Higgsfield went from $1 million to $1 billion in annualized revenue in 18 months, faster than Cursor, and almost nobody outside the AI world has heard the story. In this 20VC interview, Harry Stebbings sits down with Higgsfield CEO Alex Mashrabov, a former top-three competitive programmer from Central Asia who sold his first company to Snap, to talk about the near-death pivot, how Higgsfield counts revenue, why the company spends over $4 million a month on models internally, the economics of open versus closed models, what a moat even means in AI, and what it costs him personally to run it.

TLDW

Alex Mashrabov grew up pushed toward competitive programming by parents who told him the United States was where technology mattered, sold AI Factory to Snap for $166 million, and led generative AI there before co-founding Higgsfield. The company burned over $10 million of a $16 million seed chasing hype before talking to eight creative directors, finding that camera control was missing from AI video, and launching into immediate product market fit with no paid marketing. Higgsfield now calculates ARR as the last four weeks of live, prorated revenue times 13, gets slightly over half its revenue from businesses, sees about 30% churn in month one but net revenue retention over 300% at month 12, and has one customer who went from a $99 subscription to a $6 million deal. Alex argues Google and OpenAI will crush $20 a month consumer subscriptions, that AI benchmarks are gamed and do not reflect real video workflows, that open weights models give Higgsfield 80%+ margins versus 20 to 30% on closed models, and that routing traffic between models (“tokenomics”) is a core feature. Internally, the roughly 400-person team spends over $4 million a month on models, more than $10,000 per person. He sees only two real moats (delivering outcomes and network effects), talks about raising from Yuri Milner, hiring in Kazakhstan, Europe’s strengths, working 80 to 90 hours a week, the lesson from Snap that momentum does not last, and a personal target of over $10 billion in revenue within 12 months.

Thoughts

The most useful part of the founding story is the admission about the pivot. Higgsfield spent more than a year and over $10 million of a $16 million seed optimizing for “what’s hype today, what’s the right narrative,” and Alex takes the blame for it directly. What saved the company was not a new narrative, it was eight conversations with creative directors who all said the same thing: AI video had no camera control, and you cannot tell a story without it. That is a very specific, very unglamorous insight, and it took Higgsfield from roughly $1 million to $20 million in ARR in about three months. The lesson for any founder with a shrinking runway is that the last attempt should come from customers, not from the timeline.

The revenue section is worth reading closely because it is so unusual for a company at this stage to be this specific. Higgsfield counts ARR as the last four weeks of revenue times 13, prorates annual contracts, and excludes future contract value. Month-one churn is around 30%, which Alex openly says is below the old B2B SaaS bar, but month-12 net revenue retention above 300% is something SaaS basically never saw. Put those together with the $99 subscriber who became a $6 million a year customer, and his claim that Google and OpenAI will demolish the $20 a month consumer market, and you get a clear strategy: consumer signups are a funnel, and the business is expansion into companies that make hundreds or thousands of ads a week.

The model economics in the middle of the interview are the most concrete numbers on AI app margins I have heard in a while. Open weights and in-house models run at over 80% gross margin for Higgsfield. Closed frontier models run at 20 to 30%. And because agentic ad workflows let Higgsfield choose the model in over 40% of cases, routing becomes a margin lever, which he calls “tokenomics.” This connects to his sharp critique of benchmarks: video benchmarks test text-to-video, while real production uses 3,000-word prompts and ten or more image references per scene, closer to driving a rendering engine than writing a sentence. If the real workload does not need PhD-level intelligence to make a viral ad, the cheapest model that does the job wins, and the company that owns the routing keeps the spread.

On moats, Alex is more careful than Harry, who thinks they are mostly nonsense. Alex names two that still hold: delivering an outcome (helping businesses sell more through AI ads) and network effects, which “AI does not replace.” The system-of-record point is the less obvious part. Assets are scattered across Dropbox, Google Drive, Miro and Frame.io, and marketers want to search them semantically and check them against brand guidelines, something the pixel-first tools from Adobe and Canva were not built for. A harness that learns a brand’s visual style over time, plus a community that went from about 10 seeded open source projects to over 10,000 in eight weeks, is a more believable defensibility story than any single model. It is also a reminder that “wrapper” is not an insult if the wrapper becomes the place where the work lives.

