$500M ARR, 60 Engineers, Cash-Flow Positive: How Higgsfield Actually Runs, With CEO Alex Mashrabov

We use Higgsfield almost every day at SaaStr AI. If you’ve watched any of our short videos, the intro to The Agents with me and Amelia, the event promos, the sponsorship teasers, all of it is built on Higgsfield. We’re also a proud early investor through SaaStr Fund. So when Alex Mashrabov, the co-founder and CEO, came on stage with me, I wasn’t playing analyst. I’m a customer, an investor, and a heavy user.

Higgsfield crossed a $500M annualized revenue run rate in June, up from $200M at the end of 2025, and it’s cash-flow positive. The platform didn’t exist before March 2025, so that’s a half-billion-dollar run rate in roughly 15 months from launch. They’re now in talks to raise at a reported $5B valuation. When Alex and I sat down the number was $300M. 60 days or so later, it had crossed a $500M run rate.

Alex sold his last company, AI Factory, to Snap for $166M and ran Generative AI there before starting this. So this isn’t a first-timer stumbling into a hit. But the curve is still one of the fastest I’ve seen up close.

What I wanted out of him wasn’t the highlight reel. Everyone’s heard the “AI company hits $200M in a year” headlines. I wanted the story behind the story. Here’s what came out.

$500M ARR on a Team of ~150

They’re doing this with about 150 people. Roughly half engineering, half a large creative team. Core engineering and product is around 60 people.

Alex pegged their efficiency at roughly $5M in ARR per engineer against a typical $2M, so two to three times more efficient than a normal software company at scale. He said that when the run rate was $300M. At $500M the ratio only widens.

The reason the creative team is that large is the real lesson. Alex made a deliberate bet: pair engineers directly with 70-plus creative professionals, people who made commercials and ads for a living. Not prompt engineers. Filmmakers. Every tutorial and asset Higgsfield ships is generated on their own platform, which is how they figure out what’s actually usable versus what only works in a demo. The creatives tell the engineers where the models fall apart. The engineers fix it. That loop is the product.

He was also honest about where the efficiency stops. Vibe coding and web-based tools are great for shipping features fast. They are not good enough yet for the deep infrastructure work, the stability, the safety, the anti-fraud. As they scaled, the engineering team had to grow specifically to handle that. Speed gets you to $50M. Craftsmanship keeps you there.

They Started With One Model. That Was a Mistake.

Higgsfield launched on their own model. Alex called it “open source plus plus plus.” Within weeks they realized a new video model was landing basically every week, and no single lab was going to win every use case. Chasing that from inside a single model was a losing game.

So they pivoted to aggregation. Today you open Higgsfield and you can run Google Veo, Kling, Seedance, their own models, whatever is best, side by side, in parallel. I do this constantly. I’ll fire the same prompt across three models in seconds and pick the winner. On the surface the product got more complex. In practice it got far more powerful, because the platform picks the best model per use case instead of forcing me to.

This is where the “thin wrapper” question always comes up, and Alex had the cleanest answer I’ve heard. His view: almost every software company is going to run on AI models it doesn’t own, so the wrapper framing is mostly noise. The moat isn’t the model. It’s two things. One, collaboration and network effects, the thing that made Figma and Canva what they are. Two, helping brands sell more product, which is why they built an MCP integration and just launched a marketing agent called Supercomputer that pushes creative straight into Meta and other ad networks.

5x Canva’s ACV, Doubling Every Quarter

The average Higgsfield customer spends around $1,000 a year. Canva is around $200. That’s 5x the ACV, against a company with enormous scale and a decade-plus head start. And Higgsfield is nearly doubling ACV every quarter as they move up market.

I don’t count pennies on Higgsfield. It’s cheap relative to the value. I’ve been a Canva customer forever and I pay $18 a month and love it. I pay Higgsfield more and don’t care, because the alternative to a video I make in 60 seconds is hiring an agency, waiting two weeks, getting something mediocre, paying thousands, and deleting it because I’d never use it. That’s not a close call.

Underneath that ACV, the pricing structure is doing real work. About 40% of usage now runs through higher-level workflows like their cinema and marketing studios, not just raw model picking. The basic “mark up the model and run it efficiently” layer is real revenue, just lower margin. The upsell is the agentic workflow that replaces a contractor or an agency instead of just helping you make one asset. Alex is systematically marching customers up that stack. Marketing budgets at real companies run into the millions, so a $10K experiment against a $1M budget is a rounding error. His job is to make sure that first experiment works.

The Surprise: 70% of Revenue Is Agencies

I would have guessed Higgsfield’s early customers were web heads and AI natives. Wrong. Roughly 70% of that revenue is agencies, and it was agencies from early on.

The logic clicks once you say it out loud. Creative agencies have been struggling. AI gave them a new thing to sell and a way to become radically more efficient at the same time. A creative director who used to need a production crew, rented equipment, and a booked location can now make an ad end to end in a day. And the thing physical production can never do, swap the actor, change the lighting, generate ten variations of the same spot, is trivial in software.

I pushed him on the awkward part. Do agencies hide Higgsfield from clients? Do they white label it? His answer split the market. Smaller creative shops use it openly as their production engine. Larger agencies make their money on media consulting and media buying, so Higgsfield doesn’t threaten the core, and with Supercomputer they’re now moving into distribution too, buying and placing the ads, then reporting back on what worked.

How Alex Defines ARR

I ask every AI founder this now, because the definitions have gotten silly. Alex was straight about it.

Take annual subscriptions and divide by 12. Take the last four weeks of on-demand usage. Take monthly subscription revenue. Sum it, multiply by 12. It’s booked, recognized revenue, not cash-in-times-12, and not marketing credits given away at zero and counted as real. About 40% of revenue is on annual.

He flagged on-demand credits as the trickiest piece, because attribution gets fuzzy. My take, which he agreed with: people overcomplicate this. If Stripe says you did roughly $42M this month in real recognized revenue, you’re at a $500M run rate. Whether it’s annual, monthly, or one-off, show me the money. The dodgy version is discounting 80% and booking the list price. The honest version is cash properly recognized. He said they’ll likely start sharing Stripe dashboards directly, because otherwise everyone argues about definitions.

They Killed Most of What They Launched

The product 95% of users touch today is not the product I started on. Alex reoriented the entire company three separate times, and each pivot was driven by watching usage, not by a roadmap.

First bet: camera controls, in early 2025. Creative directors were rejecting AI video outright because there was no intention behind the camera. Controls were the wedge that got pros in the door. Second: one-click visual effects, drawing on the face-filter work from his Snap days, which took them to $10M ARR in about six weeks. Third: when he saw people making full commercial projects end to end, he reoriented around that, because AI video had stopped being a toy. Now the whole company is reoriented again around agentic workflows.