Key takeaways
- AI spend breaks normal cards for three reasons at once, which is what makes it hard to pay for AI tools on one shared number: it is usage-based and spiky, almost entirely US-dollar-billed, and spread across dozens of vendors.
- The old card stack taxes it twice. It adds 1 to 3% on every dollar-billed invoice, then declines mid-run when a token bill spikes past a shared limit.
- The fix is a payment layer built to pay for AI tools: a per-vendor virtual card with a hard limit, USD paid at $0, and funding that settles in under 5 minutes.
It is a Saturday night and your batch job is running. Halfway through, the API card declines. The limit you set in January assumed a $4,000 month; tonight's run alone is $6,200 in tokens because a customer's workload tripled. The job dies, the card is frozen for review, and your finance lead is asleep. Monday, the statement arrives with a second surprise: a 1 to 3% foreign transaction fee on nearly every line, because almost every AI vendor bills in US dollars and your card treats dollars as foreign (Capital One). You did not overspend. Your payment layer just was not built to pay for AI tools the way AI companies actually spend.
What do AI companies actually pay for?
AI spend falls into three buckets: API tokens billed per million tokens, GPU and cloud compute billed per hour, and AI SaaS billed per seat or per run. Almost all of it is priced and charged in US dollars, and most of it moves with usage rather than a flat monthly number.
The mix differs at every company, but the pattern holds. Here is what it takes to pay for AI tools at a typical AI-first team, and how each line bills.
| What you pay for | Examples | Billing model | Currency |
|---|---|---|---|
| API tokens | Anthropic, OpenAI, other model providers | Per million tokens, spiky | USD |
| GPU and cloud | AWS, GCP, CoreWeave, Modal | Per GPU-hour, bursty | USD |
| AI SaaS | Coding assistants, observability, vector DBs, eval tools | Per seat or per run | USD |
Two of the three buckets are usage-based, so the bill moves with your traffic. All three are dollar-billed. And a mid-size team easily runs 20 to 40 of these vendors at once. Those three facts, stacked, are what make it uniquely hard to fund on a normal card. If you want the underlying numbers, we break down the Claude API's pricing and GPU and cloud pricing in their own posts.
Why is it so hard to pay for AI tools?
Because the spend has three traits that a normal corporate card was never designed for.
It is spiky. A traditional SaaS subscription is $99 a month, every month. A token bill is $2,000 one week and $9,000 the next, driven by a customer launch or a runaway agent loop. A fixed card limit that fits the average month will decline the peak.
It is dollar-billed. Your team may be in London, Bangalore, or Toronto, but your model provider, your GPU host, and your eval tool all invoice in USD. If your card charges a foreign transaction fee, you pay it on the whole stack, every month, so you pay for AI tools at a premium a USD card would never trigger. See how that compounds across foreign transaction fees on your stack.
It is sprawled. One card behind 30 vendors means no per-vendor control, one number to leak, and a monthly reconciliation that reads like archaeology. Managing that sprawl is the core job of SaaS spend management.
What are the three ways the old card stack fails?
Line the three traits up against a normal card and three predictable failures appear the moment you pay for AI tools with it. Each has a specific fix.
| Failure mode | What it costs you | The fix |
|---|---|---|
| FX surcharge on USD | 1 to 3% on every dollar-billed vendor | A card that charges $0 on USD |
| Declines mid-run | Dead jobs, frozen cards, lost hours | A per-vendor card with headroom you can raise in minutes |
| No per-vendor control | Leaked numbers, blind reconciliation | A separate virtual card and hard limit per vendor |
None of these is exotic. They are the default experience when you pay for AI tools on a card built for expense reports.
How do you avoid foreign transaction fees on US-dollar tools?
Pay the invoice with a card that treats US dollars as US dollars. Most business cards add 1 to 3% on any vendor billed outside your home currency, and for a non-US team that is nearly the entire AI stack (Capital One). On a $50,000 monthly AI bill, 2% is $1,000 gone before you have run a single token. Over a year that is $12,000 you handed to your card network for doing nothing.
The fix is a card that charges 0% on USD spend. Endl cards charge $0 on USD, so you pay for AI tools at the sticker price of the invoice, not the sticker price plus a network cut. If you spend in another currency, it is the Visa rate plus a flat 1%, with no hidden markup layered on top.
How do you stop a spike from taking down every subscription?
Stop sharing one limit across everything. Give each vendor its own virtual card with its own hard cap. When a token bill spikes, only that card hits its ceiling. Your GPU cluster, your observability tool, and your source control keep charging on their own cards, because they are not sharing a pool that one spike can drain. It is the most reliable way to pay for AI tools whose bills you cannot predict from one month to the next.
This is what virtual cards for business are for: one card per vendor, a limit that matches that vendor's real spend, and an instant freeze the moment a number leaks or a trial you forgot renews. The same structure is what makes how to pay for the OpenAI API predictable instead of a monthly guess. And as more of your spend gets initiated by software rather than people, per-card limits become the guardrail for agentic payments: the agent can spend, but only up to the ceiling you set on its card.
How should an AI company set up its payments?
This is how AI companies pay for AI tools without handing a cut to the card network. Four moves, in order.
Issue one virtual card per vendor. Do not run 30 tools behind a single number. One card per vendor gives you a clean limit, a clean freeze, and a clean line on the statement.
Set each limit to real spend plus headroom. Look at the vendor's last three months, take the peak, and add a margin for the next spike. That way normal growth does not trip a decline, but a runaway loop still hits a wall.
Pay USD at $0. Route the dollar-billed stack through a card that does not surcharge USD. That single change claws back the 1 to 3% the old stack was skimming.
Fund fast, so a spike is a top-up, not an outage. Keep the funding rail quick enough that raising a limit on a Saturday takes minutes, not a Monday wire. Done together, these four moves are all it takes to pay for AI tools without FX drag or surprise declines.
Three ways the old card stack taxes AI spend
FX surcharge
1 to 3% on every dollar-billed tool, on a foreign card.
Declines on spikes
A usage spike trips one shared card and stops everything.
No per-vendor control
One number, no limits, no clean reconciliation.
Fix each with a USD-native, per-vendor virtual card and the settlement tax on your tool stack mostly disappears.
How Endl fits
Endl is the payment layer for that setup, built for how AI companies pay for AI tools. You get virtual and physical cards with per-vendor and per-card limits and an instant freeze, so each AI tool gets its own card and its own ceiling. USD spend is $0. Other currencies are the Visa rate plus a flat 1%, with no hidden markup.
The balance behind the cards is self-custodial and funded on stablecoin rails, which settle in under 5 minutes, 24/7. So when a run needs headroom on a weekend, you top up and raise the limit in minutes instead of waiting on a bank window. Adding funds runs at a flat 0.5%, and payouts reach 160+ countries when you are paying a contractor or a vendor who is not on a card. One note for the record: Endl cards are debit and spend cards drawn against your own balance, not credit, and Endl is not a bank and is not insured. It operates as a registered VASP in the EU and an MSB in Canada.
If you want the numbers before you commit, the pricing page lists every rate, and you can start free and issue your first per-vendor card to pay for AI tools today.