The last 15 minutes carry the most interesting tension. Alex’s lesson from Snap is that “the momentum doesn’t last forever,” and Snap is now under $15 billion in market cap, partly because it never told a convincing AI story. His finance team projects $4.5 billion in revenue in 12 months with deceleration built in, while he personally says over $10 billion. He also predicts that most social media content will be AI generated, but that authentic content like Harry’s will command 10 to 15x higher CPMs, and that the job of the future is a creative director talking to a computer and generating stories in real time, where taste matters most. The honest version of his outlook is that he knows the curve will flatten and is trying to capture as much as possible before it does, while working 80 to 90 hours a week and admitting he has had one full day with his son in three months. That is not a lifestyle anyone should copy, but it is a candid look at what hypergrowth actually costs.

Key Takeaways

• Higgsfield crossed $1 billion in annualized revenue 18 months after hitting $1 million. Cursor took 24 months, and Alex believes only OpenAI and Anthropic did it faster.
• Alex’s father is from Uzbekistan and both parents are mechanical engineering professors. From age eight they told him he had to get to the United States because that is where technology matters.
• His mother worked three jobs so he could compete in programming and attend training camps. By 19 he was top three in the world in competitive programming, and he was also top three in the world at checkers.
• In 2014 he worked on pre-transformer neural nets for English to Russian translation. Many teammates were hired by DeepMind and Meta, but he was drawn to a future of AI-generated video on phones.
• With co-founder Mahi de Silva he built AI Factory and sold it to Snap for $166 million, after heavy dilution in an era when AI multiples were near zero and a $12 million round was a big deal.
• At Snap, his team’s face filters ran on-device, which made them nearly free for Snapchat, drove most daily new users, and reached hundreds of millions of people.
• The original Higgsfield insight: most companies cannot keep up with the pace of social media content production, because trends change almost every day.
• By 2023 it was clear scaling laws worked, and Alex bet they would work for video too, just two to three years behind language models and coding.
• Higgsfield spent over a year searching for product market fit and burned more than $10 million of its $16 million seed. Alex blames himself for chasing hype and narrative instead of product.
• With under $5 million left, the team talked to eight creative directors. All of them said AI video lacked camera control. Higgsfield launched on March 31 of the previous year and hit immediate product market fit.
• Higgsfield does no paid advertising. It had problems after outsourcing influencer work to an agency, and the takeaway was to own distribution.
• Higgsfield partnered with a major streamer on an AI-generated replica stream. Alex expects digital replicas to become a normal way for creators to monetize given how much pressure they are under.
• ARR methodology: revenue over the last four weeks times 13, only live revenue, with annual contracts prorated to 28 days and no multi-year deal value included.
• Video AI is about two years behind coding in adoption, and its share of on-demand usage revenue is well under the 50%+ seen at leading coding companies.
• One customer went from a $99 a month subscription to a deal worth over $6 million a year in six months.
• Big demand drivers come from Asia: direct-to-consumer brands rebuilding go-to-market to produce hundreds or thousands of ads a week, and the $10 billion+ short-form drama industry, where most new shows are made end to end with AI.
• Business revenue is slightly over 50%. Pure consumer use is about 10%, roughly matching mobile’s share of revenue. The rest is aspiring creators and freelancers who are churny but tend to return within a year.
• Alex believes Google and OpenAI will demolish the $20 a month consumer subscription market, so Higgsfield focuses on upgrading users to spend over $1,000 a year.
• Consumer retention drops about 30% in month one and then stays flat. Business net revenue retention at month 12 is over 300%.
• Higgsfield has over 150 in-house creative professionals, nearly half of the company, making launch videos, tutorials and an open-sourced AI-generated movie.
• That movie needed over 100 hours of generated footage for 90 minutes of TV-quality content, so creative selection still matters a lot.
• Alex calls chasing benchmarks a mistake and says researchers at large labs game them by leaking test data into training and using LLM-as-a-judge tricks to hit bonus targets.
• Video benchmarks mostly test text-to-video, but real workflows use prompts averaging over 3,000 words and at least ten image references per scene. Video models are best thought of as modern rendering engines, like Unreal or Unity with different inputs.
• VFX and camera control took Higgsfield from about $1 million to $20 million ARR in three months. Its own image model for aesthetic photo shoots and product consistency took it from $20 million to $100 million.
• Most companies that say they build their own models post-train open weights models. The most valuable post-training uses customer decision sequences to teach models to compress ten steps into one.
• Alex cites OpenRouter data showing the open source share of usage rose from under 30% to over 60% in a year, but expects OpenAI and Anthropic to keep over 50% of the market in dollars, driven by coding.
• Gross margin on own and open weights models is over 80%. On closed models it is 20 to 30%. Higgsfield chooses the model in over 40% of cases, which it calls tokenomics.
• Internal model spend is over $4 million a month across close to 400 people, more than $10,000 per person. One creative spent over $30,000 in a week vibe coding an asset workflow tool.
• Alex expects top “10x” engineers and creatives to spend $50,000 to $100,000 a month on models, and to ask for matching raises.
• Legal and customer support have not been replaced. Higgsfield has over 10 people in legal and over 40 in customer success, all heavy AI users. AI handles over 60% of first-line consumer support but does not work well for B2B.
• High product velocity makes AI support harder: agents are only as good as their context and rules, and those change twice a week.
• Engineers moved to Claude between March and June, then largely to Codex from mid June. Alex thinks tool preference is cyclical and model release velocity will not slow down.
• Whoever builds the AI-native system of record wins. For Higgsfield that means semantic search of assets, brand guideline enforcement, and a harness that learns visual style over time.
• Alex sees two real moats: delivering outcomes and network effects. Higgsfield’s open source community projects grew from about 10 to over 10,000 in eight weeks.
• Yuri Milner was his best VC meeting. Alex has also had investors shake hands on a price and then try to syndicate the round at a 30% lower valuation the next day.
• Excluding pharma and big tech, public companies spend more on sales and marketing than on R&D, and Alex expects much of that to become personalized video.
• The West is over 70% of revenue, the US is the largest country, and Seoul is the largest city by usage. Higgsfield does not operate in China.
• About 50 employees are in California, around 50 are remote, and over 300 are in Kazakhstan, which he says is top five in the world in physics olympiads and blends Soviet math with Singaporean education.
• His management philosophy: hire the best people, empower them, retain them. He learns from Jensen Huang, Elon Musk and Nik Storonsky, who reject much of standard corporate management.
• He works 80 to 90 hours a week, aims for at least 3 hours with his wife and 5 with his son, owns no property, drives a Tesla Model 3, and spent his first exit money buying apartments for family.
• He changed his mind on HubSpot: familiar interfaces matter to go-to-market hires, so he no longer thinks everyone will build their own CRM.
• The finance model projects $4.5 billion in revenue in 12 months. Alex personally believes over $10 billion is possible and is pushing for at least 30% month-over-month growth.
• Hollywood sentiment has shifted from strictly negative to neutral or slightly negative, with AI increasingly used as a new form of CGI in hybrid production.

