Summary

Higgsfield crossed $1B in annualized revenue 18 months after hitting $1M — faster than Cursor’s 24 months — and Alex Mashrabov claims it’s “probably the third after OpenAI and Anthropic.” The methodology is trailing-four-weeks revenue ×13, prorating annual contracts and counting “only live revenue,” not multi-year enterprise bookings; he says OpenAI, Anthropic, and CloudBell use the same approach. Business revenue is slightly over 50%, mobile is under 10%, and the West is well over 70% despite the Asia-driven narrative, with Seoul the largest city by usage.
The expansion math is the tradeable headline: NRR at month 12 is over 300%, which “just never happens in B2B SaaS.” Month-one churn runs around 30% then flattens, but Mashrabov’s favorite specimen is a customer who started at $99/month six months ago and signed a deal “over $6 million a year” — driven by AI-native e-commerce advertisers printing hundreds or thousands of ads weekly and a $10B+ Chinese-dominated short-form-drama industry making shows “with AI end to end.”
Mashrabov calls model benchmarks “corporate psyops”: labs put test data into training and game scores for quarterly bonuses, while video benchmarks measure text-to-video even though real production prompts average over 3,000 words with at least 10 image references per scene. His evidence: on OpenRouter data, Google is the only relevant US incumbent while China has Tencent, Xiaomi, and Alibaba; open-source model share rose from below 30% to over 60% this year.
The margin structure is the model-layer thesis: margins exceed 80% on own and open-weights models versus 20–30% on closed-source models — and Higgsfield chooses the model in over 40% of cases, a routing layer it calls “tokenomics.” Building proprietary models from ambition was “my mistake”; they now train only where customers demand it, because “PhD-level intelligence is not necessarily needed to make a viral social media video.”
Internal model spend exceeds $4M/month across close to 400 people — over $10K per head — including a creative who spent “over $30K in a week on Astra” vibe-coding a workflow the product lacked. Mashrabov expects 10X engineers and creatives to reach $50K–$100K/month with salaries correlating upward, but pushes back on team shrinkage: legal is over 10 people and customer success over 40, because agents “are as good as the context and roles which they have” and Higgsfield ships products weekly.
Moats reduce to two things — delivering outcomes and network effects — and Mashrabov is skeptical of agent-swarm hype: “I’m not sure this is happening in the next five years.” Higgsfield’s community open-source projects scaled from about 10 seeded projects to over 10,000 in eight weeks. His contrarian bet is that Google and OpenAI “are going to completely demolish their prosumer subscription markets” — the $20/month tier Canva-type products serve.
The finance model projects “4.5” by the end of next year with “substantial deceleration” baked in; Mashrabov’s own answer is “over $10 million,” while the transcript does not specify the unit of the 4.5 figure. He cites Hollywood sentiment shifting from negative toward neutral and a target of at least 30% month-over-month growth. He argues the Silicon Valley discount is real — Harry says a Valley company at these numbers would be worth $25B+ and invokes Cognition’s $50B valuation — and cites public companies outside pharma and big tech spending more on sales and marketing than R&D.
The company is a structural anomaly: over 300 of close to 400 people are in Kazakhstan, there is no paid advertising, and a 150-plus-person in-house creative team drives most revenue through owned content. Mashrabov rejects the labor-arbitrage read — Kazakhstan is top five in the world in physics Olympiads, and its education system upgraded the Soviet math school with Singaporean principles — and he is the counter-signal on founder intensity: 80–90 hour weeks, three hours a week with his wife, no owned property, and a leased Tesla Model 3.

Deep dive

1. From Uzbekistan to top-three in the world — and a $166M exit amid severe dilution

Mashrabov’s origin sets the operating system: his father is from Uzbekistan, “where if a family of five people makes $1,000 a month, it’s considered to be wealthy”; both parents were mechanical-engineering professors who told him from age eight that he “must get to the United States ‘cause this is the place where technology matters.” His mother worked three jobs while his education centered on programming competitions and camps; by 19 he was top three in the world in competitive programming.
The pre-Higgsfield arc: pre-Transformer neural nets in 2014, including a state-of-the-art English–Russian translation system whose teammates were hired by DeepMind and Meta; then AI Factory, which he joined after meeting Mahi in 2018 and which was sold to Snap for $166M. The deal came with “severe dilution” because AI multiples then were “closer to zero,” not 200x revenue. At Snap he led GenAI; the face filters his team built drove most daily new users, ran on-device “virtually for free,” and scaled to hundreds of millions — “still probably the most-used consumer media AI product.”
His verdict on Silicon Valley after arriving: genuine meritocracy in access, but “what I see among Silicon Valley investors is that they are extremely consensus-driven.”

2. Higgsfield nearly died — camera control was the unlock

The founding insight was that “no one can keep up with the pace of production for social media, as trends change pretty much every day,” and by 2023 “it was absolutely clear that scaling laws finally work” in video — just “two or three years longer than LLMs and coding.”
The near-death is told with unusual candor: more than a year searching for a product, over $10M of a $16M seed burned, and “we felt we had just one attempt left.” His self-diagnosis: “I just lost touch with reality… I was optimizing for what’s hype today, what the right narrative was, how we could hijack attention… instead of building a good product.”
With slightly less than $5M left, they interviewed eight creative directors; everyone said the same thing: “camera control does not exist in AI, and camera control is so important to tell a story.” The product launched on March 31 of the prior year, and to Harry’s question of whether PMF is like love — when you know, you know — Mashrabov answered, “yes, it’s definitely when you know, you know.”

