Custodial vs non-custodial wallets: who actually holds your keys

You open an exchange app and see €3,400 in bitcoin. You open a wallet on your phone and see €3,400 in bitcoin. The two screens look so similar that most people assume they are the same thing with different branding. They are not, and the difference only surfaces on the day something goes wrong.

In one of them, that balance is a row in the database of a company that has promised to give it back to you. In the other, it is an unspent output on a public chain that moves only with a key you hold. Every other difference — fees, interface, support — follows from that one.

This guide is not about which is "better". It is about understanding precisely which risk you are taking on with each, using the numbers from what has gone wrong on both sides.

The technical difference, without metaphors

A transaction on a public chain requires a digital signature produced with a private key. Whoever holds the key can move the funds. Whoever does not hold it cannot, however strong their contractual rights. Everything else is a management layer sitting on top of that fact.

Custodial wallet. The provider generates and holds the keys. When you buy, nothing necessarily happens on-chain: the company records in its internal ledger that it now owes you 0.04 BTC. It might hold those bitcoin in cold storage, or pooled across a handful of wallets alongside those of 500,000 other customers, or — if it is behaving fraudulently — not hold them at all. Your relationship is contractual, not cryptographic. The upside is real: lose your password and you recover the account; there is support, there is a complaints process, and where the firm is authorised there is a regulator to escalate to.

Non-custodial wallet (self-custody). The key is generated on your device, normally from a BIP-39 seed phrase shown to you exactly once. The company that wrote the app cannot move your funds, cannot freeze them, and cannot return them if you lose the seed, because it does not have it. This is what makes a hardware wallet meaningful in the first place.

There is a three-second test: did the app show you 12 or 24 words during setup? If not, and you can regain access with an email and a code, it is custodial. Whatever the marketing says.

"Not your keys, not your coins": where it comes from and when it misleads

As a technical description the phrase is correct: if you do not control the key, you do not control the coin, you control somebody's promise. It grew out of a track record of failures that backs it up, and it remains the single most useful lesson in this field.

It becomes dangerous when it hardens into a universal command. Three reasons:

1. It moves the risk, it does not remove it. You swap counterparty risk for your own operational risk. If your backup procedure is bad, you have made things worse.
2. It ignores the asymmetry in capability. An authorised exchange has a security team, audits, and cold storage with multi-person controls. You have a drawer and good intentions. In plenty of cases, the honest comparison does not favour the drawer.
3. It pushes people into a leap they have never rehearsed. Moving €20,000 into a brand-new wallet without ever having tested a restore from the seed is taking on more risk than you think you are avoiding.

What the custodial failures cost

These are the numbers that created the distrust in the first place. They are worth quoting accurately, because inflated versions circulate widely.

Two clarifications almost nobody makes. First: a great deal of FTX money came back. The "$8 billion evaporated" narrative is incomplete; the real damage was that customers spent two years without access and repayments were computed at November 2022 prices rather than at the value on the date they were paid.

Second: the $4.7 billion figure often attributed to Celsius as a "shortfall" is wrong. That $4.7 billion was the scale of obligations to customers; the documented balance-sheet hole was around $1.2 billion. It is a difference of nearly four times, repeated endlessly without checking.

The other side: self-custody fails too

If you only look at the exchange column, you draw a biased conclusion. In its 2026 crypto crime report covering 2025 data, Chainalysis counted $713 million stolen from personal wallets, 20% of everything stolen that year, spread across 158,000 incidents with more than 80,000 unique victims. These are small, numerous losses rather than few and colossal ones, which is exactly why they never make headlines.

The mechanisms differ from a bankruptcy: seed phrases photographed or saved to cloud storage, malicious approval signatures, targeted phishing, compromised browser extensions and — increasingly — physical coercion. The Ledger Connect Kit supply-chain attack of 14 December 2023, which drained roughly $484,000 by injecting code into a library used by thousands of decentralised applications, is the canonical example: every victim held their own keys, and every victim signed with them.

Multisig and MPC: the middle ground almost nobody uses

The choice is not binary. Two architectures distribute control:

Multisignature (multisig). The address requires m of n signatures to spend. A typical 2-of-3 scheme puts one key on your hardware wallet, a second on a different device in another location, and a third with a professional recovery service. Lose one: you still have access. Have one stolen: it is not enough to move anything. It is transparent on-chain and well supported on Bitcoin. The cost is complexity, more moving parts to document, and slightly higher fees because the transactions are larger.

MPC (multi-party computation). The key never exists as a whole: it is generated and used in shares held by several parties that cooperate to produce a signature. All of this is invisible off-chain; the network only ever sees an ordinary signature. It is what many institutional custodians and some consumer wallets run internally. The advantage over multisig is that it works on any chain and does not reveal the scheme. The disadvantage is that you depend on the provider's specific implementation, which you can rarely audit yourself.

