Name the Peril. Price the Architecture.
Why wide cover with vague edges buys disputes, not certainty — and three rules for custody that turn clarity into capacity.
BRIEF
Insurance is contingent capital: a promise to deliver a defined sum on a defined event, at a price that reflects what is knowable about the risk. Standard forms work where decades of similar losses let underwriters sort risks into buckets. Digital infrastructure has no such history, and wide cover with vague edges buys disputes, not certainty. I propose three working rules for digital asset custody: name the perils, price the architecture, and let limits follow evidence. Together they shrink what a capital provider has to assume. That is what unlocks capacity, and capacity is what moves penetration.
In many of the large digital-asset losses I’ve traced, the payer was the holder’s own balance sheet or a lender’s bridge, not an insurer. This piece is about why that happens, and three rules that would influence who pays.
Who pays for a loss today
Every loss has a payer. The only question is whether the payer was chosen in advance.
For dollars, the answer is layered and visible: the US Government insures a bank deposit up to $250,000, and SIPC protects brokerage customers up to $500,000 if the firm fails, but says it does not protect digital assets that aren’t securities.1,2
For digital assets, the public record shows several kinds of payer, and none of them was chosen in advance.
WHO PAID: SELECTED PUBLIC LOSSES, 2016 TO 2025
| Payer | Case | Loss | How the gap closed |
|---|---|---|---|
| Own balance sheet | Coincheck, Jan 2018 | ~$530M (~¥58B) | Refunded about $400M (~¥46.3B) to 260,000 customers from its own funds.3 |
| New investors | Ronin, Mar to Apr 2022 | ~$540M–$625M (173,600 ETH + 25.5M USDC, price-dependent) | Sky Mavis raised $150M, led by Binance, and committed to reimburse all users.4 |
| Debt and parent equity | DMM Bitcoin, May to Jun 2024 | ~$306M (4,502.9 BTC) | About $352M (¥55B) through borrowing, a capital increase and subordinated debt, with parent-group support. Closure announced six months later.5,6 |
| Customers | Bitfinex, Aug 2016 | ~$72M (119,756 BTC) | Balances cut 36% across all accounts, including accounts not breached. Customers received BFX tokens in proportion to their loss.7 |
Dollar figures above are approximate, converted at or near the date of each event; DMM Bitcoin uses the rate reported alongside the yen figures when its funding plan was announced (5 June 2024). Ronin is given as a range, not a point estimate, because the token price moved materially between the theft and Sky Mavis's own disclosure.
Insurers rarely announce a payout, so silence in the press is weak evidence either way. That is why I say “many,” not “most.” None of these four cases even reached a wording dispute — there was no policy on risk to argue over. That’s the more basic failure, and the one this piece is mostly about. The wording fights come later, once someone is willing to write the risk at all. What the table does show is who steps in when the cover is absent: whoever has the deepest balance sheet, and sometimes the customers.
Then there’s the cash under the mattress. Self-custody is the ethos of this asset class, and I don’t have a sourced split of how much value sits there. Whether it should be insured is a separate question from whether people should self-custody: in my view, for most of it there is no viable insurance product. A holder who is the only control is a single point of failure by definition — a fact about design, not a verdict on the culture.
Insurance is contingent capital
Insurance does one thing: it converts a capital problem into a price. A firm that might lose $100 million on an event can hold $100 million idle, or pay a premium and let someone else hold it. I use “contingent capital” loosely, as a promise to deliver a defined sum on a defined event, not as the bank-capital instrument.
That makes insurance a tool in the buyer’s financial stack, not a compliance cost. It also makes the quality of the promise the whole game: which event, how much, how fast, and how certain.
Certainty is where frontier technology struggles. The promise has to be defined well enough to be priced and honoured. Everything below follows from that.
Why standard forms work, and where they stop
ISO aggregates premium and loss data across the industry by class and drafts the standard forms most carriers write from.8 A small manufacturer, shop or office reports into the same class code as thousands of similar risks, so the exclusions were tuned by those losses, and both sides know where the edges are. Underwriting here is sorting: which class, which bucket, which rate. Nobody is architecting the risk. The risk is being filed.