Detailed Summary

From Competitive Programming in Central Asia to Snap

Alex describes a childhood built around competition. In Uzbekistan, a family of five earning $1,000 a month is considered wealthy, and his parents, both engineering professors, saw international rankings as the only way out. His mother worked three jobs and his father traveled with him to camps and competitions. By 19 he ranked top three in the world in competitive programming, but instead of academia he went into startups. After working on pre-transformer translation models, he became convinced that phones would become the dominant device and that AI would produce much of the video people watch on them. That led to AI Factory, co-founded with Mahi de Silva, which sold to Snap for $166 million. The dilution was heavy because AI companies were valued at close to nothing back then, but the deal got him to the US. He found San Francisco genuinely meritocratic, though its investors are more consensus-driven than he expected.

The Near-Death Pivot and Product Market Fit

At Snap, his team’s on-device face filters drove much of Snapchat’s new user growth. Afterward he focused on a different gap: most companies cannot produce social media content fast enough to stay relevant. Early tools like photo slideshows and long-to-short video clipping were not good enough. Believing scaling laws would come to video, he bet on AI video, but Higgsfield wandered for over a year and burned through most of its seed round. With under $5 million left, the team went back to basics, interviewed eight creative directors, heard that camera control was the missing piece, and launched. Product market fit was immediate, and Higgsfield still does no paid advertising. Asked about an influencer controversy, Alex says outsourcing creator work to an agency was a mistake and that owning distribution matters more than ever.

How Higgsfield Counts $1 Billion in Revenue

The interview was recorded the day Bloomberg reported Higgsfield crossing $1 billion in annualized revenue. Alex explains the method: the last four weeks of revenue times 13, which he says matches how leading AI labs report. Annual contracts are prorated so only one 28-day slice counts, and only live revenue is included. Business revenue is slightly over half, and pure consumer use (roughly the mobile share) is under 10%. A large middle group of aspiring creators and freelancers churns but often returns, which is why Higgsfield invests in education like Higgsfield Academy. The standout metric is expansion: one customer went from $99 a month to a $6 million a year contract, largely driven by e-commerce brands producing ads at massive scale and by AI-made short-form dramas.

