The best run in gambling history ended at zero

In December 1992 a Greek immigrant and former waiter named Archie Karas drove to Las Vegas with $50 in his pocket. He had just lost almost everything he owned in a poker game in Los Angeles.

Over the next two and a half years he ran that stake, plus a $10,000 loan he repaid within hours, up to a peak of around $40 million. He beat Doyle Brunson. He beat Stu Ungar. He beat Chip Reese. At one point he was reportedly holding most of Binion's supply of $5,000 chips because there were not enough left in the building for anyone else.

In 1995 he lost $30 million of it in three weeks. Eleven million at craps, then two million back to Reese, then seventeen million at baccarat. He went to Greece, came back, and lost the remaining twelve. He finished where he started, which is to say at zero.

That is the most successful gambling run anyone has ever documented, and it ended at exactly the same number as the least successful one.

This is why.

The oldest question in the subject

The maths that explains Karas is older than the United States.

In 1656, Blaise Pascal wrote to Pierre de Fermat with a problem about two players betting against each other until one of them had nothing left. Christiaan Huygens published a version of it in 1657, as the last of five problems he left for the reader at the end of the first printed textbook on probability. Abraham de Moivre gave the general solution in 1712.

It is called the gambler's ruin problem, and it is not a warning or a moral. It is a theorem.

Two players bet repeatedly on an even-money proposition. Neither has an advantage. The coin is fair, the dice are clean, nobody is cheating. Each player keeps going until one of them is broke.

The probability that you are the one who ends up broke is not 50%. It is your opponent's bankroll divided by the two bankrolls combined. Play against someone with the same money as you and it is a coin flip. Play against someone with ten times your money and you go broke about 91% of the time. Play against someone with effectively unlimited money, and keep playing, and your probability of ending at zero approaches 1.

Not "you will probably lose". You will lose. In a game with no house edge at all, against an opponent doing nothing but showing up.

Why zero is different from every other number

The reason is not complicated once you see it, and it is the single most important idea about gambling that almost nobody is taught.

Your balance moves up and down. Most numbers it passes through are ordinary: you can be at $500 and later be at $900 or $200. Every number is a place you can leave in either direction.

Except one. At zero you stop. Not because of a rule, but because there is nothing left to bet. Probability calls this an absorbing state, meaning a place the process can enter but never exit.

Now count the absorbing states in a casino. You have one. The house has none.

That asymmetry is the entire game, and it survives everything else being fair. You can have identical odds, identical skill, identical luck, and identical information, and still lose with certainty, because you are the only participant who has a way to be permanently removed. Given enough bets, a random walk will visit every level, including the one you cannot come back from. The only question is when.

Karas did not run out of luck in 1995. He ran out of the thing that had been protecting him from the maths the whole time, which was money.

How much bigger they actually are

"Effectively unlimited" sounds like a figure of speech. It is not, and the numbers are public because these are listed companies that have to file them.

Flutter, which owns FanDuel, guided to full-year 2025 revenue of roughly $16.6 to $17.5 billion. DraftKings guided to $6.3 to $6.6 billion, and reported taking $36.8 billion in wagers over just the first nine months of 2025.

Set that against a bettor with $2,000 to their name. The ratio is not ten to one or a thousand to one. It is millions to one.

And the house does not even need an unlimited bankroll for the theorem to bite. It only needs one large enough that its own ruin is not a realistic possibility within your lifetime of betting. Every operator cleared that bar a long time ago.

There is a second layer to it, though, and it is the more interesting one.

The casino is not playing your game

Here is a question that sounds like a trick and is not.

Can a bet have a positive expected value and still ruin every single person who takes it?

Yes. And the demonstration is short enough to check on your phone.

Flip a fair coin. Heads, your wealth goes up 50%. Tails, it goes down 40%. You bet everything, every round.

The expected value is clearly positive. Half the time you multiply by 1.5, half the time by 0.6, so the average is 1.05. Each round is worth 5% more than the last. Any textbook says take this bet all day.

Now actually play it. Over a long run you get roughly as many heads as tails, so pair them up. A win and a loss in either order gives you 1.5 times 0.6, which is 0.9. You lost 10% of everything across the pair. Per round that is about a 5% decline, compounding, forever. Every individual who plays this bet long enough goes broke, with certainty, on a bet whose expected value is positive.

Both numbers are correct. They are answers to different questions.

The +5% is what you get by averaging across many people playing simultaneously. Physicists call that an ensemble average. The -5% is what you get by following one person through time, which is a time average. For anything multiplicative, meaning anything where the next bet's size depends on what the last one did to your balance, those two numbers are not the same. The economist Ole Peters built a body of work on exactly this gap, and it has a name: these processes are non-ergodic.

Sit with what that means for a betting app.

The operator experiences the ensemble average. They have millions of customers betting at the same moment, and their result is the average across all of them, which is stable, predictable, and positive. That is the number in their quarterly filing.

You experience the time average. You have one balance, and it moves through one sequence, one bet after another, and your result is what happens along that single path. That number is worse, and it is worse even when nothing is rigged.

Same game. Two different numbers. Only one of them is yours.