Digital infrastructure has no class code to sort into. ISO hasn’t published loss costs for a Tier IV data center or an institutional digital-asset custodian — the class is too nascent for the pooled loss history a rate depends on — so there’s no bucket with real experience to file the risk into.
None of this makes the forms bad. Risk is priced by regime, and the regime follows the data — the three rows below are a practitioner’s framework, not an industry taxonomy.
| Regime | How it is priced | Where it fits |
|---|---|---|
| Classify | A rate from class and exposure. Thousands of similar risks, so the bucket does the work. | Small commercial. The base property of a data center built to a known standard. |
| Model | Catastrophe models for natural perils. Stochastic models for bespoke perils where a few years of data and known physics allow it. | Natural catastrophe on any property. Many bespoke data-center perils. |
| Architect | No population and short history. Price from the control design, and write the wording to match. | Digital asset custody. In data centers, the seam between property and cyber. |
The three rules below are for the third row.
Name the perils
Start with a question a data-center operator should be able to answer. If a firmware exploit bricks a GPU cluster, is that a property loss, a cyber loss or a tech E&O loss? The hardware is physical. The event is cyber. Silent cyber cover that exists by the absence of a clear exclusion rather than by explicit inclusion sits in that gap. Depending on how the program is built, the same loss can be covered, excluded or disputed across property, cyber and tech E&O at once, with no carrier accepting primary responsibility.
Even all-risk paper doesn’t settle it. NotPetya hit Merck in 2017, and its all-risk property policies excluded only hostile or warlike action. Eight insurers held out, disputing about $700 million of the claim. In 2023, a New Jersey appeals court ruled the exclusion didn’t apply — "hostile" meant military action, not any government-linked act.9
Nobody drafting those words had NotPetya in mind. It took six years for a court to say what they meant.
Naming the perils moves the argument from claim time to quote time, when both sides can still walk away. It costs something: a named list leaves things out. For custody, where loss history is a few years deep, the named peril is the honest promise: this event, this limit, this price.
Price the architecture
Naming the peril fixes the promise. The price comes from what the underwriter can see. For custody, the question that replaces temperature, from my last piece, is this: what is the smallest set of failures that moves the funds? If it’s one, the system has a single point of failure. If it takes several independent failures, it’s resilient. The CryptoCurrency Security Standard asks versions of this across key generation, storage and usage.10 A rating approach built on that question moves price with quorum, signer independence and verification of what gets approved, not with the label on the wallet.
The question has to be asked of the book as well as the insured. When many insureds rely on the same signing vendor, wallet interface, oracle or bridge, one failure moves them all. Each can look resilient on its own while the portfolio holds a single point of failure. That is common-mode failure. It has to be tracked as a dependency across the book and capped like any other limit.
Clarity compounds into speed, too: a wording and a price that are already defensible don’t wait on a committee to bind — and in a sector that moves this fast, a quote that turns around in days is itself a form of capacity.
Let limits follow evidence
Limits are where evidence becomes a number. I worked through a velocity-limit example — drain rate times time-to-freeze as a probable maximum loss — in my piece on hot and cold storage controls. The short version holds only if three things hold. The same compromise can’t change the limit. One control set can’t reach every wallet. And the freeze works when tested. Each is a question about evidence, not intent.
That gives a ladder. A stated control earns a small limit. A tested control earns more. A control verified by an independent party and monitored in operation earns more again. Audits and penetration tests belong on the ladder as evidence that a control is enforced, not as the control itself.
Why this unlocks capacity
Uncertainty and ambiguity limit capacity supply. Three things sit in the dark today: what the words mean, what fails together, and how bad one event can get. Each rule addresses the fog.
| Rule | Unknown it removes | What the capital provider gains |
|---|---|---|
| Name the perils | What the wording means at claim time. | Known scope, with disputes at quote time instead of claim time. |
| Price the architecture | What fails together. | Visible frequency drivers and a book-level aggregate it can cap. |
| Limits follow evidence | How bad one event can get. | A bounded loss it can size a line against. |
A provider that can bound severity, see correlations across the portfolio and read the promise can put more capital behind each risk, and behind more risks. That is how capacity grows. It is a hypothesis, not a result. The test is whether capacity followed clarity in the classes that got it first.