3. $1B annualized in 18 months — and exactly how the number is built

Announced the day of recording via Bloomberg: $1B annualized, 18 months from $1M versus Cursor’s 24. Mashrabov says the calculation is trailing four weeks ×13, with annual contracts prorated to the 28-day slice and “only live revenue” counted; three-year enterprise deals are not baked into the figure. He says OpenAI, Anthropic, and CloudBell use the same methodology.
Mix: business revenue is slightly over 50%; Mashrabov says around 10% is pure-consumer use cases and separately that mobile’s share of revenue is under 10%. The remainder includes aspiring freelance creators whose behavior “is a little bit of a journey”: most come back within a year, and Higgsfield educates them through Higgsfield Academy and its YouTube channel to become “this new AI-native workforce.”
The expansion story is the punchline: about a 30% month-one drop followed by flat retention, which he concedes is not B2B-SaaS-grade, but business-segment NRR at month 12 is over 300% — “it just never happens in B2B SaaS.” The best specimen: $99/month to a deal worth over $6M a year in six months, powered by Asia-originated trends — DTC e-commerce rebuilding go-to-market around hundreds or thousands of weekly ads, and short-form dramas, a $10B-plus industry owned primarily by Chinese companies where most new shows are made with AI end to end.

4. No paid advertising, 150-plus in-house creatives — and the death of the $20 subscription

Competitors told Harry this was “the most impressive influencer campaign in tech,” but Mashrabov says “we don’t do any paid.” Consumer growth runs on an in-house team of over 150 creative professionals — almost half the workforce — producing launch videos and tutorials, including what he calls the first AI-generated movie, which was fully open-sourced. The revealing ratio: 90 minutes of TV-quality output required over 100 hours of generated content — “creative decision-making, picking the right piece is still very important.”
On the influencer controversy, a partial mea culpa: Higgsfield had only two people on the creator and customer-success sides and outsourced distribution to an agency. “That was not a good experience,” and the takeaway he generalizes is “it’s very important to own distribution. Distribution is now more important than ever.”
His stated contrarian bet: Google and OpenAI’s focus on ads means “they are going to completely demolish their prosumer subscription markets” — the $20/month tier. Canva’s low-hanging consumer-design use cases are his “most apparent example” of the effect already emerging. Higgsfield’s answer is to get users past $20 and “make them upgrade and spend over $1,000 a year with us.”

5. Benchmarks are “corporate psyops” — the own-model mistake and tokenomics

The walked-back architecture decision, owned in full: “this was my mistake… at some point I really was thinking that chasing benchmarks is valuable, but I do believe this is just a sort of corporate psyop.” His account from inside large labs: researchers put test data into training, use LLMs as judges, and apply other tricks to game scores for quarterly bonuses.
His evidence for benchmark irrelevance: on OpenRouter data, Google is the only relevant US incumbent, while China — where benchmark obsession is “probably less” — has Tencent, Xiaomi, and Alibaba all relevant, with ByteDance trying to catch up. Video benchmarks test text-to-video when a video model is really “a modern rendering engine… like Unreal Engine or Unity”: Higgsfield’s open-sourced movie used prompts averaging over 3,000 words and at least 10 image references per scene.
The economics: margins on its own models and open-weights models are over 80%, versus probably 20–30% for closed-source models. Most businesses “don’t necessarily need Astra specifically” or the newest Fable model, which is why OpenRouter reports open-source share rising from below 30% to over 60% this year. Higgsfield chooses the model in over 40% of cases — “we call it tokenomics.”
His hedge on the dollar split is that, “just because capitalism works,” OpenAI and Anthropic will still have over 50% of market dollars, especially in coding. Higgsfield still builds models when customers demand a specific use case: VFX and camera control took ARR from roughly $1M to $20M in three months, while its own image model for aesthetic photoshoots and product consistency helped scale from $20M to $100M ARR. The strategy is customer-driven, “not just by an ambition to conquer the world and build the best model in the world.”

6. $4M a month on internal model spend — and why headcount didn’t shrink

The number Harry calls “the best question of the whole show”: internal model usage is over $4M/month across close to 400 people, more than $10K per head. The emblematic story: a creative frustrated with asset-organization and auto-editing workflows “went five nights straight on Astra” and spent over $30K in a week vibe-coding. It was not production-ready, “but we learned a lot. This was actually a net-positive experience.”
His forecast: 10X engineers and 10X creatives could reach $50K–$100K/month in model spend — “unfortunately, I also expect that these people will ask for comparable salary raises” — while many other functions, such as legal and finance, stabilize around $500K/month very quickly.
The change of mind on labor: he expected legal and customer support to be “mostly replaced” and calls failing to ramp those teams quickly “one of the main operational mistakes.” Today legal is over 10 people and customer success over 40, all AI-heavy; over 60% of first-line support requests can be AI-handled, “but when it especially comes to B2B, AI just doesn’t work.” Harry’s counter — Revolut’s claimed 92% consumer resolution rate — draws a concession plus the mechanism: Higgsfield ships new products weekly, and “agents are only as good as the context and roles they have.”
On tooling: from March to June, everyone moved to Claude, including the creative team; coders then moved from Claude to Codex around mid-June, with 10X creatives following. Mashrabov says the pattern is cyclical.

