly alarming share of global traffic gets funneled through a handful of geographic pinch points.
The Red Sea is the obvious one. Something like 15 submarine cables thread through the narrow Bab el-Mandeb strait between East Africa and the Arabian Peninsula. That's the cord tying Europe to Asia and East Africa, and in September 2025 several of those cables got cut, most likely by a ship dragging its anchor across the seabed. (Anchor drags cause roughly 30% of all cable faults, according to the International Cable Protection Committee, so this is a feature of the system, not a freak event.) Traffic didn't stop. It rerouted, which is what's supposed to happen. ThousandEyes caught packet loss showing up in Riyadh afterward, because Saudi traffic that normally reached the sea through Jeddah suddenly had to take the long way around. That's the catch with rerouting: it keeps you online by swapping a clean short path for a saturated long one, and loss tends to live on the long one.
Island nations don't get to treat this as an occasional problem. Comparitech counted ten island nations with no direct continental cable backup, and several more leaning on a single international cable system, which means one bad afternoon can take an entire country dark. Tonga is the example everyone cites for a reason. In 2022 an undersea volcano erupted and snapped its one cable, and the country was off the fiber internet for more than five weeks. Even Taiwan, which is far better wired with a couple dozen cables, keeps having the lines that feed its smaller Matsu and Penghu islands sliced by passing vessels, stranding residents for weeks at a time.
The point for our map is that a region can post perfectly respectable average numbers and still be sitting one dragged anchor away from a miserable month. Redundancy is what separates sturdy internet from fragile internet, and it never shows up on a speed test.
The Trombone Problem in Sub-Saharan Africa
If you want the most structurally lossy region on the map right now, it's sub-Saharan Africa, and the reason behind it is one of those failures that's almost beautiful once you understand it.
Picture two ISPs, one in Nairobi and one in Lagos, that need to swap traffic. You'd assume the data travels between Nairobi and Lagos. A lot of the time it doesn't. African networks have a long history of peering with each other at European exchange points, in London and Amsterdam, rather than connecting locally. So a packet headed from one African city to another can ride all the way up to Europe and back down again. It slides out and back like a trombone, which is exactly what the engineers who study this call it.
The cost is real. The Internet Society, looking at Kenya and Nigeria, found ISPs reporting latencies north of 200 milliseconds and sometimes pushing 600 when there was no local exchange point to keep domestic traffic at home. Studies of intra-Africa routing keep turning up traceroutes between two African endpoints that detour through Europe or North America for several hops before circling back. Every extra hop is another queue, another buffer, another opportunity to drop something, and the continental TCP healthy rate shows the damage. Africa averages well under 60% in the Cloudflare data, more than twenty points behind every other region on earth.
The good news is that this is the most fixable pocket on the whole list, and people are fixing it. Exchange points like LINX Accra and LINX Mombasa, both opened in 2025, exist specifically to keep local traffic local, and where they catch on, latency and loss come down while the bills for international transit shrink. One regional analysis framed Africa's bottleneck as a political and coordination failure more than a technical one, and that rings true. The fiber is increasingly there. The handshakes between networks to use it well sometimes aren't.
The Landlocked Trap
Take "depends on submarine cables" and stack "has no coastline" on top of it, and you arrive at the quietest pocket of the bunch. There are 44 landlocked countries, and every one of them reaches the global internet through a neighbor, which means its connection quality is hostage to that neighbor's infrastructure, and occasionally to that neighbor's mood.
The single-point-of-failure stories here are wild once you start collecting them. Bhutan has historically routed its connectivity through one town in India, Siliguri, a chokepoint sitting outside its own borders that it can't do much about. Armenia, which has no diplomatic relations with two of its neighbors, has watched its internet vanish for hours because a fiber backbone broke in Georgia. Chad currently hangs on a single working link through Cameroon, after its backup route through Sudan went dead courtesy of that country's war. The Democratic Republic of Congo shares a border with Rwanda, and Rwanda's backbone reaches the line, but the connection only works in one direction because the DRC never built its own backbone out to meet it.
