Higgsfield went from $1M to $1B in annualized revenue in 18 months. Alex Mashrabov says Cursor needed 24.
His team of close to 400 people spends over $4M a month on AI models, and one of them burned over $30K on a single model in a week. Alex still calls that week a net positive.
Before any of it, he sold AI Factory to Snap for $166M, then burned more than $10M of a $16M seed searching for a product. The first number he hands you is a cost:
“On average at Higgsfield, a person on the team spends over $10,000 a month on various models.”
Growing that fast also means hiring close to 400 people at the same pace, and that’s usually where fast companies stall.
Finding the right people to build your company takes time you don’t have. And when the best candidates move fast, you can’t afford to move slowly.
Meet fillmore by Metaview, your AI recruiting coworker that finds great candidates, researches their backgrounds, and runs personalized outreach and follow-ups. Interested candidates get booked straight into your calendar. Just approve meetings in Slack and show up for the calls.
You build the company. AI finds the people
Ambitious teams like throxy (YC P25) and Ambience Healthcare are seeing epic results:
▫️ Up to 49% candidate response rates
▫️ 7 minutes from calibration to live outreach
▫️ 24/7 sourcing, outreach, and follow-ups
Whether you don’t have a dedicated recruiter or need to kick hiring into high gear, fillmore delivers every time.
Book a free demo to see fillmore in action:
Alex sat down with Harry Stebbings on 20VC. Every figure below comes from Alex himself, so treat it as his claim unless I say otherwise. I went through the full interview and pulled out the 10 calls that matter:
1. How Higgsfield Counts $1B in ARR (18 Months, 2 Strict Rules)
Anyone can repeat the $1B. The counting rule behind it is the part to copy.
What’s very important is that we take revenue, not sales. If it’s an annual subscription or an annual enterprise contract, we prorate it across 12 months and take only the piece that corresponds to 1 month, 28 days to be precise. That’s the first piece. Second, it’s only live revenue. We are not taking 3-year enterprise deals and baking them into the 1 billion figure.
The formula is 4 weeks of revenue times 13. Two rules stacked on top make it hard to inflate: annual deals spread across 12 months, and live revenue only.
The climb took 18 months, third behind OpenAI and Anthropic and ahead of Cursor at 24. The Cursor breakdown’s timeline runs from October 2023 to November 2025, about 25 months, so the comparison holds.
The forecast is the better story. Asked about going from $1B to $10B in 12 months, Finance projects $4.5B, with substantial deceleration baked in. Alex pushes for at least 30% growth month over month and, asked for his own number, says over $10B.
The distance between the founder’s number and his own finance team’s is where I would start any diligence call.
Benchmarking your own climb against the fastest ARR curves, start here:
▫️ The Lovable growth playbook, stage by stage
▫️ Zero Ads. Zero VC. The story of Magnific
▫️ The growth playbook behind one of AI’s fastest ARR climbs
Run the live-only test on your own ARR before an investor runs it for you.
2. From $99 a Month to $6M a Year in 6 Months (And the 30% Month-1 Leak)
$99 a month. 6 months later, $6M a year.
1 customer started 6 months ago spending just a subscription, $99 a month. And now we just signed a deal over 6 million a year. This level of acceleration is mind-blowing to me.
Month-12 net revenue retention in the business segment is over 300%, a figure that never happens in B2B SaaS. Retention is the hardest metric for startups, and a cohort analysis model with an Excel template shows how to track it. The catch, which he names before you can: the company is 18 months old.
I would want cohort-level detail before I trusted that number.
What does $6M a year buy? 2 trends, both from Asia: direct-to-consumer brands publishing hundreds of ads, if not thousands, a week to A/B test, and short-form drama studios making new shows with AI end to end, an industry sized above $10B.
The top of the funnel leaks. About 30% of users drop in month 1, then churn flattens, a weaker start than the 80% month-1 logo retention the B2B SaaS era expected.
