Seven card apps, one order book

Outside the exchange-issued cards, 9 of 139 obtainable crypto cards let you open a leveraged position in the same app that issues the card. 7 of them execute on Hyperliquid — 7 front ends onto one order book. Every one of them documents it. None of them counts.

A crypto card used to be a spending accessory. Over the past year a different shape has appeared: the wallet app that issues the card also carries an order book, and the same login that pays for coffee can open a forty-times-leveraged short. We went through all 139 cards a person can actually get today, set aside the ones issued by exchanges — where a derivatives desk is the parent company’s main business and tells you nothing — and found 9 built that way. That exclusion is a judgement about what is worth reporting, not a filter the data applies, and the group it removes has no published size: no column in the dataset records issuer type, so any count of it would be a brand-name guess.

The finding is not that wallets are becoming exchanges. It is narrower than that: 7 of the 9 route to the same venue, so what the cardholder picks is a front-end onto one market.

Seven out of nine is not, by itself, a surprise. Hyperliquid carries the deepest on-chain perpetuals book, and an app adding a desk routes to the venue that can fill the order. Each time it is the obvious call, and better for the trader than a thin book of the wallet’s own. What the ratio adds up to is the finding: one dependency, unlabelled, behind 7 products a reader would name as different companies.

Key findings

1. 7 of 9 desks execute on Hyperliquid — Hyperbeat Card, Based Card, Tria Card, MetaMask Card, Ledger Card, xPortal Card and Coin98 Fusion Card. These are interfaces onto one order book, not 7 venues. A halt or oracle failure there reaches every one of them, and only Tria has somewhere else to go: it ships a second route, on Decibel.
2. The same 7 are running someone else’s venue, and every one of them says so. Ledger, MetaMask, Coin98 and Based name the venue on the same page as the feature; xPortal and Tria a line or a page away; Hyperbeat in its Terms, which define perp trading as executing “from your Pay Wallet, directly on Hyperliquid”. This is not a secret being kept, it is a dependency nobody has added up, which is why the count is the finding and not any single disclosure. The two that run their own book are Jupiter, on its own Solana program, and Revolut, through a brokerage authorised by the Bank of Lithuania.
3. “Trade and spend from one account” is true of 1, not 9. Only Hyperbeat Card defines the card and the collateral against one wallet in its own terms, whatever the other landing pages imply.
4. The card’s country list is not the desk’s country list. Ledger blocks its desk in five jurisdictions — the UK, the US, Ontario, France and Belgium — and sells the card in all five, so a holder in any of them gets a card and no desk. Two of the five are EEA states, which is easy to miss when the card’s row says only “EEA”.
5. 13 candidate groups were rejected, and the reasons are published below rather than left implicit. The two largest rejections are whole classes: collateralised borrowing dressed as leverage, and tokenised shares held at 1:1.

The 9 cards

Shared balance first, then the rest. On the first card the trading margin and the card sit in one wallet. On 5 the issuer states that they do not. On the remaining 3 the app has two places money can sit and nobody has published which one a liquidation reaches — the last column says so rather than guessing.

Where the orders actually go

Counting front-ends per venue. Tria appears twice because it ships two routes, which is why the bars sum to more than 9.

We have written this shape before, in the BIN-sponsor concentration research: several consumer brands, one regulated entity underneath, and a failure mode that only becomes visible on the day it fires. The difference here is that a BIN sponsor failing stops your payments, while a venue failing strands an open position with money in it — and this venue has fired, three times, before most of these card apps had a desk to be affected.

What is at the other end

Seven of these cards route to a venue none of them is, so it is worth saying what it is. Hyperliquid describes itself as a layer-one blockchain, not a broker: its own docs call it “a performant blockchain built with the vision of a fully onchain open financial system”, running a custom consensus called HyperBFT, with the order books living in a component called HyperCore where “every order, cancel, trade, and liquidation happens transparently with one-block finality”. There is no licensed intermediary in that sentence, which is why the compensation schemes and the negative-balance rule that come with Revolut’s brokerage have no equivalent here.