On top of all that fragility, landlocked countries pay what amounts to a toll. The intermediary nations charge to carry the traffic, so the connectivity ends up both pricier and shakier. It's no accident that so many of these countries keep satellite around as a backup. When your only terrestrial path is a single thread somebody else is holding, you want a plan B that doesn't run through their territory.
The Satellite Fringe
Anywhere fiber gives up, which is to say the open ocean, the deep desert, mountain villages, polar stations, and ships, the internet shows up from orbit instead, and orbit brings its own kind of loss.
The old approach is geostationary satellites parked 36,000 kilometers overhead. The geometry is merciless: a round trip to that altitude and back costs you 600-plus milliseconds before anything useful happens, which is enough on its own to make gaming or a video call an exercise in patience. Then there's rain fade. At the Ku and Ka frequency bands these systems use, raindrops are close enough in size to the signal's wavelength that they scatter it, so a heavy storm can knock the link down. Operators build in a fade margin and adaptive coding to fight it, but a tropical downpour is still a packet-loss event you can set your watch by.
Low-earth-orbit constellations like Starlink solve the latency problem by flying much closer, dropping ping into the 20-to-60-millisecond range, and they shrug off weather better, with packet loss reportedly under 1% in clear skies. What they add is a brand-new loss source: the handover. As one satellite slips out of view and the link jumps to the next one, there's a hiccup, and maritime testing clocks anywhere from half a second to two seconds of degraded performance per handover. Above 60 degrees north the satellites are spread thinner, so coverage gets patchy. And on the consumer plans, "best effort" is the polite phrase for the fact that your bandwidth can quietly get deprioritized during busy hours with nothing in the contract to stop it.
So the satellite fringe isn't one pocket so much as a sliding scale. Geostationary users pay in latency and rain. LEO users pay in handover stutters and rush-hour throttling. And the regions that have only satellite are, by definition, the regions fiber decided weren't worth the trench.
The Middlebox Tax
Back to that second fingerprint, the connections that open fine and then keel over. This is the dominant failure across a tight little cluster of countries: Kenya, Uganda, Tanzania, Morocco, South Africa, and, for darker reasons, Myanmar. Uganda and Myanmar each see better than a third of their connections reset after the handshake, which is a staggering number once you let it sink in.
This usually isn't censorship and it isn't a shortage of raw bandwidth. It's stateful equipment running past what it was specced for. Carrier-grade NAT (the gear that lets a crowd of users share a thin supply of IPv4 addresses), deep-packet-inspection boxes, and session-tracking middleboxes all have to remember the state of every connection passing through. Pile on enough connections and those state tables overflow or the timeouts misfire, and the box starts dropping live sessions without a word. The geographic clustering is the giveaway. When five neighboring countries fail the same way, you're looking at shared infrastructure habits, common vendors, similar deployment patterns, rather than a handful of unlucky ISPs.
What's instructive about this pocket is how stubborn it is. Across a six-week re-measurement in the Cloudflare data, this cluster barely twitched, which is the signature of an infrastructure problem rather than a policy one. You don't fix an overloaded NAT pool with a config tweak overnight. That makes it the chronic condition of the packet-loss world, the ache that doesn't go away, as opposed to the next pocket, which can clear up almost as fast as it appeared.
The Filtering Fingerprint
The first fingerprint, resets the instant a connection tries to open, is the censorship-and-filtering signature, and it throws off some of the most dramatic numbers on the map. It also comes with a caveat I want to be honest about.
China is the steady example, posting handshake-stage reset rates that line up with years of documented Great Firewall behavior against traffic bound for filtered destinations. The real eyebrow-raiser in the 2026 data was Bulgaria, an EU member, whose handshake-reset rate got worse over the year until it rivaled or beat China's, while its neighbors Serbia and Romania sat at perfectly normal levels. If you route traffic through Bulgarian transit, that's worth a look.