What MiCA actually protects

In the European Union, Regulation (EU) 2023/1114 — MiCA — has applied to crypto-asset service providers since 30 December 2024, and the national transition periods ran through 2026 (Spain's ended on 1 July 2026). Providing custody in the EU now requires a CASP authorisation from a national competent authority, or a passport from one granted in another member state. Older national registers created purely for anti-money-laundering purposes have been closed or superseded and no longer entitle anyone to operate.

What the framework adds:

• Segregation of client assets. The provider must keep clients' crypto-assets separate from its own and may not use them on its own account. This is the obligation that directly targets the FTX failure.
• Governance, capital and custody requirements, plus provider liability for losses attributable to its own operations.
• Supervision and a public register you can search. National authorities publish these — the Spanish CNMV register is a representative example, and every member state has an equivalent maintained by its financial supervisor. Check the register of the country where the entity is authorised, not the one where you happen to live, since a passported firm appears in its home state's list. This is the first filter when deciding where to trade.

Outside the EU the picture is patchier. In the United States there is still no single federal custody regime for crypto: a joint SEC and CFTC interpretation of 17 March 2026 clarified the taxonomy of tokens and acknowledged that most crypto-assets are not themselves securities, while market-structure legislation remained pending in the Senate as of August 2026. US custodians are therefore supervised through a patchwork of state money-transmitter licences, state trust charters and, where securities are involved, SEC rules. In the UK, firms must be registered with the FCA under the money-laundering regulations. Wherever you are, the same question applies: which specific legal entity holds my assets, under which authorisation, and can I find it in that regulator's own list?

Side by side

Worked example: three profiles

Ana, €480 and monthly purchases of €50. Buying on an exchange, withdrawing to a wallet and going back in costs her network fees and time every month. Her risk of losing a seed phrase is higher than the risk that an authorised platform fails in the next few years. A sensible call: custody on a licensed exchange, with app-based or hardware-key 2FA, a unique password and a withdrawal allowlist. Once she passes roughly €1,500 and the habit is established, make the jump.

Bruno, €14,000 he does not intend to touch for three years. The arithmetic reverses completely. A $79 hardware wallet plus a €60 metal backup comes to about €150, or 1.07% of his holdings, paid once and amortised over five years (roughly €30 a year, 0.2% annually). In exchange he removes counterparty risk from 100% of his balance. Self-custody, with a mandatory restore rehearsal before moving the bulk: send €20, wipe the device, restore, confirm the funds reappear.

Carla, €260,000 and a business partner. A single device is a single point of failure, and a single point of coercion. 2-of-3 multisig, with one key hers, one her partner's and a third held by a professional service or in a safe deposit box, plus a written and tested recovery document. Setting that up — devices, advice, hours — will run to about a thousand euros: under 0.4% of the holdings, and it is the only one of these three configurations that survives Carla disappearing tomorrow.

Decision checklist

• Establish whether your wallet is custodial using the seed-phrase test, before deciding anything else.
• If you use custody, confirm the provider appears in its home regulator's public register, and note which legal entity is actually named.
• Write down a threshold: how much you are willing to hold with a third party. Revisit it when you cross it.
• Enable 2FA with an app or a hardware key, never SMS, and add a withdrawal address allowlist.
• Before moving a serious amount into self-custody, run a full wipe-and-restore rehearsal.
• Write an inheritance plan: what exists, where it is, and how to access it, without the seed inside the same document.
• Once the numbers would cost you sleep, evaluate multisig before buying a second device of the same type.

• Do not leave your whole balance on an exchange just because the app is convenient to check.
• Do not treat self-custody as a moral switch: it is an operational decision with concrete costs.
• Do not confuse "regulated" with "insured". There is no guarantee fund for crypto-assets.

Where things stand in 2026

The European map has simultaneously narrowed and hardened. With MiCA fully applicable, the list of platforms that may legally serve EU customers is short and searchable — something that simply did not exist in 2022. When you decide where to leave assets in custody, the licence is now the first filter, ahead of fees. The trade-off is that some large platforms have ended up outside the perimeter, and their European users have had to move funds on tight deadlines.

On the self-custody side, the trend Chainalysis recorded through 2025 and 2026 is uncomfortable: fewer mass technical hacks against individuals and far more social engineering, impersonation and coercion, with impersonation scams growing 1,400% year on year. That shifts the problem from "where do I keep the key" to "how do I verify what I am signing", and neither architecture protects you there on its own.

The sensible split for almost any holding remains what it always was: what you need to move, somewhere regulated; what you do not intend to touch, under a key only you control; and anything beyond that, split across several keys. What has changed is that checking the first part now takes thirty seconds in a public register, and that the second part is being attacked through your judgement rather than through your hardware.

Frequently asked questions

Sources and references

1. Chainalysis — 2026 Crypto Crime Report
2. Chainalysis — Crypto hacking and stolen funds 2026
3. CNMV — Crypto-asset regulation under MiCA
4. CNMV — Register of crypto-asset service providers (example national register)
5. SEC — Staff clarifies application of federal securities laws to crypto assets
6. FBI — Cryptocurrency and AI scams bilk Americans of billions (IC3 2025)
7. Ledger — Security incident report
8. CoinDesk — Ledger exploit drained $484K

Related guides