It also means the costs and the failure mechanics are published once and apply to all seven, whatever the card app is called. Perps fees start at 0.045% taker and 0.015% maker and fall with rolling 14-day volume. Funding is charged every hour, paid between traders rather than to the platform, with the interest component fixed at 0.01% per eight hours — 11.6% a year, paid by longs to shorts before the premium component moves it either way. And the card app can add a fee of its own on top, through what the venue calls builder codes: Coin98 publishes 0.04%, which takes the entry cost from 0.045% to 0.085% — nearly double, disclosed in the docs and nowhere near the app screen.

Those percentages are quoted on the position, and the reader’s money is the margin, so they are worth converting once. At the 40x ceiling several of these apps advertise, a position is 40 times the margin behind it: the 0.045% taker fee is 1.8% of your margin each way, about 3.6% for the round trip before the market has moved at all, and the fixed part of funding alone runs 1.2% of margin a day — 0.01% of the position every eight hours, three times over.

What each front end adds on top

Coin98 is the clear case and was, in an earlier version of this page, the only one reported — which left the impression that the other six add nothing. They may or may not. Two publish a number, one says it charges none in the desk’s own terms and the opposite in its general ones, one publishes a total without saying whose it is, and three publish nothing at all.

The liquidation numbers are the ones worth reading twice, because they are what the leverage ceilings on the front pages actually mean. Maintenance margin is half the initial margin at maximum leverage, which the docs put between 1.25% on 40x assets and 16.7% on 3x assets — on a 40x position, a 1.25% move against you starts the process. What happens next depends on the order book: if market orders close the position and bring the account back above maintenance margin, whatever collateral is left stays yours. If they cannot, and equity falls below two-thirds of that threshold, a backstop liquidation moves the position and the cross margin to a liquidator vault, and the docs say plainly what the trader is left with then: “if the trader has no isolated positions, the trader ends up with zero account equity”. That last clause is also the argument for using isolated margin rather than cross for a position you are not watching — isolated confines the loss to the margin behind that one trade, and it is a setting in the app, not a property of the card. And if an account still goes negative, auto-deleveraging closes positions on the winning side at the previous mark price to keep the platform solvent. So “a liquidation costs you the margin” is the good case, and a profitable position being closed for you is a documented outcome rather than a hypothetical.

None of that is hypothetical either, which is the part a concentration argument usually cannot show. On 26 March 2025 the venue’s validators voted to delist the JELLY market and every open position in it was settled — the documented behaviour, since “when an asset is delisted, all positions are settled and open orders are cancelled”. The venue then announced that long holders would be refunded “as if their position settled at the closing price of 0.037555”, which is itself the admission that the actual settlement went against them. On 10 October 2025 the first cross-margin auto-deleveraging in the venue’s history closed positions on the profitable side. That account of it comes from the founder’s personal post rather than the venue’s own channel — “during recent volatility, Hyperliquid had 100% uptime with zero bad debt. This was Hyperliquid’s first cross-margin ADL in more than 2 years of operation” — which is worth knowing when you go looking for it later. And order entry has failed outright: the status page records 37 minutes on 29 July 2025 when “orders were delayed in being sent to the nodes”. A card app cannot route around any of it.

One thing these three do not establish: a loss to anyone reading this. All three predate most of the desks on this page — MetaMask Perps, Hyperbeat and xPortal’s perps are 2026 products — so on none of those days did a single one of these 9 cards have a position at risk. They are evidence of how the venue behaves under stress, not a casualty list. We also have not computed an uptime denominator: three incidents over roughly a year and a half is a rate we have not measured against anyone else’s.

One pot or two

This is the property that decides how far a bad trade can reach, and it is the one thing the marketing blurs. “All from the same account you save and spend from” is Hyperbeat’s own line, and Hyperbeat is the one card here where the terms back it: they define both card spend and perp trading against the same Pay Wallet. Even there, the issuer stops short of the sentence a reader wants — card credit is collateralised into lending protocols while perp margin sits on the venue, and nothing published says a liquidation on one side reaches the other.