7. Moats, the Silicon Valley discount, and the forward-growth dispute

Echoing Harry’s “moats are largely bullshit” framing, Mashrabov allows exactly two: delivering outcomes — businesses selling more through AI ads — and network effects. He hedges on agent hype: “when people talk about a swarm of AI agents talking to each other, I’m not sure this is happening in the next five years.” Higgsfield’s community open-source projects went from about 10 seeded projects to over 10,000 in eight weeks.
His system-of-record vision is a semantic asset library that can search content, enforce visual brand guidelines, and learn visual style over time. He says Higgsfield has invested heavily in Harness for that purpose, while Claude and OpenAI cannot necessarily learn a company’s visual style in the same way. This is his alternative to pixel-first Adobe and Canva, not merely another model.
Fundraising war stories: Yuri Milner “gets it” and deeply understood the transformation of content, but “sophisticated investors can play games”: investors have shaken hands on a price and then called other investors to pull syndicates in at a 30% lower valuation.
On Harry’s charge that Higgsfield trades at a non-insider discount, Harry says a Silicon Valley company growing this quickly at $1B in revenue “would easily be a $25 billion company” and invokes Cognition’s $50B valuation; Mashrabov agrees there is upside. He then pivots to market size: excluding pharma and big tech, public companies spend more on sales and marketing than R&D — a figure he says four team members verified — and much of that work could become personalized video.
The finance model projects “4.5” by the end of next year with “substantial deceleration”; the transcript does not specify the unit. Mashrabov’s own answer is “over $10 million.” He ties the upside to Hollywood sentiment moving from strictly negative toward neutral and to private conversations in which more people ask whether AI can help sell new stories or overcome budget limits.
Long term, Alex says “Hexcell” has the potential to be bigger than “Amplivine” and Shopify. He describes the company as infrastructure for DTC distribution: Shopify is one infrastructure layer, while Higgsfield aims to become another.

8. The Kazakhstan anomaly, loyalty over arbitrage, and no shortcut to hard work

On the over-300-person Kazakhstan base, he rejects the arbitrage frame: Kazakhstan is top five in the world in the recent International Physics Olympiad, “on par with the United States, China, and India”; its system upgraded the Soviet school of math with Singaporean principles; the government subsidizes thousands of students abroad who return; and personal income tax is 15%. “I just hope we’re going to bring more dollar millionaires in Kazakhstan, in Central Asia, and in this part of the world than any other company.”
His Europe-is-competitive riff names Anscale, Airan, Nereus, and Crusoe among relevant neoclouds, and Legora, ElevenLabs, and Lovable on the application layer, with ASML as essential infrastructure. His loyalty metaphor: “if you’re a Fulham fan, you’re not going to root for Arsenal” just because Arsenal is winning, whereas in the United States people may switch allegiance to whichever team is on top.
Management philosophy, stripped down: Jensen, Elon, and Nick “completely abandon all the management principles”; the only rule is “hire the best people, empower them to do the best work, and figure out how to retain them. Everything else is frankly secondary.”
The personal cost is stated without varnish: 80–90-hour weeks, “at least three hours a week with my wife, at least five hours a week with my son,” and one full unplugged day with his son in three months. The deeper layer: his father dedicated his life to Alex’s competitions and has had Parkinson’s since Alex was 21 — “even having access to capital and exits cannot fully change things.”
He owns no property; after the company sale, he spent over $1M buying apartments for his parents, relatives, and his wife’s parents. He drives a leased Tesla Model 3.
Quickfire signal: he reversed on HubSpot — “I was thinking that HubSpot was going to become obsolete,” but a familiar interface matters when hiring GTM talent; Harry still says it will “get fucked.” His contrarian belief is that most social content will become AI-generated while authentic shows command 10–50x higher CPM. The job with no name yet is a creative director “talking to computers and generating stories in real time,” where taste becomes more important than the existing chain of scriptwriting, storyboarding, and directing roles.
His preferred board addition is Frank Slootman; Harry’s pushback is that Slootman’s Snowflake GTM machine was offset by Databricks prioritizing product, with Harry recommending Chad Peets instead. The Snap lesson closes it: with market cap below $15B, “momentum doesn’t last forever,” so positive momentum cannot be taken for granted.

Verification Notes

The transcript states the finance projection only as “4.5” and Alex’s personal answer as “over $10 million”; no larger unit is supplied for either claim, so the digest does not normalize them to billions.

Jensen Pushes Back on Doomers, Xi & Trump Talk AI, and “AI” Gets a Rebrand | #294 MOONSHOTS Live