How to stop fake startups claiming your startup program credits
One fake company can cost what 500 free trials do, and the credits are spent before anyone looks. Which basic fixes break, and how to run the program anyway.
You've stopped reading the credits ledger and started opening the company websites, and the sixth one has the same photo of two people at a whiteboard as the second.
Both companies are in the current cohort of your startup program. In this scenario, you run a GPU inference API, and the program gives an early-stage company $5,000 in credits the day it's approved. Both of these were approved on October 1, and both balances hit zero by October 11. By the end of the 40 approvals there are 14 you can't explain: a page with no product behind it, founders with no history, credits gone within days. You can't say whether that's one operator or four. You can say it wasn't 14 startups.
Nobody writes a runbook for startup program credits abuse, so you open a spreadsheet. Swap in your own rates as you read. This scenario assumes a GPU-hour sells for $2.50 and costs you $1.50 to run.
$5,000 grant / $2.50 = 2,000 GPU-hours
2,000 hours x $1.50 = $3,000 of your cash per grant
14 grants x $5,000 = $70,000 in credits
Balance you can still void = $2,180
Credits already spent = $67,820, or 27,128 GPU-hours
27,128 hours x $1.50 = $40,692 you've already paidYou void the balances, and $2,180 disappears from the ledger. The other $67,820 stays gone. Those GPUs ran, your compute invoice is due on the 1st, and there's nobody to bill, because the companies don't exist. If your company spends $85,000 a month, $40,692 is about two weeks of runway (40,692 ÷ 85,000 × 30 = 14.4 days).
It also explains October. A grant burned in 10 days is 2,000 hours ÷ 240 hours, or about 8 GPUs running without a break. Fourteen of those at once is about 117 GPUs. Say your pool is 400 GPUs and paying customers normally keep 310 busy: 310 + 117 = 427, which is 27 more than you have. For those 10 days, paying customers' jobs waited in line behind companies that don't exist. Two customers on your $2,000 plan complained, you gave each a $500 service credit, and now you know what you were apologizing for.
The tab you've been avoiding is the waitlist. The cohort was capped at 40, so every one of those approvals took a place from someone else, and there are 31 names on that list. One of them wrote last week: "You said the cohort was full. Any chance a spot opened up? We've got about six weeks of compute money left." That's a two-person team with a working demo. Say 1 in 4 startups in your program goes on to pay, at $1,500 a month:
- 14 lost places × 25% = 3.5 customers
- 3.5 × $1,500 = $5,250 a month
- $5,250 × 12 = $63,000 a year that will never appear on any report
The next cohort opens January 5, and it's already on your website. You can limit it to accelerator referrals, which shuts out the team that sent that email. You can halve the grant, which every honest startup pays for. Or you can open it as it is and come back to this spreadsheet in April.
Why startup programs attract fake companies
The size of one claim is what makes this different. A free trial on this API is $10 in credits, or 4 GPU-hours, which costs you $6. A real startup spending $250 a month costs you $150 a month, or $1,800 over the year, and never finishes its grant. A fake company draws the whole thing at once.
That first bar isn't missing. It's $6. One fake company costs what 500 free trials do ($3,000 ÷ $6), and it takes one form.
The evidence is also thin by design. You built this for companies with no revenue, no track record and a website a few weeks old, so real applicants and invented ones hand you the same form. And the reward goes out before anyone learns anything, because a grant that's been spent is compute you've already paid for.
Not every bad claim is an invented company, either. A real company whose grant ran out can come back under a new name, much like one company starting trial after trial.
The obvious fix is to ask for more proof. But every extra proof lands hardest on the applicant with the least to show, and that's who you built the program for.
The basic fixes, and how each one breaks
Start with email, since that's what the form collects. A unique-email rule compares strings, and one mailbox can have many. Gmail ignores dots, and plus-tagged mail reaches the same inbox. That's simply how email works.