This is also why the phrase "in the long run" is a trap when you hear it about your own betting. The long run belongs to whoever is still solvent at the end of it. For the house, the long run is a business plan. For you, the long run is the thing you have to survive to reach, and the whole difficulty is that you cannot.

The system that should fix this, and why it doesn't

Everyone who understands the ruin problem eventually reinvents the same solution, so it is worth killing properly.

Bet $10 on red. Lose, bet $20. Lose, bet $40. Keep doubling. The first time you win, you recover everything and finish $10 ahead. This is the martingale, and it feels airtight, because it is airtight, given one assumption.

The assumption is that you can always place the next bet.

A normal roulette table runs a $10 minimum and a $500 maximum. Your sequence goes 10, 20, 40, 80, 160, 320. The next one is 640, and the table will not accept it. So you get six attempts, and by then you have $630 on the table trying to win $10.

On a single zero wheel your chance of losing six in a row is about 1 in 55. Which sounds rare, until you notice you only need it to happen once, and that you are running the sequence over and over. Do the arithmetic on the whole thing and you lose about $1.74 per completed sequence, which is exactly the house edge applied to everything you wagered along the way. The system did not beat the edge. It reorganised it into lots of small wins and one catastrophe, which is a much more comfortable way to lose money and not a cheaper one.

The table limit is the part worth staring at. It is not a cap on the casino's exposure in any meaningful sense. It is the mechanism that guarantees your bankroll runs out before your strategy does. The house wrote a rule that makes its own bankroll effectively infinite and yours definitively finite, and it did it in a way that reads like a house-keeping detail on a placard.

Why it always ends on one bet

There is a version of this that skips all the slow parts, and it is the one people actually describe when they talk about the night it ended.

Double or nothing is a fair bet. Take it once and you have a 50/50 shot. There is nothing wrong with the odds.

But taking it repeatedly is not a fair strategy, because the two outcomes do different things. A win gives you a bigger balance and the option to do it again. A loss ends the sequence permanently. Only one of those results is final, so the sequence can only terminate one way.

Ten in a row leaves you standing about 1 time in 1,000. And that is at even money. Even if you somehow had a 90% chance every time, which nobody has ever had at anything, ten straight leaves you standing 35% of the time and twenty leaves you at 12%.

This is why the story is never "I lost it gradually". It is always "I was way up, and then I put it all on one thing". The wins were never going to end it. They were the mechanism that kept it running long enough for the loss to arrive.

Karas told an interviewer afterwards that he had no fear of losing because he had done it before. That is not a character flaw. It is what the structure produces. Every previous recovery is evidence that recovery is possible, right up until the recovery that does not come, which is also the only one that counts. Whatever the house edge does to you slowly, this does in one move.

What to actually take from this

• Stop thinking about odds and start thinking about survival. Whether a single bet is fair is close to irrelevant. What decides your outcome is whether any sequence of losses can remove you. If one can, enough time will find it.
• Never bet a fraction of a shrinking balance. That is the multiplicative trap. It is the difference between the +5% and the -5%, and it is the only part of this you actually control.
• Treat every all-in as the last bet of your life, because statistically it is. Not this one necessarily. But the sequence has exactly one ending and it is not a win.
• Understand what "I'll stop when I'm up" requires. It requires stopping. The structure gives you no signal to stop, and every previous win is evidence you should not.
• Count from where you started. Karas at $40 million and Karas at $50 were the same distance from zero in the only sense that mattered, which is that one bad run got him there either way.

If the difficult part is not understanding this but acting on it at 1am with a balance on the screen, that is a different problem and it needs different tools: a plan you wrote when you were calm, a person you can call, and something between the urge and the app. Ungamble is built for that gap, with a panic button and Rux, an AI friend trained on the science of quitting who will happily work through any of this with you at the hour it actually matters.

The one thing worth remembering

Archie Karas was arrested in 2013 for marking cards at a blackjack table in San Diego, was convicted, and was banned from every casino in Nevada in 2015. He died in 2024. The peak figure of $40 million rests mostly on his own account and on the people who played him, so treat the exact number loosely.

The shape of it is not loose at all, and the shape is the point. The best run in the history of the activity, tens of millions of dollars, against the best players alive, over two and a half years, finished at zero. Not because he got unlucky at the end. Because he kept playing, and the only player at that table who could be removed was him.

You cannot fix that with a system, a bankroll, or a hot streak. There is one move that takes you out of the theorem entirely, and it is the one nobody sells you, because it is the only outcome the house has no product for.

You stop.

Sources: history of the problem: Pascal to Fermat, 1656; Huygens, "De ratiociniis in ludo aleae", 1657 (the fifth of the problems appended to it); general solution by de Moivre, 1712. Operator financials: Flutter Entertainment and DraftKings Inc. full-year 2025 guidance and quarterly SEC filings. Time averages versus ensemble averages in multiplicative processes: Ole Peters, "The ergodicity problem in economics", Nature Physics 15, 2019, and Peters and Gell-Mann, "Evaluating gambles using dynamics", Chaos 26, 2016. Archie Karas biography, the 1995 losses, the 2013 conviction and the 2015 Nevada exclusion: contemporaneous Las Vegas Review-Journal reporting and Nevada Gaming Commission records. Martingale and table limit arithmetic computed for a single zero wheel at a $10 minimum and $500 maximum.