Sell the $99 plan as the door, not the business.
together with Metaview:
When one customer goes from $99 a month to $6M a year in six months, the team behind it has to grow just as fast.
fillmore by Metaview finds candidates and runs the outreach for you, then books the interested ones straight into your calendar.
3. Higgsfield Burned $10M of a $16M Seed Before Finding Product-Market Fit (8 Calls Fixed It)
The company that hit $1B nearly ran out of runway first.
We burned more than 10 million out of 16 million raised in seed fundraising. So we felt we had just 1 attempt left. And frankly, I feel I’m responsible, because I was focusing on the wrong things.
The year went to hype: the right narrative, hijacked attention, anything but a good product.
With less than $5M left, Alex stopped writing the story and started calling customers. 8 creative directors gave him the same answer: camera control does not exist in AI, and it decides whether a video can tell a story. Higgsfield shipped that gap as the product, and the fit was immediate.
He doesn’t blame the market, which is why I trust the fix.
“I think I just lost the touch with reality back then.”
Before the money gets thin, run the model-market fit and pivot checks, and keep a 13-week cash flow forecast open. Spending $40K on features no one wanted is the expensive way to learn what 8 customer calls tell you first.
Book those calls before you write another line of code.
4. Why Higgsfield’s Founder Says $20 AI Subscriptions Will Get Destroyed
Alex says his own company will never win the $20-a-month market.
We will never be able to win in a market of subscriptions of $20 a month. Today, Google and OpenAI pursue ads so much, but fundamentally I think they are going to completely demolish all the consumer subscription markets, the $20 a month subscriptions.
His logic: a strong horizontal product from Google or OpenAI absorbs the vertical $20 plan. It is a contrarian bet, and the host’s Canva example is the most apparent case. He is betting against the cheap tier while selling above it, so weigh the prediction that way.
Google and OpenAI will likely win the direction. The timing is the part I would discount.
Higgsfield’s answer is to climb. When a user starts at $20 a month, the goal is to show enough value that they upgrade to over $1,000 a year. Business is slightly over 50% of revenue and mobile is under 10%, and the freelancers and aspiring creators learning video AI are a little churny, though most come back within a year to try again.
Founders who price as if serving a customer costs nothing get squeezed first. How to pick the right monetization strategy is the next read.
Price against the free horizontal tool, not last year’s competitor.
5. Why Higgsfield Stopped Chasing Benchmarks (And Earns 80% Margins on Open Models)
Benchmarks score text-to-video. Production runs on something else.
The way to think about video models today, it’s just a modern rendering engine. Think about this as Unreal Engine or Unity, but just different types of inputs.
A prompt can’t define a visual output alone. The average prompt in Higgsfield’s open-sourced film runs over 3,000 words with at least 10 image references per scene, and the people labeling benchmark outputs can’t read prompts that long. Prompts this heavy are context engineering by another name, and a masterclass on prompting GPT-6 Astra covers the habit.
Alex calls his own benchmark chase a mistake. Higgsfield now trains models only where customers ask, such as photo shoots and product consistency, and the margins explain why: over 80% on its own and open-weight models, 20 to 30% on closed ones.
That gap turns routing into a core feature. Higgsfield picks the model itself in over 40% of cases, a lever called tokenomics, and NVIDIA cut agent costs by half without touching the model.
Alex reads OpenRouter data as showing open-source share rising from below 30% to over 60% this year, though he still expects OpenAI and Anthropic to hold over 50% of market dollars. That is 1 marketplace’s view, and I would not size a model bet on it alone.
Judge a model on your workflow, your references, and your prompt length, never on a leaderboard.
6. $4M a Month on AI Models: The Employee Who Spent $30K in a Week
Over $10,000 a month per person goes to models. Some weeks run far past that.
This month I just caught a guy who spent over 30k in a week on Astra. He was frustrated that asset organization and auto editing are still not very good in production, and he said, “I’m just going to do this myself.” He went 5 nights straight on Astra and it works. I wouldn’t say it was production ready, but we learned a lot.