| Basic fix | What it catches | What it misses | Who it wrongly blocks | What it costs to keep working |
|---|---|---|---|---|
| Email verification | Typos and dead addresses | Any real inbox on a new domain | Founders whose code lands in spam | Deliverability work |
| Work email on the company's domain | Applicants on free mailboxes | Anyone who bought a domain | Founders who haven't set one up | Explaining the rule on every no |
| Disposable-domain blocks | Known throwaway services | A domain bought this week | People using privacy forwarding | A list someone updates forever |
| Gmail dots and plus addressing | One Gmail inbox written several ways | A second real mailbox | People who sort their mail with plus addresses | Upkeep as mail services change |
| CAPTCHA on the form | Simple scripts | One form filled in by hand, which $5,000 easily pays for | People using screen readers | Fees and lost applications |
| Card or phone required | Casual applicants | Anyone with another card or number | Founders with no company card yet | A charge per text, fewer applications |
| One application per IP address | Repeats on one connection | Any other network | Coworking spaces and accelerator cohorts | Exceptions and tickets |
| VPN blocks | Known VPN ranges | Everything off your list | Remote teams on company VPNs | Another list to maintain |
| Incorporation documents | Invented names | Registered shells | Teams still incorporating | Reading paperwork from many countries |
| Accelerator referrals only | Most walk-in applicants | Partner codes that leak | Every bootstrapped founder | Partners to manage, codes to rotate |
| Shorter credit expiry | Credits sitting unused | A grant spent in 10 days | Startups that spend slowly | Fewer reach a paid plan |
| Manual review | What a person can spot | A well-made page | Whoever looks thin to a tired reviewer | Hours every cohort |
Manual review is where most programs end up. At 300 applications a quarter and 20 minutes each, that's 100 hours (300 × 20 ÷ 60), or two and a half working weeks. The page that fails is often the real one, from a founder who's been writing code instead of a landing page.
Say you add incorporation documents for January. A two-person team halfway through incorporating drops out without telling you. A registration proves a company exists on paper. It doesn't tell you whether anyone's building anything, or whether those same people had your credits last cohort.
How to limit startup credit abuse without closing the program
None of those fixes is enough, so change what a single bad approval can cost.
Price the grant in cash. A cohort at full draw here is 40 × $3,000 = $120,000. Work out your own unit cost before you set the size.
Release credits in stages. Give $500 on approval and the rest in monthly amounts once there's real usage. A startup spending $250 a month won't notice. Fourteen fake companies would each cost you $300 at that stage, so $4,200 in total.
Give program accounts their own capacity, with a hard cap, so a bad cohort can't push paying customers to the back of the line.
Keep fit and repeat claims apart. A reviewer can judge whether a company fits your program. Whether this claim is a repeat is a separate question, so keep one record of every grant across every cohort.
Put it in your terms. Credits can't be transferred or resold, and unused credits can be voided, so you can take back what's left.
Decide what a no looks like. Real founders will get caught, so give them a reason and a way to reply.
Repeat claims are the hard part. Your records only know a name or domain they've already seen, and a new application arrives with nothing attached to the last claim. At 300 applications a quarter, keeping up with that is a job of its own.
Evaluate the claim before credits move
You don't have to build that job yourself. A bought domain clears your work-email rule, a new application gets a fresh monthly cap and a 20-minute review only ever sees one form, but with AbuseZero, creating a fresh account isn't enough to erase previous activity. AbuseZero flags abuse that single checks miss. Its multi-layer AI risk analysis examines 140,000+ signals and detects linked claims. Where a card-on-file rule ran out, AbuseZero detects repeated payment fingerprints across users, even when new accounts and emails hide the connection. It recognizes repeat devices across new accounts, even when email addresses keep changing, and its email rules cover mailbox aliases, disposable domains and mail routing. You also get a proprietary algorithm, with AI that checks the claim again.
Your server makes one API call before a grant is released and gets a clear allow or deny, under rules you control. Your reviewer keeps the question only a reviewer can answer, which is whether a company fits. Eligibility can be once per user, ever, you can import past cohorts as claim history, and Observe mode shows what would be denied before any founder is turned away. The quickstart protects your first benefit.
Frequently asked questions
What is startup program credits abuse?
It's claiming credits meant for early-stage companies without being one, or claiming them more than once, usually through invented companies or repeat applications.
How large should a startup credit grant be?
Can a basic check tell whether a startup is real?
No. A new company has little to show, so that judgment stays with your reviewer. AbuseZero evaluates whether each claim should go through and returns allow or deny.
Keep your credits for the startups you built this for
Start with Builder
Pay as you go. Protect your first benefit today.
See pricing
Builder, Pro, Scale and Enterprise.
Talk to us about Enterprise
High-volume programs and custom terms.