Five more document a split themselves, and three of those give it away in the small print: a minimum deposit to open the trading account. MetaMask asks $10 into a Hyperliquid account, Coin98 5 USDC into a dedicated futures account, Tria $10. A minimum deposit is a reliable tell, because money that has to be moved was never in one place. Based says it outright instead: card and trading wallets cannot be converted into one another, and card withdrawals are not enabled. Revolut needs no tell at all, because the separation there is legal before it is a product decision. Its CFD cash is client money, which Revolut says is segregated and safeguarded, with the caveat it prints in the same breath: up to 20% of that value can sit on accounts inside Revolut Group, and only the remainder goes to a third party.

Based is the card that most looks like an exception and is not. Its web terminal offers a “Unified Account” where “a single balance per asset covers everything: perps, spot, and cross margin” — which is a setting on the trading side, not a bridge to the card. The FAQ answers the actual question directly: card and trading wallets are separate, cannot be converted into one another, and card withdrawals are not enabled at all. Read the FAQ, not the product page.

Note how often “self-custody card” describes the wallet rather than the money already loaded to spend. Coin98 pairs a self-custody wallet with a card balance held by its issuer, which it says plainly. Ledger and xPortal are the other kind of answer: Ledger’s card is provided by Baanx and Ledger does not say where the card balance sits, while xPortal documents topping up a perps account without ever stating that the card balance is a different pot. Reading those two as separated is our inference, not the issuer’s statement. Tria’s card is not a debit balance at all, but a 0% APR credit line against collateral.

The two country lists

Two of these cards publish a coverage map for the card and a different one for the desk, and a third card was cut from the research when the two maps turned out not to overlap anywhere. Quoting one number for both is how a comparison page tells a reader they can trade when they cannot.

The table below carries 4 rows because two different problems live in the same field. Ledger and Revolut are the genuine article: card and desk sold to different countries. Hyperbeat and Tria are a second kind of trouble, where the card’s own documents disagree with each other about whether a US resident may hold it at all.

Where each desk is closed, in the issuer’s words

Telling a reader the region beside a card is the wrong list and then not supplying the right one is half a finding. Four issuers publish a desk restriction, one publishes it only by reference to sanctions regimes, and four publish nothing we could find — which is itself worth knowing before you fund an account. Restrictions belonging to a different product of the same issuer are excluded here on purpose: Hyperbeat’s 24-country list is for its raffle product, Based’s 77-country list is for the Visa card, and Jupiter’s one region sentence is about prediction markets.

Two run their own book, one is regulated

Two entries run their own book rather than someone else’s — Jupiter on its own Solana program, and Revolut — but only Revolut trades under a licence written for the job. Its CFDs run through Revolut Securities Europe UAB, authorised by the Bank of Lithuania, at rates it says it generally offers: 1:20 on major indices and on gold, 1:10 on other commodities, 1:5 on single stocks. Negative-balance protection comes with that regime, and the same page states it: the margin-close-out rule “makes sure your account doesn’t fall into a negative balance”. It applies because the licence requires it for retail clients, which makes it firmer than a promise a firm writes for itself. The other 8 desks carry no equivalent and no compensation scheme. The seven of them that publish a ceiling at all run from 40x to 250x — a different instrument from a 1:5 CFD on a single stock, and not a multiple of it.

The low rates are a feature of the regime, not a verdict on the product — but they are the reason these 9 do not belong in one undifferentiated table.

What we left out, and why

13 candidates or candidate classes were cut, and the reasons are published because most of them are cards a reader would expect to find here.

The honest counter-arguments

A few things should be said against all of this.

The denominator is ours. 9 of 139 is a rate over a population we defined, and the group we took out, the exchange-issued cards, is the group that would dilute the finding, because every exchange runs its own book. We still think the cut is right. An exchange having derivatives is not news. But it is a judgement, its size is unpublished because no column in the dataset records issuer type, and a reader is entitled to treat the absolute number as the sturdier one. The two closest calls under our own rule are Bitget and BloFin, both cut on the centralised-parent test. Count them and it becomes 7 of 11, and the concentration ratio falls from 78% to 64%.