The math holds: close to 400 people, over $4M a month, over $10,000 per head. Many people spend over $10,000 in a week, and Alex forecasts that top engineers and creatives are heading for $50K to $100K a month within 12 months.
His finance team would call him too stubborn about controlling spend. He still counts the $30K week as a net positive: a frustrated employee built something that works, even if it is not production ready.
A $30K week that teaches the team something beats a committee that never lets anyone try.
Higgsfield isn’t an outlier. Tokens got 600x cheaper in 6 years, yet Uber burned a year of AI budget in 4 months, and GitHub compressed its coding agent’s output and watched the total bill climb.
“30k in a week is quite a lot. But we learned this.”
Before you set your own model budget, read these:
▫️ Your AI bill is mostly wasted tokens
▫️ You are overpaying for intelligence
▫️ Inference engineering: the cost cut most teams miss
Give every model budget an owner, and reserve a slice for experiments you expect to fail.
7. Higgsfield Expected AI to Replace Support and Legal. It Runs 50+ People Across Both (Here’s Why)
Alex expected agents to replace support and legal. His org chart disagrees.
I don’t see any elimination. It’s true that probably over 60% of customer support requests, especially the first line of defense, can be handled with AI. But when it comes to B2B, AI just doesn’t work.
Alex calls not ramping these teams quickly a main operational mistake. Legal is over 10 people now and customer success is over 40.
The bottleneck is context. Higgsfield launches a new product pretty much every week, and an agent is only as good as the context and rules it holds.
Memory alone doesn’t fix that: the bigger an agent’s memory, the worse it gets. Deflecting a first-line ticket is easy. Owning the answer is not.
Outside Higgsfield the pattern holds. The AI job apocalypse piece attributes to the Atlanta Fed a finding that over 90% of US firms that adopted AI in the past 3 years reported zero change in headcount. The AI layoff trap explains why firms keep cutting anyway.
Before an agent touches your support queue, read these:
▫️ 3 founders shipped Claude Managed Agents to production in 2 weeks
▫️ The agent system around the model is the product
If your rules change 2 times a week, someone owns updating the agent 2 times a week.
8. The Only 2 AI Moats Alex Believes In: Outcomes and Network Effects
Everyone has a moat theory. Alex has 2.
We do believe there are only 2 ways of modern value creation, or moats, today. First is when you deliver the outcome, and for us it’s allowing businesses to sell more through AI ads. The second thing is network effects. Unfortunately, AI does not replace network effects.
His network-effect bet is open source. Higgsfield went from about 10 seeded open-source projects 8 weeks before the interview to over 10,000, the same fork-and-build snowball that runs GitHub. 4.7 million people downloaded OpenClaw which shows how fast that snowball can roll.
I find the network moat the harder one to build, because an outcome shows up in the first quarter and a network takes years.
Alex doubts swarms of AI agents trading with each other arrive in the next 5 years, which leaves human networks as the ones that count. The same logic drives his bet on the system of record, the place a company’s creative assets, search, and brand checks live. Whoever builds the AI-native one wins, and the SaaS defense playbook covers the defensive side.
That sits close to Marc Andreessen’s argument that the AI moat is not the model and to the case for where AI moats live now.
Name your moat. If it’s neither an outcome nor a network, it’s a feature.
9. Zero Paid Ads, 150 In-House Creatives (What “No Paid” Includes)
Higgsfield sells AI video, and it keeps 150 creatives on staff.
We have an in-house team of over 150 creative professionals. They make product launch videos, they make tutorials. For example, we made the first generated movie, and we open sourced all of it. What we learned is that for 90 minutes of TV quality content, it was over 100 hours of generated content. So creative decisioning, picking the right piece, is still very important.
Consumer growth comes from Higgsfield’s own content and distribution, with no paid ads, by Alex’s account. 150 creatives against close to 400 people is about 38% of the company.
