The Twenty Minute VC · 28 Sep 2026 · From the week of 28 September
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
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Harry Stebbings interviews Alex Mashrabov, founder and CEO of AI video company Higgsfield, which he says went from $1M to $1BN in annualized revenue in 18 months with roughly 400 people, most of them in Kazakhstan. Mashrabov describes burning more than $10M of a $16M seed on the wrong things before refocusing on product and PLG around camera control. He also covers how Higgsfield calculates run-rate revenue, its retention and expansion numbers, its no-paid-ads growth driven by a 150-person in-house creative team, its $4M-a-month internal model spend, and the margin difference between open-weight and closed models. His central argument is that video AI will transform the advertising and sales-and-marketing industry. He believes defensibility comes only from delivering outcomes and from network effects, and that horizontal players like OpenAI and Google will wipe out $20-a-month prosumer subscriptions.
For founders
- Mashrabov says Higgsfield burned more than $10M of a $16M seed chasing hype and narrative, then found immediate product-market fit after interviewing eight creative directors who all named the same gap: camera control.
- Higgsfield calculates annualized revenue as the last four weeks of live revenue times 13, pro-rating annual contracts and excluding multi-year deals, and Mashrabov says OpenAI and Anthropic use the same method.
- Mashrabov reports roughly 80%+ gross margin on post-trained open-weight models versus about 20-30% on closed-source models, which makes model routing ('tokenomics') a core feature; Higgsfield picks the model in over 40% of cases.
- He believes the only moats today are delivering an outcome and network effects, and points to community open-source projects growing from 10 to over 10,000 in eight weeks; Harry Stebbings argues moats are largely BS and that speed of execution is what matters.
- Internal model spend is over $4M a month, more than $10K per employee, and Mashrabov expects top '10x' engineers and creatives to reach $50K-100K a month each.
For revenue leaders
- Higgsfield runs no paid ads and grows through owned content from an in-house team of over 150 creative professionals; an influencer program outsourced to an agency was 'not a good experience', and Mashrabov's takeaway is that you must own distribution.
- Mashrabov says one customer went from a $99/month subscription to a deal over $6M within six months, and business-segment NRR at month 12 is over 300%, despite a roughly 30% drop in the first month.
- Mashrabov expects OpenAI and Google to demolish the $20/month prosumer subscription market, so Higgsfield's priority is upgrading $20 users to more than $1,000 a year.
- He estimates AI can handle over 60% of first-line customer support, but says it doesn't work for B2B, and weekly product launches make support agents hard to keep current; he calls not ramping legal and CS teams quickly one of his main operational mistakes.
- Mashrabov changed his mind on HubSpot: he no longer thinks companies will build their own CRMs, because B2B go-to-market hires value a familiar system of record.
What was said 34, most useful first
Higgsfield calculates annualized revenue as the last four weeks of live revenue multiplied by 13, pro-rating annual contracts and excluding multi-year deals. Listen
Mashrabov says he believes OpenAI, Anthropic and others use the same four-weeks-times-13 method. Higgsfield counts revenue, not sales: annual subscriptions and enterprise contracts are pro-rated so only the 28-day share is counted. Only live revenue is included, and multi-year enterprise deals are not baked into the $1BN figure.
“What we do is we look revenue over the last four weeks and multiply it by 13.”
Higgsfield's business-segment NRR at month 12 is over 300%, while about 30% of users drop in the first month before retention flattens. Listen
Mashrabov acknowledges a large first-month drop of maybe 30%, which he attributes to people not fully realizing the value. He contrasts this with the B2B SaaS expectation of month-one logo retention above 80% and says user education is a core priority. He caveats that the company is young, but calls the expansion unprecedented.
“NRR at month 12 is over 300%. It just never happens in B2B SaaS, right?”
Listen to the episode Retention & customer success Link to this
Mashrabov expects OpenAI and Google to demolish the $20/month prosumer subscription market, so Higgsfield focuses on upgrading users to over $1,000 a year. Listen
He calls this a contrarian bet and says horizontal products will absorb many verticalized $20-30/month tools; he agrees that much low-end consumer design work Canva served can now be done in OpenAI. He says Higgsfield will never win at $20 a month. Its priority is showing value fast enough to move a $20 subscriber to spending more than $1,000 a year.
“How can we make them to upgrade to over, to spend over $1 ,000 a year with us?”
Higgsfield earns over 80% margin on its own and post-trained open-weight models versus roughly 20-30% on closed-source models, which makes model routing a core feature. Listen
Mashrabov says social media ad production doesn't need 'PhD level intelligence' and favors cheaper, more steerable models, which can also be more cost-efficient for customers. As agentic ad workflows grow, Higgsfield chooses the model in over 40% of cases. He calls this optimization of token count and token cost 'tokenomics'.
“The margin on own models and open weights models is over 80%. And then it almost doesn't matter. And for close source models, it's probably between 20 and 30%.”
