Eleven hours for on top di Atlantic ELEVEN HOURS OVER THE ATLANTIC SEAT 24B WATCHES A FILM AGAIN SEAT 24C REVIEWS FOUR SERVERS GEO ADDS 703 MS TO EVERY KEY MOVE THE WORK NOT THE KEYSTROKES

Eleven hours, from New York go somewhere for east. Di cabin light turn amber. Seat 24B open di screen wey dey behind di seat for im front, scroll four hundred titles, and pick film wey e don already see three times. We no dey criticise am. Na di most sensible thing person fit do wit eleven hours inside aluminium tube wey dem pump air enter, and most of us don do am.

Seat 24C bring out phone.

Di two seat dey on top di same Wi-Fi. Di two of dem dey di same distance from ground. By di time wey di wheel go touch ground, one of dem go don watch film wey e fit already quote, and di odda one go don merge three branches, clear test suite wey dey fail, and restart job wey jam for machine wey dey four thousand miles away.

Di interesting part no be say e possible. Na why e possible — and why di obvious approach, di one wey most people dey try once and abandon somewhere on top Greenland, no dey work at all.

Two different different tins dey wey airline fit mean wen dem talk "Wi-Fi," and na which one dem bolt for your own fuselage go decide everything wey go follow.

Di old arrangement na geostationary satellite wey dem park for 35,786 km. Dat altitude no be design choice wey person make for your sake; na di exact distance wey orbit go take twenty-four hours, so di satellite go look like say e just hang for one place. Di result na arithmetic wey no engineering budget fit argue wit: go up and come back down, two times, na 477 ms round trip for light isef, before even one router enter di matter.

Di new arrangement na low-Earth-orbit constellation wey dey around 550 km — like sixty-five times closer, and 7.3 ms propagation round trip.

Wetin dem measure for air match di physics. Wen dem analyse Ookla im Q1 2025 in-flight speed tests, dem find median round-trip time of 44 ms for Starlink against 703–757 ms for Hughes, Intelsat and Viasat — gap wey be like sixteen times. Anoda team put instrument for real devices across 25 flights on 7 airlines for ACM IMC 2025, and dem measure di same shape by demsef: under 40 ms for LEO, above 550 ms for GEO.

For long-haul ocean crossing, di honest number small worse pass di headline. Di same research measure median of 62 ms on top di Pacific, wey dey range from 19–86 ms. Na dat figure you suppose carry for head. E still correct well well. But e no be 44.

Na here blog post for product site suppose tell you say di future don land.

E never land for everybody di same way, and di aircraft wey e reach last na your own.

Na narrowbody dem dey refit first — dem plenty, dem dey turn around quick, and dem dey fly short domestic hop where nobody dey try do eleven hours of anything. Di widebody wey actually dey cross ocean na dem dey last for di queue. As of mid-2026, United get pass 400 aircraft on Starlink and dem expect say e go near one thousand by year end — but na only around sixty of dem be widebody, dia first Starlink widebody transatlantic flight (Newark–Heathrow) fly only for June 2026, and di full widebody fleet no go ready until summer 2027. Virgin Atlantic don finish dia twelve A350s — around 28% of dia widebody fleet — wit 787s wey dem schedule through di second half of 2026 and A330neos wey go enter 2027.

So: if na inside intercontinental flight you dey read dis one for 2026, di odds still dey against you. Na probably di 700 ms link you dey on top, no be di 62 ms one.

Wey make di rest of dis post more useful, no be less.

Film and shell dey fail for opposite direction, and in-flight Wi-Fi na exactly di wrong network for one of dem.

Film na bandwidth problem wey no send latency. E dey buffer. If di first frame land 700 ms late, nobody go notice, because nothing about di experience depend on say di film go react to you. Press play, wait, watch. Di round trip happen once.

Interactive shell na di opposite of am. E no need bandwidth at all — dem dey measure terminal session for kilobits, no be megabits — but e dey pay di round trip for every single keystroke wey you expect make e echo back. For 700 ms, dat one no be slow connection. Na conversation wey get two-third-of-a-second pause afta every character.

Na dis one people dey get wrong. Dem go see "50 Mbps" for di in-flight portal, conclude say di connection correct, open terminal, and discover say "correct" no get anything to do wit am. Bandwidth no be di currency here. Na round trips be am.