Then the host pushes on “no paid,” and the edges get fuzzy. Influencer deals typically carry a production fee plus cost-per-click attribution, and Higgsfield also outsourced creator and customer-success work to an agency, a bad experience in Alex’s words. “No paid” is narrower than it sounds.
I would ask what share of growth he counts as organic before repeating the claim.
Alex also predicts most social content will be AI-generated, with authentic shows earning a 10 to 15x higher CPM. On distribution, The VC Corner argues that the only moat left is motion, with speed, story, and surface area deciding who wins in AI, and a GTMfund guest post calls distribution the final moat.
Own the content engine before you rent anyone’s audience. Start with the free AI content engine that beats a content team or the founder version.
10. Why Higgsfield Hires in Kazakhstan: 300+ People, Top-5 Physics, 15% Income Tax
Everyone asks about labor arbitrage. Alex answers with physics rankings.
Kazakhstan is top 5 in the world in physics. You look at the recent International Physics Olympiad for high schoolers, they’re top 5 in the world, on par with the United States, China, India. This is also the core of our team: people who won international competitions in math and physics.
About 50 people sit in California and over 300 in Kazakhstan. Arbitrage hires for price. Alex hires for a pipeline, and an olympiad ranking is a sturdier argument than a tax rate.
A ranking describes the top of a pipeline, not 300 hires, so test that claim first.
The pipeline comes from a Soviet math tradition upgraded with Singaporean education principles, plus a government that subsidizes thousands of high schoolers to study abroad, many of whom return. The 15% personal income tax comes after the talent argument, and the stated hope is that Higgsfield prints more dollar millionaires in Central Asia than any other company.
Most founders still run a hiring playbook written for a market that closed in 2022, and startup location strategy in 2025 asks whether it still matters where you build.
Before you hire for cost, find the country that already trains the skill you need.
The Higgsfield Playbook
Customers pick the product. Margins pick the model. Your own content picks the audience.
▫️ Founders: Count revenue Alex’s way: 4 weeks times 13, prorated, live only. Start the $99 plan, then book 8 customer calls before your next build sprint.
▫️ Investors: Ask for month-12 net revenue retention and the margin split between open and closed models, not just the ARR headline. Then ask for the assumptions behind the gap between $4.5B and over $10B, and read Stop Asking What Your Startup Is Worth before you price a round on this curve.
▫️ Operators: Give every agent a rulebook with a named owner and a change log, and budget model spend per head like salary. Start with the queue where a wrong answer costs the most.
▫️ Everyone else: Higgsfield has seen no elimination of support or legal roles so far, so the work shifts to context and judgment. Train there first.
The 5 Principles to Steal
Count revenue the strict way. 4 weeks times 13, prorated, live revenue only.
Talk to 8 customers first. Burn the $10M after the calls, not before.
Route models by margin. Over 80% on open weights, 20 to 30% on closed, and a model picked by you in over 40% of cases.
Budget model spend like payroll. Over $10,000 a month per head, with room for the $30K week.
Treat momentum as borrowed. Alex’s biggest lesson from Snap is that momentum doesn’t last forever: it was worth $80B when its gap with Meta was under 10x, and he says it sits below $15B today. Run the startup momentum diagnostic while yours is still good.
$1M to $1B in 18 months, and the turn came from 8 customer calls.
If this breakdown saved you an hours of watching, send it to one founder or investor who needs it.
The clips above are the best moments, and the full 20VC episode has the rest, including the Snap lessons and the public-company question.
Growth, moats, and positioning
▫️ Granola’s growth playbook: wedge, moat, pricing, context layer
▫️ The new SaaS playbook for the AI agent era
▫️ Clone your voice into Claude in a weekend
Agents and operating models
▫️ The one-person company operating system
▫️ Inside OpenAI: the operating model that makes 2 engineers beat 200
Cost, reliability, and shipping
▫️ The Claude Opus 5 playbook: benchmarks, pricing, effort dial