Back to Blog

How to protect a free AI video tool from GPU credit farming

GPU credit farming burns cash before revenue arrives and pushes paying customers down your queue. The arithmetic, the queue fix and which basics break.

This message says 140th in line, and it came from someone who pays you.

Hi, I pay you $49 a month and my 5-second clip says it's 140th in line, about 42 minutes. Yesterday the same clip was done in six. I have to send this to a client at 3:30. Who are the 139 people in front of me? If it isn't done in time, please refund this month and I'll find something else.

Say that reaches you at 2:45 p.m. on a Tuesday. It's a scenario, so swap in your own numbers. You open the render queue.

Every new account gets 100 free credits, enough for five 5-second clips. A clip holds one GPU for 6 minutes, and you pay $3 per GPU hour.

  • $3 ÷ 60 minutes = $0.05 per GPU minute
  • 6 minutes × $0.05 = $0.30 a clip
  • 5 clips × $0.30 = $1.50 for every account that spends its credits

Since 12:45 p.m., a new account has appeared every two minutes. Each one renders its five clips and never comes back. That's 30 accounts and 150 clips an hour, on top of the 120 clips a normal afternoon brings. Your fleet is capped at 20 GPUs, which finish 200 clips an hour (20 × 60 ÷ 6).

  • 120 + 150 = 270 clips arriving each hour
  • 270 minus 200 = 70 clips added to the line each hour
  • 70 ÷ 200 an hour = 21 more minutes of waiting, every hour

Two hours of that is 140 clips, and your customer is standing behind them.

Wait for a new render, illustrative0204060801001201400123456Minutes in lineHours since it started
Wait for a new render, illustrative0501000123456Minutes in lineHours since it started

You raise the cap, because nothing else is within reach. It takes 27 GPUs just to stop the line growing (270 × 6 ÷ 60), and the line doesn't get any shorter. The extra clips cost you $45 an hour (150 × $0.30), which is $1,080 a day. That's what 22 customers on your $49 plan pay you in a month (22 × $49 = $1,078).

This is the part of GPU credit farming that stings. Your GPU bill arrives before any revenue does, and none of it comes back. You can delete all 60 accounts and zero their credits, but the clips they finished ran on time you've already bought. A farmed credit turns into spent cash the moment it renders.

Say you have 300 customers, or $14,700 a month. A full month of this is $32,400 ($1,080 × 30), more than twice your revenue. If you have $60,000 in the bank and lose $15,000 a month, you had four months of runway. Add the farm and you have about 38 days ($60,000 ÷ $47,400 × 30).

By 5 p.m., 14 customers have written in, and refunding them costs $686 (14 × $49). The one who wrote at 2:45 sent their client an apology instead of an ad, and they don't renew. That evening you're deciding whether to switch off the free credits that brought you nearly all 300 of those customers.

That was a scenario. This wasn't:

From our early beta

Fake users drained one team's shared model provider quota. Their paying customers couldn't use the product until the team raised its budget to cover the abuse.

Your 27 GPUs would be the same trade: capacity for paying customers, bought back at your expense.

Why GPU credit farming hits video tools so hard

Credit farming means collecting your free credits across many accounts. Someone might want clips for a channel they run, or clips they can sell. Either way, you pay.

A chat reply holds a model for seconds, while a clip holds a whole GPU for minutes. So the cost of one free user is high, and each of those minutes is a slot a paying customer can't have.

150 free clips an hourEvery GPU stays busyOne 5-second clip200 clips finish an hourStarts 42 minutes later140 clips aheadFarmed accountsRender queueGPU fleetPaying customerFarmed accountsRender queueGPU fleetPaying customer
150 freeclips anhourEveryGPUstaysbusyOne 5-second clip200clipsfinishanhourStarts 42 minuteslater140clipsaheadFarmedaccountsRenderqueueGPUfleetPayingcustomerFarmedaccountsRenderqueueGPUfleetPayingcustomer

None of those accounts broke a limit. Each one spent exactly the credits you gave it. That's why more GPUs don't hold as a fix: another account shows up two minutes later with another 100 credits, and extra capacity only finishes its clips sooner. You have two separate problems here, who waits and who gets the credits. Start with who waits, because you can fix that today.

Protect paying customers with a separate GPU queue

Price a credit in GPU time. Use real minutes from your logs, failed renders included.

cost per free account=free clips×GPU minutes per clip×price per GPU minute\text{cost per free account} = \text{free clips} \times \text{GPU minutes per clip} \times \text{price per GPU minute}

Reserve GPUs for paying customers. About 70 of your 120 hourly clips are paid, and that needs 7 GPUs (70 × 6 ÷ 60). Reserve 9 so a busy hour fits, and never let a free job touch them. The two spare GPUs cost $144 a day (2 × $3 × 24), which is less than three $49 refunds.