Di failure specific pass "e dey feel slow," and di research clear well well about where di pain dey land.

For one controlled study of 31 participants wey compare 20 ms against 200 ms end-to-end input latency, bulk typing barely move: 290.9 ms per character against 302.7, difference wey di authors no fit call significant. Error rate be di same.

But di correction task — to reposition cursor wit arrow keys and fix text, di thing wey you actually dey do inside editor — jump from 38.9 seconds go 46.5 seconds, like 20% slower, wit very big effect size (p < 0.001, d = 1.673). Di authors talk di mechanism plain: wen you dey navigate, you must dey wait for di screen, because where di cursor land na im dey decide wetin you go do next.

Di load wey person dey feel worse even where speed no change. For di typing task, where performance no change statistically, di vex wey dem report demsef still rise from 20.9 go 31.3.

Carry dat one enter terminal and e go name exactly wetin dey scatter for up there: no be to type command comot. Na vim navigation, tab completion, to scroll through less, and di poke-and-look rhythm of REPL. Every single one of dem na loop where di next thing wey you go do depend on wetin just come back.

Once you accept say na round trips be di resource wey scarce, di fix stop to be "find faster plane" and e turn architecture question: take fewer round trips, and make each one carry more.

One well-known tool dey wey face dis matter head-on. Mosh dey predict wetin your keystrokes go do and e dey echo dem locally before di server confirm am. Di effect wey dem measure big well well: for link wey get like half second round-trip time, median keystroke response fall from 503 ms go under 5 ms. For trans-oceanic path from MIT go Singapore, 273 ms go under 5 ms.

But read di small print, because na im be di whole point. Mosh dey predict like 70% of keystrokes. Di around 30% wey e no fit predict na, for di authors dia own word, principally navigation — and for dose ones, im latency profile go look like plain SSH again. Wetin dat mean be say: di one interaction wey high latency dey punish pass na exactly di one wey local echo no fit save.

So prediction dey help, but e no dey save you. Di thing wey actually dey save you na to refuse to dey inside di loop dat often.

Stop to dey type enter latency. Put work wey go run long for di far end and just dey check am. Build, test suite, migration, coding agent — none of dem need make you dey present for di round trip. Dem need you for di beginning and for di decision. Everything wey dey between na di machine im own problem, and di machine dey for ground wit good connection.

Na also why parallelism dey beat interactivity for up there. You no fit type quick through 700 ms. But you fit supervise four servers at once well well, because supervision no be latency-bound — na attention-bound, and you get eleven hours of attention and nothing else to spend am on.

To talk am concrete, for phone, on top bad link:

Dem no invent any of dis one for aircraft. Na just how remote work over hostile link dey always look, and plane na just hostile link wey honest pass di rest.

Two things dey wey we no fit verify, and we no go form say we verify dem:

Whether your airline im network go even allow SSH comot at all. We no see any reliable public evidence about port blocking, VPN policy or idle timeout across airline Wi-Fi providers — and plenty confident forum claim dey wey no dey survive checking. E dey vary by carrier, by provider and by aircraft. Test am for short flight before you plan long one around am.

How di cabin dey actually spend im time. Di in-flight entertainment statistics wey people dey quote pass turn out say na advertising rate card dem trace back to, wey no get methodology, so we throw dem comot. Di best survey evidence wey we see — Inmarsat/Censuswide, 11,231 passengers across ten countries, even though na 2022 dem field am and na intent dem ask about instead of measuring time — put entertainment way ahead of work: 41% wan watch film wey dem don download and 35% wan stream, against 26% wey wan do work.

So treat seat 24B as character, no be statistic. Di point no be say to watch film bad. Na say di alternative no dey unavailable again.

Di film end for di same place wey e end di last three times. Di plane don start to dey come down. Seat 24B don get flight wey correct well well, and e go land wit rest, wey get value.

Seat 24C land wit di work don finish — because di work never dey inside di seat. E dey for machine wey dey ground di whole time, and na window to look am di phone be from start to finish.

Eleven hours na long time to spend for 700 ms. But e na very short time to spend supervising four servers wey no send di satellite at all.