The Custody Compare Desk: Who Holds Your Bitcoin, Compared

The one-sentence take

Custody is the first question, not the last. Before any rate, yield, or feature, ask: who can move your bitcoin, and what happens if the custodian fails? Pledge scores 11 custodians on an 8-factor rubric — but a score only means something read within a custody model, never across models. A self-custody 8.0 and an institutional-custody 8.0 are not the same 8.0.

The first question

A wallet does not store bitcoin; it stores the private keys that authorize spending. "Custody" is the answer to two questions: where do the keys live, and who can touch them? Every custodian is a different answer to those two questions, and the difference decides what can go wrong.

The two model groups — compare within, never across

Pledge's publishing rule is explicit: do not rank collaborative custody, self-custody tooling, and institutional qualified custody as if they solve the same job. They are different products for different people. The desk groups by model, then compares within the group.

Group 1 — Collaborative / self-custody (5): Unchained, Casa, Swan, AnchorWatch, Onramp. You keep meaningful control; a partner (or partners) co-signs. The job: self-custody with a safety net.

Group 2 — Institutional qualified custody (6): BitGo, Fidelity Digital Assets, Anchorage Digital, NYDIG, Gemini, Coinbase Custody. A regulated fiduciary holds the keys. The job: institutional-grade safekeeping with a legal wrapper.

The table (11 scored providers)

Scores are Pledge's 8-factor weighted composite (0–10), point-in-time 2026-05-24 data-layer snapshot (re-verified 2026-08-20). Confidence: only BitGo + Anchorage = verified; the other 9 are estimated (entity verified, some sub-factors inferred from public disclosure). Render the confidence flag, not just the number.

Within-group reading — what actually separates them

Collaborative tier (control + recovery are the axes). Unchained (2-of-3, client holds 2 keys) and Onramp (2-of-3 across three firms) structurally prevent any single party from moving the funds — the closest to "you keep control" while someone else helps. AnchorWatch adds an insurance wrapper on top of a miniscript vault (Lloyd's-syndicated, AM Best A+ paper) but has been live under two years — its loss-history factor reflects a short window, not a long record. Casa is the self-custody coordinator: it helps you run your own keys, and recovery/inheritance is the product. Swan is the guided option — the most service-dependent, and the only one with prior-custodian exposure (2023 Prime Trust receivership + Fortress Trust hack). That's a flag, not a bury.

Institutional tier (charter + insurance + segregation are the axes). The legal wrapper is the bankruptcy-remoteness proxy: BitGo and Fidelity are now OCC-chartered national trust banks (conversion approved Dec 12 2025 — OCC's action was conditional, preliminary and contingent, with the OCC reserving the right to modify or rescind; BitGo's own press describes it as full/unconditional), Anchorage was the first OCC-chartered crypto bank, and Gemini/Coinbase/NYDIG are NYDFS-chartered trust companies — fiduciaries under New York Banking Law. Coinbase Custody is the exception that needs the fine print: its bankruptcy-remoteness claim rests on a UCC Article 8 clause ("custodied crypto is not property of Coinbase and not subject to the claims of its general creditors") that Coinbase's own 10-K admits courts have not yet tested — a legal argument, not a settled legal fact. And the majority of US spot BTC ETF assets — 84% ($77B of the ~$91.7B market, Apr 2026, Reported — secondary market-data sources; Coinbase itself disclosed custody of 9 of 11 at launch) — sits with Coinbase, so it's a market-level concentration, not a niche.

The load-bearing facts (read these twice)

1. No custodian is government-insured. Not FDIC, not SIPC, not SIPA (SIPA excludes non-security digital assets). Every primary page says it; the page should too.
2. SOC-2 is not insurance. SOC-2 is an AICPA controls attestation — an audit that controls were designed and operating. It pays nothing if assets are lost. "Insured" with no amount and no named underwriter is a marketing zero.
3. Private insurance is capped and partial. BitGo: $250M (Lloyd's of London + European syndicate; BitGo pays its own deductibles) — verified 2026-08-22 against bitgo.com. Gemini: $125M as of Mar 1 2024 ($25M hot-wallet commercial crime + $100M cold storage) — verified against gemini.com/institutions/custody. Coinbase: commercial crime/specie/cyber policies with no disclosed limit; its 10-K admits cold-wallet losses could exceed coverage and corporate assets may be insufficient.
4. Insurance covers theft/loss events, not market loss. A capped backstop, not a balance guarantee.
5. A charter is stronger than a contract clause. Fiduciary trust/bank charters (OCC, NYDFS) are the structurally strongest legal wrapper; the untested UCC Article 8 claim is the soft underbelly of the exchange-custodian tier.

The honest caveats

• Scores are category-relative and point-in-time. A custody 8.0 is not a loan 8.0 or an ETF 8.7. The data layer is a 2026-05-24 snapshot (re-verified 2026-08-20); re-stamp at publish.
• 9 of 11 scores are estimated, not verified. Only BitGo and Anchorage carry verified confidence. The number is a starting point, not a verdict.
• Don't let brand familiarity substitute for legal clarity. Fidelity and NYDIG score on TradFi trust signals with relatively thin public disclosure; the score reflects what's provable, not what's assumed.
• Don't conflate Coinbase Custody Trust with the Coinbase consumer platform — separate legal entities, different protections. Same for Gemini Custody vs the Gemini exchange's retail lending history (Genesis Earn 2022).
• Unchained is not a qualified custodian. It coordinates a multisig where you hold 2 of 3 keys; the custody job is partly yours. That's the point of the model — and the responsibility.
• Vendor security claims are self-published. "Secure Element," "EAL6+," "audited" — vendor claims, not certifications of your safety.

The take

Compare within a custody model, on three axes: control (who can move it), recovery (what happens if you lose access), and insurance (what actually pays out, with an amount and a named underwriter). The score is the map; the agreement and the charter are the territory. Risk is the product.