Why $20 a Month Subscriptions Are Doomed

Alex’s contrarian view is that Google and OpenAI, with strong horizontal products, will destroy the consumer $20 a month subscription market for vertical apps. Harry offers Canva as an example of a company whose low-end consumer design use has been eaten by OpenAI. That is why Higgsfield judges success by whether a $20 subscriber can be shown enough value to spend over $1,000 a year. Month-one consumer retention drops about 30% before flattening, which Alex concedes is below the 80% logo-retention bar from B2B SaaS, but business net revenue retention at month 12 is over 300%.

Content as Distribution and the 150-Person Creative Team

Competitors told Harry that Higgsfield ran the most impressive influencer campaign in tech. Alex frames it differently: the goal is for the best commercial video content to be made on Higgsfield, with every workflow shown publicly. Over 150 in-house creative professionals, nearly half the workforce, produce launch videos, tutorials and even a fully AI-generated, open-sourced movie. The movie showed how much curation still matters: 90 minutes of TV-quality output required over 100 hours of generated footage.

Benchmarks, Own Models and Open Weights

Alex calls his early focus on benchmarks a mistake and describes how large labs game them. By OpenRouter usage, he argues, Google is the only relevant US incumbent in models (later adding Nvidia), while in China Tencent, Xiaomi and Alibaba are all relevant, with ByteDance catching up. For video specifically, benchmarks miss the real work: long prompts, many reference images, and precise control of characters and backgrounds. Higgsfield still builds its own models where customers need them, like its image model for product photo shoots that helped take it from $20 million to $100 million ARR. He says most companies that claim to build models are post-training open weights, and that the most valuable post-training teaches models to collapse multi-step customer workflows into one.

Tokenomics and the $4 Million Monthly Model Bill

Most social media marketing does not need frontier intelligence, so customers want cheaper, stable models. Open weights and own models give Higgsfield over 80% margins, while closed models give 20 to 30%, and Higgsfield picks the model in over 40% of workflows. Internally, the team of nearly 400 spends over $4 million a month on models. The creative team started vibe coding tools that do not exist in production, including one person who spent $30,000 in a week building an asset organization workflow over five straight nights. Alex admits spend sometimes goes out of control but calls that case net positive. He expects top engineers and creatives to reach $50,000 to $100,000 a month, while support functions like legal and finance will stabilize at much lower levels.

Where AI Has Not Replaced Jobs

Alex expected legal and customer support to be largely replaced and says not staffing them quickly enough was an operational mistake. Today Higgsfield has over 10 people in legal and over 40 in customer success. AI can handle over 60% of first-line support, but not complex B2B requests. Harry notes Revolut’s 92% AI resolution rate for consumers. Alex points out that shipping new products every week makes support agents harder to keep accurate, because their context and rules keep changing. His engineers moved to Claude from March to June and then mostly to Codex, and he expects model releases to keep coming fast, including specialized models like OpenAI’s legal push, which Harry is skeptical of.

System of Record and the Two Moats

Using Solve Intelligence as an example, Alex argues whoever builds the AI-native system of record wins. Creative assets are spread across many tools, and marketers want to search them in natural language and check them against brand identity and guidelines. Adobe and Canva built the best software for a pixel-first era, but that is not where the world is heading. Harry says moats are mostly nonsense and cites Lovable and Instinct as wrappers that won on speed. Alex responds that there are only two real moats now: delivering outcomes and network effects. Higgsfield’s open source community projects grew from about 10 to over 10,000 in eight weeks, which he hopes becomes a compounding moat much like forking on GitHub did for software.

Fundraising, Valuation and the Size of the Market

Alex names Yuri Milner as his best VC meeting because Milner understood that content trends like AI-native ads and short dramas flow from Asia to the West. He has scars from investors who agreed on a price and then shopped the deal at a 30% discount. Harry argues Higgsfield is discounted for not being a Silicon Valley insider and would easily be worth $25 billion otherwise. Alex says Higgsfield is building for the long term, with aspirations to be the distribution infrastructure for direct-to-consumer businesses the way Shopify became their commerce infrastructure. The West drives over 70% of revenue, and the biggest lesson from Asia is how hard companies there push for direct relationships with customers.

