20VC: $1BN ARR in 18 Months; The Untold Story of Higgsfield | Spending $4M Per Month on Models | Why Moats in AI are BS | Scaling a Content Team to 150 People with Alex Mashrabov

Summary

Alex Mashrabov frames Higgsfield as one of the fastest-scaling AI companies, claiming it reached $1B in annualized revenue in 18 months using a live-revenue methodology rather than booked sales. The company’s early product-market fit came from repeated customer interviews that revealed a missing primitive in AI video: camera control. He says Higgsfield initially burned over $10M searching for a product, then pivoted to distribution, owned content, and product-led growth. Mashrabov is skeptical of common AI benchmarks, especially for video, arguing that real workflows require long prompts, many image references, and a rendering-engine mindset rather than simple text-to-video scoring. He also says moat creation comes from outcome delivery and network effects, with open-weights models often delivering much better margins than closed models. Operationally, he describes very high model spend, a large in-house creative team, and a business that is expanding rapidly in enterprise while still wrestling with churn, education, and the broader challenge of turning AI video into a durable category.

Notes

• Mashrabov says his family pushed him toward the U.S. because technology mattered more there than in Uzbekistan.
• He reached top three globally in competitive programming by age 19 before choosing startups over academia.
• He describes early work on pre-transformer neural nets and state-of-the-art English-Russian translation.
• His prior company AI Factory sold to Snap for $166M, and he later led GenAI there.
• He recalls a much smaller AI funding market before today’s high multiples and notes severe dilution.
• Higgsfield’s first product attempts were weak approximations: slideshow-from-images and vertical video clipping.
• The startup spent more than a year and over $10M of its $16M seed round searching for a viable product.
• Customer interviews pointed to camera control as the missing feature in AI video.
• Higgsfield deliberately avoided paid acquisition and instead bet on owned distribution.
• He says outsourcing distribution to an agency was a bad experience, reinforcing the need to own go-to-market.
• Mashrabov says Higgsfield crossed $1B in annualized revenue and defines ARR as live revenue over the last four weeks multiplied by 13.
• He claims Higgsfield hit $1B ARR in 18 months from $1M, faster than Cursor’s roughly 24 months.
• Video AI adoption is still early, roughly two years behind coding.
• A customer moved from $99/month to a $6M deal in six months, showing extreme expansion.
• He says DTC advertisers in Asia are generating hundreds or thousands of ads per week and using AI to A/B test creative.
• Business revenue is slightly over 50% of total revenue, with consumer use cases around 10% of that consumer side.
• Mobile revenue remains under 10%, underscoring that the company is not primarily a consumer app.
• Month-12 net revenue retention in the business segments is over 300%, despite weak month-zero logo retention.
• Higgsfield maintains an in-house creative team of over 150 people.
• Creating 90 minutes of TV-quality content required more than 100 hours of AI-generated material, so creative selection still matters.
• He argues video benchmarks are closer to corporate theater than real product quality measurement.
• Real video workflows are not simple text-to-video; they need long prompts and many image references.
• He says benchmark labelers cannot evaluate the full prompt complexity, and each scene may use at least 10 image references.
• He compares video models to rendering engines like Unreal Engine or Unity.
• An in-house image model helped scale revenue from about $20M to $100M by improving aesthetic photo shoots and product consistency.
• He says many companies saying they built their own models really mean they post-trained open weights on proprietary data.
• OpenRouter’s open-source model share rose from below 30% to over 60% in a year, which he uses to support the open-weights thesis.
• For social ad creative, he says cheaper and more steerable models are often enough; Higgsfield sees 80%+ margins on open-source runs.
• Open-weights and own models reportedly produce over 80% margins, while closed models are closer to 20% to 30%.
• Model selection itself is a product lever; Higgsfield chooses the model in more than 40% of cases.
• He says state-of-the-art models are expensive and each employee spends over $10K per month on models.
• Internal model usage exceeds $4M per month.
• The company has close to 400 people, making the per-person model bill meaningfully above $10K.
• One engineer reportedly spent $30K in a week on Astra to unblock a workflow.
• He says spend can feel out of control, but the expensive experiment proved net-positive.
• He expects top engineers and creatives to justify $50K-$100K monthly spend, while some legal and finance tasks settle much lower.
• He says AI is already handling much of first-line support, but B2B support is still hard.
• Higgsfield launches new products almost every week, which forces constant agent updates.
• He says agents degrade when context and rules change twice a week.
• The stack migrated from Claude to Codex, first in engineering and then in creative workflows.
• He expects future AI products to become more specialized rather than simply larger.
• AI-native systems of record, not generic tools, are the thing he thinks will win in workflows like patents and asset libraries.
• Higgsfield is investing in a harness that improves over time and learns visual style.
• He rejects classic AI moat talk and says value comes from outcomes plus network effects.
• Open-source distribution on GitHub jumped from about 10 seeded projects to over 10,000 in eight weeks.
• He says the West contributes well over 70% of revenue, with Seoul the top city and the U.S. the top country.
• The company has about 50 people in California, about 50 remote, and more than 300 in Kazakhstan.
• He argues Kazakhstan is unusually strong in physics and math, helped by Soviet schooling, Singaporean education ideas, and subsidized study abroad.
• He says the Kazakhstan team benefits from a 15% personal income tax.
• He works roughly 80–90 hours a week and tries to preserve a few hours with his wife and son.
• He says he only got one full uninterrupted day with his son in the last three months.
• He says LLMs are now being used for hacks, fraud, bot credit abuse, and auto-refunds against AI companies.
• He changed his mind on HubSpot: familiarity and system-of-record functionality still matter in B2B.
• He still sees value in HubSpot as a system of record for diagnosing data-flow mismatches.
• He thinks most social content will be AI-generated, but authentic content will command much higher CPMs.
• He predicts creative directors will soon talk to computers in real time to generate stories and video variants.
• He admires Frank Slootman’s no-bullshit operating style.
• He says Snap’s big lesson is that momentum does not last forever.
• He says many public companies still have not figured out their AI story.
• He argues Meta belongs in a separate category because of Instagram and WhatsApp.
• He says Meta AI is still early and that public storytelling and launch consistency matter.
• Management’s internal model projects 4.5 by next year, but he believes the company can exceed 10.
• He explicitly says his own view is “Over 10.”
• He says adoption in creative AI is driven by monetization, DTC ads, and aspirational cinematic content.
• He thinks Hollywood is moving toward AI as hybrid production and a new form of CGI.
• He says Higgsfield could become bigger than AppLovin and Shopify, with distribution as the key long-term advantage.