Routing to the deepest book is also good for the trader. A wallet that stood up its own thin order book would fill worse and slip more. Each of the seven made the call that serves its own users, and the aggregate is a fact about where the liquidity is, not about anyone behaving badly.

And almost nobody holds two crypto cards for trading redundancy. People hold a second card because the first one gets declined abroad. The diversification this page says you do not have is one hardly anyone was looking for. The argument survives in a narrower form: if you hold two and assume the desks are independent, they are not.

The harm has also not happened to anyone here. The three incidents are real and dated, and all three predate most of these integrations, so nobody in this dataset lost money to venue concentration. It is a tail risk written down before it fires, which is both the reason to publish it and the reason to keep it in proportion.

None of that moves the count. Seven of the nine still execute in the same place, and none of the nine says how many of the others do.

Questions this research answers

Methodology

We started from every card in the Sweepbase dataset that a person can obtain today — 139 of them — rather than from a shortlist, because the prose in our own database is not a usable signal here. The word “trading” appears on 16 live rows, seven of them exchange cards whose desk was never in question, and it is missing from rows that do belong here. Matching on it would have produced a different list, wrong in both directions.

Each card was checked against its issuer’s own site, docs and help centre, and every inclusion carries a verbatim quote with the URL it came from. Claims that only third parties made were rejected — a claim about an issuer has to come from the issuer.

Two things had to be true of every row, and the second is the one that keeps catching us out: the quote has to exist on the page cited, and that same page has to carry the figures the row publishes. A real sentence attached to a page that does not support the row is still a sourcing error. Every row was read three times against that standard, each time by someone who had not seen the previous read.

What we have not done, in two parts. We have opened a position on none of these desks, so every fee, liquidation and balance claim here comes from documents rather than from trading. And we have not installed all nine apps: two are ours to look at — Tria, whose card the author holds, and Jupiter, where the Perps tab was checked in an account we own — and the other seven are described from what the issuer publishes about the screen.

Corrections

Four belong here, all from 5 August 2026, and naming them is the only reason to believe the rows we did not have to correct.

The largest is Ledger’s geography. This page said the desk names five restricted jurisdictions and the card is sold in two of them. It is sold in all five: France and Belgium are EEA states, Ontario is in Canada, and the card’s own row lists the EEA and Canada. So the group that gets a card and no desk is larger than we printed, and the error ran against this page’s own argument rather than for it.

The second is an attribution. We wrote that the venue’s own account posted only that uptime held, while the description of the October auto-deleveraging came from the founder’s personal account. Both statements sit in the same post from that personal account. The facts are unchanged; the split was ours.

Then three numbers. The fixed part of funding was given as 1.3% of margin a day and it is 1.2%, because our figure re-compounded a rate the same paragraph had already called simple. Jupiter’s row had positions collateralised by the JLP pool, which is backwards: the trader posts the collateral and borrows from the pool. And “trading” appears on 16 rows of our database, not 15. It became 16 the day an unrelated edit added the word to Revolut’s countries field, which is the argument against matching on prose in one line.

Last, Revolut’s client money. We described CFD cash as sitting in a segregated client-money account at an eligible third party and stopped there. Revolut’s own legal page carries the rest of the sentence: up to 20% of that value may be held on accounts inside Revolut Group. The omission flattered the one card here we hold up as the legally separated case.

Sources: issuer sites, product documentation and help centres, linked and dated on every row of the card table above, all re-read on August 5, 2026. Card-level facts come from the Sweepbase dataset; scoring is explained in the rating methodology. Spot an error? Report a correction.

About the author: Mihail B. is the founder and editor of Sweepbase. He holds the Tria Card, which appears on this page and is one of two cards here that Sweepbase has an affiliate arrangement with — the other is Hyperbeat. This page carries no affiliate links; the category listing does. How we verify is described on how we test.