Family, Sacrifice and Work Ethic

The conversation turns personal when Harry asks if Alex will regret the time away from his son. Alex talks about his mother’s three jobs, his father’s devotion to his education, and his father’s Parkinson’s diagnosis, which no amount of money can fix. He works 80 to 90 hours a week, tries to spend at least 3 hours a week with his wife and 5 with his son, and has had one full disconnected day with his son in three months. He says there is no shortcut to hard work, citing the product leaders he worked with at Snap, and credits his wife’s patience and a cultural emphasis on mutual sacrifice.

Hiring, Kazakhstan and Europe

Over 300 of Higgsfield’s employees are in Kazakhstan. Alex rejects the idea that it is mostly labor arbitrage: Kazakhstan ranks top five in physics olympiads, combines Soviet math with Singaporean teaching methods, sends thousands of students abroad who often return, and offers a 15% personal income tax. He hopes Higgsfield creates more dollar millionaires in Central Asia than any other company. He also pushes back on Harry about Europe, pointing to neoclouds like Nscale and Nebius, application companies like Legora, ElevenLabs and Lovable, and ASML as proof that Europe competes at every layer. He values loyalty and contrasts it with Silicon Valley’s two-year job hopping. On management, he learns from Jensen Huang, Elon Musk and Nik Storonsky, and boils it down to hiring the best, empowering them and retaining them.

Quick Fire: HubSpot, AI Content, Snap and $10 Billion

Alex changed his mind on HubSpot, because experienced go-to-market hires value a familiar system of record. His contrarian belief is that most social media content will be AI generated, while authentic content will earn 10 to 15x higher CPMs. The job that does not exist yet is a creative director who talks to a computer and generates stories in real time, where taste is the key skill. He would most like Frank Slootman on his board after reading Amp It Up, though Harry warns that Snowflake’s go-to-market focus lost ground to Databricks’ product focus. His lesson from Snap is that momentum does not last, and he thinks Snap fell behind Meta because it never told a convincing AI story, while Mark Zuckerberg did. Higgsfield’s finance model projects $4.5 billion in revenue in 12 months, but Alex personally says over $10 billion, pointing to monetization-driven adoption and a Hollywood that is slowly warming to AI as a new kind of CGI. He wants to take Higgsfield public and believes it can be bigger than AppLovin and Shopify because distribution is what matters.

Notable Quotes

“We burned more than 10 million out of 16 million raised in seed fundraising. So we felt we have just one attempt left.”

Alex Mashrabov, on the year Higgsfield spent searching for product market fit

“I was so much optimizing for what’s hype today, what’s the right narrative, how we can hijack the attention, all these things really like everything instead of building a good product.”

Alex Mashrabov, taking responsibility for the failed first year

“One customer started 6 months ago spending just subscription $99 a month. $99 a month. And now we just signed a deal over 6 million.”

Alex Mashrabov, on Higgsfield’s revenue expansion

“The way to think about video models today, it’s just modern rendering engine. It think about this as like Unreal Engine or Unity but just different types of inputs.”

Alex Mashrabov, on why text-to-video benchmarks miss real workflows

“The margin on own models and open weights models is over 80%.”

Alex Mashrabov, comparing it with 20 to 30% on closed source models

“The agents are as good as context and rules which they have. And if context and rules change pretty much twice a week, it gets a little difficult.”

Alex Mashrabov, on why fast product velocity makes AI customer support harder

“Unfortunately, AI does not replace network effects.”

Alex Mashrabov, on the two moats he still believes in

“So yeah, I don’t believe that there is any shortcut to hard work.”

Alex Mashrabov, on working 80 to 90 hours a week

“People just still don’t fully appreciate that most of the content on social media is going to be AI generated.”

Alex Mashrabov, on what he believes that others think is crazy

“The momentum doesn’t last forever.”

Alex Mashrabov, on his biggest lesson from Snap

Watch the full 20VC conversation with Alex Mashrabov here.

Related Reading

• Higgsfield AI the AI video platform at the center of the conversation.
• 20VC with Harry Stebbings the venture podcast where the interview first aired.
• OpenRouter model rankings the usage data Alex cites on the rise of open source model share.
• Neural scaling law (Wikipedia) background on the scaling laws Alex bet would carry over from language models to video.
• Network effect (Wikipedia) the moat Alex says AI cannot replace.