Cap the free lane. The other 11 GPUs finish 110 clips an hour. Allow one render at a time per free account.

Give free work a daily budget. 11 GPUs at $3 is $33 an hour, so $792 is the most a day can cost. Add a hard spending cap and a switch that pauses new free jobs and tells people why.

Make free clips cheaper. If GPU time scales with length, a 3-second clip takes 3.6 minutes and costs $0.18. A watermark makes a clip harder to sell, but the GPU has already run, so it won't lower your bill.

Hold renders until the email is confirmed. Then an unconfirmed signup costs you one email and no GPU time.

Here's the same afternoon, two hours in:

One shared lineTwo lanes
Paying customer's wait42 minutesUsually none
New free user's wait42 minutesAbout 98 minutes (180 ÷ 110 × 60)
Most you can spend on free jobs each hour$60$33
Free GPU time going to the farm75% (150 of 200 clips)75% (150 of 200 clips)

Your paying customers are safe and the bill has a ceiling. But an honest person trying your tool for the first time now waits 98 minutes behind the farm, and they'll close the tab. The last row is still $594 a day ($33 × 75% × 24). Lanes decide who goes first, and the farm still gets its credits.

The basic checks against repeat accounts, and how they break

Each of these catches something, and each one has a price.

Email checks. Verification catches typos and dead addresses, but it only proves someone can open an inbox, and a second real inbox passes, while real people lose codes to spam. A unique-email rule catches the same address twice and almost never hurts anyone. It compares text, though, and one mailbox can have many spellings. Gmail ignores dots, so sam.lee@gmail.com and samlee@gmail.com are one inbox, and samlee+clip2@gmail.com arrives there too. That's how aliases work. Disposable-domain blocks catch known throwaway services, miss new domains and ordinary mailboxes, turn away people using privacy forwarding, and need a list someone updates forever.

Signup friction. CAPTCHA stops simple scripts. It doesn't stop a person opening accounts by hand, it trips up people using screen readers or privacy tools, and it has other limits. A phone number or a card stops casual second accounts, and a card requirement can cut farming a lot. Anyone with another number or card still gets through, you pay for every text, and students and people who want to try first walk away at the card form.

Network rules. One account per IP address catches repeat signups on one connection and nothing on any other. It also blocks agencies, classrooms and households that share an address. VPN blocks catch known ranges, miss everything off your list and shut out staff on company VPNs.

Limits. Rate limits stop bursts from one account, and a farmed account only needs five renders, so it never gets near one. Set one too tight and it hits real users in a busy hour. A per-account cap has the same gap, because a new account starts with a new cap. Cutting free credits from 100 to 40 means two clips. Video often takes a few tries to get right, so honest people run out before they've seen a good result, while someone after five clips opens three accounts and spends the same GPU minutes.

Manual review. A person spots things rules can't, until the volume shows up. At 30 accounts an hour you have 720 a day, and at one minute each that's 12 hours of someone's day. The clips have rendered before anyone looks, and a tired reviewer turns away people who just look odd.

If you're doing this yourself, start with lanes, the budget and confirmed email, then add the email rules and treat them as a job that never finishes. Write down what one person gets, too, because none of these checks can tell you whether the person behind a new account has already had their credits.

Evaluate the claim before the GPU starts

Whether this claim should go through is the question AbuseZero exists to answer. The lists and limits above each look at one small thing. AbuseZero puts 140,000+ signals through multi-layer AI risk analysis, which detects linked claims and flags abuse that single checks miss. There's also a proprietary algorithm, with AI that checks the claim again. Those signals are read together, never one at a time. Where your domain list goes stale, AbuseZero checks mailbox aliases, disposable domains and mail routing. Where your one-per-IP rule stops, it sees beyond an IP address with VPN, hosting, network reputation, Tor and proxy intelligence. It recognizes repeat devices across new accounts, even when email addresses keep changing. Creating a fresh account isn't enough to erase previous activity.

You make one API call before the first render is queued and get a clear allow or deny, under rules you control. Observe mode shows you what would be denied before anyone is blocked. You don't have to build your own abuse prevention system, and the quickstart protects your first benefit. Keep your lanes and your budget either way, because they protect capacity whatever causes the next spike.

Frequently asked questions

What is GPU credit farming?

It's collecting a tool's free credits across many accounts, so GPU time you pay for keeps going to someone who won't pay you.

Does a priority queue stop credit farming?

No. It protects paying customers from the wait, and the free credits still go out. AbuseZero evaluates each claim before they do.

Should I remove free video generation?

Keep that as a last resort, because it's often how people find you. Work through the decision before you switch it off.

Save your GPUs for the people you built this for

Keep reading

Copying your prompt…

Copied the to your clipboard. Give it to your coding agent & it will handle the rest.