The prize is large. I can’t give you a sourced figure for how much digital-asset value is insured, and I doubt anyone can. My estimate from placement work is that the insured limit sits at or below 1% of the value at risk. Treat that as a practitioner’s estimate with a loose denominator, not a statistic. AM Best put the crypto market above $3 trillion in June 2025, citing Coinbase.11 The grid below uses total market value as the denominator, not an insurable base — it says nothing about what share of that value could or should be insured. The grid is arithmetic, not a forecast.
LIMIT NEEDED ($ BILLIONS) AT A GIVEN MARKET VALUE AND PENETRATION
| Market value | 0.5% | 1% | 2% | 5% |
|---|---|---|---|---|
| $2 trillion | 10 | 20 | 40 | 100 |
| $3 trillion | 15 | 30 | 60 | 150 |
| $4 trillion | 20 | 40 | 80 | 200 |
| $6 trillion | 30 | 60 | 120 | 300 |
At $3 trillion and 1%, the market means $30 billion in aggregated limit. Not all of the value is insurable. But every point of penetration needs capital providers who trust the promise. The rules are how they come to be.
The steelman
Named perils leave gaps. A buyer who reads a short list may buy cover that doesn’t respond to the loss they suffer. The answer is plain schedules and advisers who explain them, not vague wording.
The strongest objection to all this is that buyers don't choose all-risk because it's architecturally better — they choose it because they can't enumerate every way they might lose, and a named list always misses something. That's true, and it's not new. Lloyd's wrote long-tail asbestos liability cover for decades on Names who carried unlimited personal liability. Claims surfaced years after the policies were written, under doctrines that let one claim reach every policy a company ever bought. Membership had peaked near 32,000 in 1988; by 1993 Lloyd's admitted limited-liability corporate capital for the first time in its history and reinsured the back book through a new vehicle, Equitas.12 What sold didn't fail because buyers were wrong to want broad cover. It failed because nobody had priced what the architecture of that promise actually was.
Even the industry doesn't fully trust the name it's defending. "All risk" spent decades drifting toward blander language — "open perils," then "causes of loss – special form" — because courts kept reading "all" as a promise that everything was covered, a reading insurers found indefensible once a claim actually tested it. If the people who draft the word "all" don't trust it, a buyer shouldn't either.
Capacity is also not the only reason penetration is low. Price and demand matter, and not every buyer should insure. A firm with a deep balance sheet and a trusted way to raise capital may rationally self-fund. The market’s job is to make cover cheaper than idle capital for the buyers who need it.
An invitation
The market has been selling reassurance. It should be selling a promise.
That promise only holds as a shared standard. If you underwrite, broker or provide capacity for custody, help write it — tell me where this breaks on LinkedIn or X.
Sources
- FDIC: Understanding deposit insurance
- SIPC: What SIPC protects
- The Daily Star (AFP): Japan sanctions Coincheck after massive cryptocurrency heist
- Stockhead: Axie Infinity creator Sky Mavis raises US$150 million to compensate victims of massive Ronin hack
- Kanto Local Finance Bureau: Business improvement order confirming DMM Bitcoin’s BTC outflow
- Decrypt: Japanese crypto exchange to shut down after $300M hack, transfer accounts to SBI
- Bitfinex: Security Breach & BFX Token FAQ
- IRMI: Insurance Services Office, Inc.
- Claims Journal: N.J. appeals court rules war exclusion doesn’t apply to NotPetya attack
- CryptoCurrency Certification Consortium: CCSS v9.0 release
- AM Best: Sketching a Risk Profile for the Emerging Crypto Market
- Economic History Society (Robin Pearson): Delusions of Competence — The Near-Death of Lloyd’s of London, 1980–2002
I work at a Lloyd’s coverholder. Views are my own, not those of my employer or any carrier. Nothing here is an offer of coverage or a statement of how any policy will respond.