The “Cheapest AI Inference” Provider Was Just a 20x OpenRouter Markup

CrofAI (crof.ai, nahcrof.com) was one of those ads. It looked like a miracle for indie devs, an inference provider selling access to frontier models like kimi-k3 at prices that undercut even the most aggressive OpenRouter listings.

There was just one problem. The “custom inference engine” was a Chrome tab pointed at OpenRouter.

On September 13, 2026, an investigator named KTibow dropped a detailed exposé showing CrofAI wasn’t running any inference at all. It was a wrapper. Every “cheap” token was just a passthrough from OpenRouter, often with a 20x markup on the model you thought you were renting. The fallout was swift: CrofAI denied everything, then “shut down”, then promised a refund, then deleted its entire online presence within 72 hours.

Let’s break down exactly how this happened, why it almost worked, and what it means for anyone buying inference credits from providers that don’t own their hardware.

The Tell: An OpenRouter Fingerprint Where It Shouldn’t Exist

The investigation started with a simple observation. KTibow noticed that CrofAI’s API accepted the openrouter:advisor tool, a server-side utility that only exists on OpenRouter’s platform.

That’s a bit like finding a McDonald’s wrapper in a gourmet burger kitchen. It’s possible there’s a reasonable explanation, but you should probably start asking questions.

What KTibow found was that CrofAI’s claims of “custom inference engines” were fiction. The service was silently routing requests to different, and often much weaker, models. The full mapping was documented in the exposé’s GitHub attestations:

The irony is almost poetic. CrofAI’s entire pitch was that they were undercutting competitors by running their own optimized inference stack. The reality? They were just buying credits from OpenRouter, pretending the models were self-hosted, and charging you a premium for the privilege of being lied to.

And it wasn’t just the models that were fake. CrofAI’s own “Greg” model family, marketed as proprietary innovations, was routed to other open-weight models:

In DMs, CrofAI admitted the entire greg family was a lie.

The Five-Shot Cover-Up: How an Exposé Turned Into a Comedy of Errors

The most damning part of the investigation wasn’t the initial proof. It was what happened next.

KTibow gave CrofAI a heads-up and a lengthy grace period to fix the issue. Then they tested again. Then again. Five times. Here’s the play-by-play from the investigation:

After each attempt, the code got more elaborate at hiding fingerprints. But the smoke always remained.

The Hardware Math That Never Added Up

If the fingerprints were the “how”, the hardware claims were the “why” anyone should’ve been suspicious earlier.

CrofAI claimed to run kimi-k3 on RTX PRO 6000s rented via Vast. That model requires ~802 GiB of VRAM even at the “lobotomy level” (Q2_K) quantization. The largest RTX PRO 6000 machine on Vast has 8 GPUs total, 765 GiB. Not enough. At full precision, K3 would need to be split across multiple machines with zero NVLink between them, which would make inference catastrophically slow.

CrofAI also claimed to be running deepseek-v4-flash-0731 on a local DGX Spark during the “investigation.” The Spark has 128GB of RAM. The model weights are 167GB. At the Q8_0 quantization CrofAI advertised, it’d be 323GB.

It’s not just implausible. It’s physically impossible.

Meanwhile, CrofAI’s May blog post had already admitted that daily costs above $1,000 killed their subscription model. In September, they were claiming 2-3x more hardware with lower prices.

The math never worked, because the “hardware” never existed.

Why This Is Bigger Than One Scam

Zach Moskow, an AI infrastructure commentator who called out CrofAI on X (and subsequently deleted his account), put it bluntly in a post about the incident:

“If you cannot tell a customer exactly who is serving their inference, what model they are receiving, and where their data goes when capacity fails, you are not selling inference.”

He’s right. And the problem isn’t just CrofAI. The AI inference market is currently full of what one commenter called “drug dealers who step on their tokens”, providers who undercut the market and hope their cheapness attracts enough volume to avoid a forensic audit.

What CrofAI did was obvious in retrospect. Instead of a subtle bait-and-switch that might’ve gone unnoticed for years (like serving a slightly cheaper but better model, which some commenters suggested was the rational approach), CrofAI got greedy. They served much cheaper, much weaker models and pocketed as much as a 20x margin. They got caught because the fingerprints were easy to find, the openrouter:advisor tool, gen- response IDs, images served from Alibaba Cloud IPs.

The question is: how many other “cheap” providers are doing this slightly more carefully?

What CrofAI’s Collapse Teaches Us

The CrofAI saga becomes a useful checklist for anyone choosing an inference provider:

One commenter noted that “the clean way to run this scam was to serve better, cheaper models on the Pareto frontier”, e.g., serving GLM 5.3 Flash when someone requests an older, worse model. That’s harder to catch. That’s also a reason for buyers to be even more skeptical of below-market prices from unproven providers.

The Flimflam Legacy

At the end of the day, CrofAI’s owner confirmed his own guilt in DMs, then deleted everything. The domain now returns a 404. The Twitter account is gone. The subreddit is private.

One detail worth reflecting on: the owner’s Discord name was “Devious Flimflam.” Merriam-Webster defines flimflam as “deception, fraud.”

He literally named himself “Devious Fraud” and still got hundreds of developers to hand him money. In a market where tokens are priced like a commodity and trust is the only differentiator, that’s a sobering reminder.

If you bought CrofAI credits, you should file a chargeback for every transaction, fraud victims are entitled to full refunds even for used services. And if you’re evaluating a “cheap inference provider” right now, ask yourself: if the model weights don’t fit on the hardware they claim to rent, what exactly are you paying for?

Probably just a wrapper around someone else’s API with a 20x markup.