Here's the caveat, and it's a big one for anyone trying to draw conclusions from a single snapshot. Filtering loss is often temporary. Early in 2026, Iraq briefly posted the worst connection quality on the planet, with nearly two of every three connections dying at the handshake, and then recovered to above the global average six weeks later, the interference simply gone. Brazil pulled the same disappearing act on a smaller scale. Whatever was injecting the resets stopped. So a single measurement window catches filtering while it's happening, which is an event and not a permanent trait of a country. Anyone who wrote Iraq off as hopeless in April would have looked foolish by June. Measure twice before you reroute the world around a place.
The Congestion Hour
The last pocket is the one closest to wherever you're sitting right now. That third fingerprint, loss deep in an established stream, is plain congestion, and it shows up in places you wouldn't put on a "bad internet" list: Ecuador, Peru, Ukraine, Bangladesh, and a surprising run of EU members including Slovenia, Croatia, Hungary, and Greece. Ecuador's mid-stream loss rate runs more than double the global average.
None of these are filtered or censored networks. They've got congested last-mile links and skimpy peering, and the loss tends to keep a daily schedule, invisible at 3 a.m. and ugly at 9 p.m. when everyone's streaming at once. Its close relative is buffer bloat, where oversized buffers in your router or modem hoard excess packets in long queues instead of dropping them promptly. The connection barely loses anything, but latency under load swells up, and anything time-sensitive suffers as though packets had gone missing. It's the failure that hides best from speed tests, because the pipe really is fat. It just gags the second you put it to work.
The cure for this pocket is the least exciting and the most expensive one on the list: more fiber, better peering, smarter queue management. No routing trick or policy change makes congestion evaporate. Somebody has to dig.
Why the Map Looks the Way It Does
Stand back from all seven and the world's packet-loss map sorts itself into a few overlapping forces, none of which reduces neatly to "this country is poor."
Some of it is geography nobody can move: chokepoint straits, island isolation, landlocked dependence on the neighbors. Some of it is relationships that never formed, the trombone belt where the fiber is in the ground but the local interconnection isn't. Some of it is equipment shoved past its limits, the middlebox tax of saturated NAT and DPI gear. Some of it is deliberate interference, the filtering that's real but flightier than it looks in a single reading. And a lot of it is ordinary congestion, the universal tax that bites hardest wherever last-mile investment fell behind demand.
That gap is why the speed map and the loss map argue with each other so loudly. Saudi Arabia tops mobile speed charts and lands in the bottom ten for connection reliability, so its users get blistering downloads sprinkled with dropped sessions. Vietnam, running a fraction of South Korea's broadband speed, beats it on connection integrity because it spent its money on cable diversity and domestic peering. Speed measures how wide the pipe is. Loss measures whether the pipe holds while you're using it. They are not the same question, and most of the time they don't have the same answer.
What to Do With All This
If you run anything global, a game server, a SaaS product, a video platform, the loss map should weigh on your decisions more than the speed map does. The playbook that falls out of it is fairly consistent. Terminate connections at an edge close to your users, so there are fewer hops where a middlebox can stick its nose in. Lean on HTTP/3, which runs over UDP and slides past a lot of the TCP-killing gear without engaging it. And measure your reliability from where your users actually live, not from inside your own cloud region, because a ping out of us-east-1 will never feel what a player in Nairobi or Karachi puts up with every night.
If you're just trying to work out which pocket you're stuck in, the fingerprints scale all the way down to your own router. Loss that comes and goes with the clock is congestion. Latency that balloons only when the line is busy is buffer bloat. Connections that won't even open are pointing upstream at filtering or a middlebox that's having a bad day. The internet drops your data for specific, diagnosable reasons, and naming the reason is most of the battle. A good next step is to run a packet loss test from your own connection and see which fingerprint, if any, shows up on yours.
Sources and Further Reading
Cloudflare Radar TCP connection-quality data, as analyzed by TechnologyChecker in 2026; Internet Society reporting on internet exchange points and landlocked developing countries; ThousandEyes' analysis of the 2025 Red Sea cable cuts; Comparitech's study of island-nation cable vulnerability; and academic work on circuitous routing in Africa from AFRINIC and others. The figures come from measurement windows in 2025 and 2026, and as the Iraq and Brazil examples show, they can move fast. Treat them as snapshots rather than verdicts.