Mt. Gox: When an Exchange Became a Single Point of Failure

An exchange customer did not control the private keys to individual bitcoins. They saw an internal account and trusted the company to hold corresponding reserves, process withdrawals correctly and protect its wallets. As long as withdrawals worked, the difference between a coin and a claim to a coin was almost invisible.

In February 2014, Mt. Gox halted withdrawals and then sought protection from creditors. The company announced that roughly 850,000 BTC were missing — about 750,000 belonging to customers and 100,000 of its own. Bitcoin’s consensus had not failed. The records, security and management of a centralized intermediary had failed.

The exchange had shown warning signs before its final collapse: withdrawal problems, delays and explanations customers could not independently verify. An internal number on a screen changed instantly, but an actual transfer to a personal wallet was the only test of whether the company’s obligation corresponded to coins it could move.

For many users, Mt. Gox was Bitcoin. Its login page displayed the price, its order book set the rate quoted by the media and its Buy button turned dollars into a digital balance within seconds. Mark Karpelès appeared to be a technical founder capable of keeping a complex machine running. A customer could not see the wallets, internal permissions or whether the sum of all accounts matched the actual reserves. The convenient interface concealed the fact that they had sent property to one company in one jurisdiction.

As withdrawals slowed, people compared experiences on forums and some traveled to the Tokyo office. British creditor Kolin Burges stood outside the building with a sign demanding to know where customers’ money had gone. Inside, the problem was still described in technical language; outside were people who could no longer pay bills or recover their savings. The crisis of trust did not begin on the day of insolvency. It began every time the exchange promised a withdrawal and offered another explanation instead of a transaction.

In March 2014, Mt. Gox announced that it had found approximately 200,000 BTC in an old wallet format. This reduced the net shortfall, but deepened the question of how a company handling such large sums could lose track of its own wallet.

The earliest public explanations emphasized transaction malleability, a property of transactions at the time that could change their identifiers. Later investigations and charges linked a large share of the stolen coins to a long-running drain and laundering through BTC-e. One technical label did not explain years of poor controls.

The discovery of an old wallet was itself an accounting alarm. A company managing other people’s property did not know that it controlled a reserve of 200,000 BTC until it searched after the collapse. Without continuous inventory, separation of roles and comparison of internal liabilities with on-chain funds, even the recovered coins became evidence of the problem they were supposed to lessen.

In the March 20 announcement, Karpelès explained that the coins were in an old-format wallet the company believed to be empty. This was good news for creditors and a frightening description of internal control. The two hundred thousand bitcoins were not found through a sophisticated investigation of an attacker, but while searching the company’s own old storage. Only after falling did the company begin to determine what it actually owned.

Later analyses traced large outflows through addresses associated with the BTC-e exchange, and U.S. authorities charged its operator, Alexander Vinnik, with laundering stolen funds. An indictment is not the same as a complete technical reconstruction of every missing coin, but it shifted the explanation from a one-time bug toward prolonged extraction. Transaction malleability may have complicated withdrawal accounting; it could not excuse years in which management failed to see its liabilities diverging from the coins it controlled.

The recovered funds did not return to customers at the press of a button. Japanese insolvency and civil-rehabilitation proceedings followed, with claims verification, choices about payment methods and legal disputes. The first substantial payments in bitcoin and bitcoin cash did not begin until 2024.

The trustee repeatedly moved the final deadline for some payments; the current official notice sets it at October 31, 2026. That is the second lesson of centralized custody. A Bitcoin transaction can become final within hours, but a claim against a bankrupt company may accompany a person through a substantial part of life.

Nor were creditors waiting for a simple return of identical balances. The proceedings had to determine eligible claims, valuation methods, currencies of payment and treatment of the limited pool of recovered assets. Bitcoin’s price rose substantially over the years, creating an unusual situation: the property became far more valuable, while each creditor still bore the time, uncertainty and rules of insolvency.

Japanese lawyer Nobuaki Kobayashi entered the center of the story as rehabilitation trustee. His work looked almost like the opposite of a fast internet exchange: collect claims, verify identities, negotiate a plan, administer recovered coins and inform creditors through legally binding steps. Every postponed deadline felt like another lost year to the person behind the screen, but the trustee could not simply distribute keys according to an old database export. He had to respect the order and claims of thousands of people in many countries.

Some creditors sold their claims to funds at a discount during the wait because they needed certainty or money immediately. Others kept their claims, waited ten years and watched the recovered coins rise in value. The same collapse led to different outcomes: selling ended the uncertainty at the cost of accepting a loss; waiting left open both the possibility of a larger recovery and the risk of further delays. Behind the anonymous balances were people for whom money available today and money promised after the proceedings were not the same thing.

After Mt. Gox, “not your keys, not your coins” became the lived experience of a generation. Self-custody reduces the risk that a company will use, lose or freeze the funds. It adds the risk of personal mistakes illustrated by Howells and Thomas.

The lesson is therefore neither that every intermediary always steals nor that every user can safely manage keys. It is more precise: a balance at a custodial exchange is a counterparty liability. Anyone who mistakes it for bitcoin controlled by a personal key overlooks the most important line on the balance sheet — whom they must ask for permission to withdraw.

Even proof of reserves is incomplete if it does not also show liabilities and the rules governing use of the reserves. A snapshot of addresses may confirm certain coins at a certain time; by itself it does not reveal whom the company owes, whether the assets are pledged or whether its processes are secure. Mt. Gox teaches us to examine not only the existence of reserves, but the entire relationship among keys, accounting and customer claims.

Karpelès’s story is often presented as a portrait of a single culprit, but failure on this scale also required a crowd willing to entrust an ever larger volume to the exchange. That was not the customers’ fault; they reasonably expected the service to hold what it displayed. Yet a network effect formed a trap: the more liquidity Mt. Gox had, the more useful it became and the harder it was to go elsewhere. Centralization did not arise by command. It grew from convenience, a brand and thousands of decisions to trade where everyone else already traded.

After the collapse, two different ledgers remained. The Bitcoin blockchain continued agreeing with every valid transaction signed by the key holder. The Mt. Gox database contained promises to customers that the company could no longer fulfill. The difference was not a philosophical slogan but a decade of waiting. When users move coins to their own wallets, risk does not vanish; it changes from the decisions of Mark Karpelès and an insolvency court to their own ability to protect and recover keys.