Higgsfield spends over $4M a month on internal model usage, more than $10K per employee across roughly 400 people. Listen
Mashrabov says creative staff began vibe coding. One person spent over $30K in a week building an asset-organization workflow that wasn't production-ready but taught the team a lot, and many spend over $10K a week. He acknowledges his finance team thinks he is too stubborn about not controlling the spend, but calls the experience net positive. He expects top '10x' engineers and creatives to reach $50K-100K a month and to ask for comparable salary raises, while spend for functions like legal and finance stabilizes quickly.
“I do believe we are gonna get to spend close to 50k and 100k a month for those who can call 10x engineers, 10x creatives.”
Mashrabov estimates AI can handle over 60% of first-line customer support but says it does not work for B2B, and high product velocity makes support agents harder to maintain. Listen
He says expecting legal and customer support to be mostly replaced, and so not ramping those teams quickly, was one of Higgsfield's main operational mistakes. Higgsfield now has over 10 people in legal and over 40 in customer success, all heavy AI users, and he sees no elimination in those roles. Because the company launches products roughly weekly, agents' context and rules change about twice a week, which makes a smart, coordinated human team important.
“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 especially comes to B2B, like, AI just doesn't work.”
Listen to the episode Retention & customer success Link to this
Public companies outside pharma and big tech spend more on sales and marketing than on R&D, and he sizes Higgsfield's market against that spend. Listen
He says at least four people on his team verified the figure. He argues that sales and marketing aim to deliver the personalized offer that converts best, and much of that will become personalized video. He cites Shopify as a model for becoming infrastructure for DTC businesses, says Higgsfield aims to be infrastructure for their distribution, and names AppLovin as another aspiration.
“when we look at public companies, and we exclude pharma and big tech, spend on sales and marketing is higher than spend on R &D.”
Mashrabov predicts most social media content will be AI-generated, with authentic content commanding a much higher CPM. Listen
He calls this something people still don't fully appreciate. He expects authentic shows to be far smaller in volume but to command a 10-50x higher CPM, and so to create more value than AI-generated content. He suggests even authentic creators will use AI for overlays on existing videos.
“It's going to be 10-50x higher CPM, whatever, than AI -generated content.”
Higgsfield's finance model projects $4.5BN in revenue by the end of next year, but Mashrabov personally believes it will exceed $10BN. Listen
He says the $4.5BN model assumes substantial deceleration, which his quant-minded finance team says is how the business works. The company is still pushing to grow at least 30% month over month. He bases his higher view on monetization-driven adoption by DTC brands and on Hollywood sentiment shifting toward using AI as a hybrid-production tool, which he describes from private conversations.
“Our current business model projects 4 .5.”
Higgsfield burned more than $10M of a $16M seed chasing hype before finding product-market fit by focusing on product and PLG. Listen
Mashrabov says Higgsfield spent more than a year searching for a product that worked, and he takes responsibility for optimizing for hype, narrative and attention instead of building a good product. With slightly less than $5M left and feeling they had 'one attempt left', the team committed to product and PLG on the belief that the best product would win. That led to the launch that took off.
“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. Everything instead of building a good product.”
Interviewing eight creative directors surfaced one consistent gap, camera control, which became Higgsfield's breakthrough product. Listen
After the pivot, Higgsfield asked eight creative directors what was missing from AI video, and every one of them said camera control, which they called essential to storytelling. The product launched on March 31 and Mashrabov describes product-market fit as immediate. He says VFX and camera control took Higgsfield from roughly $1M to $20M ARR in about the first three months.
“We spoke to eight creative directors about their experience with AI and what's simply missing. Everyone told us that camera control does not exist in AI.”
Higgsfield went from $1M to $1BN in annualized revenue in 18 months, versus 24 months for Cursor by Mashrabov's account. Listen
On the day of recording, Bloomberg reported that Higgsfield had crossed $1BN in annualized revenue. Mashrabov says this probably makes Higgsfield the third fastest to that milestone after OpenAI and Anthropic. He says image models for aesthetic photoshoots and product consistency took the company from $20M to $100M ARR.
“actually it took us 18 months from one million to one billion. For Cursor, it took 24 months.”
One Higgsfield customer went from a $99/month subscription to a deal over $6M a year within six months. Listen
He studies the stories of the largest customers on the platform and calls this level of acceleration mind-blowing. He attributes demand like this to direct-to-consumer e-commerce companies producing hundreds or thousands of ads a week, and to AI-made short-form dramas. He expects other customers to follow the same expansion path.
“So one customer started six months ago, spending just subscription $99 a month, $99 a month. And now we just signed a deal over 6 million.”
Listen to the episode Retention & customer success Link to this
Business customers make up slightly over 50% of Higgsfield's revenue, pure consumer use is around 10%, and mobile is under 10%. Listen
Mashrabov says the remainder of non-business revenue comes largely from aspiring creators and freelancers learning video AI to earn money. He describes them as 'a little churny' but says most come back within a year. Higgsfield invests in Higgsfield Academy and a YouTube channel to educate them as a future AI-native workforce. The West accounts for well over 70% of revenue; Seoul is the largest city by usage and the US the largest country.
“It's true that their behavior is a little churny. Within a year, most of them actually come back to try again.”
Listen to the episode Retention & customer success Link to this
Higgsfield runs no paid advertising and drives growth through an in-house team of over 150 creative professionals, nearly half its workforce. Listen
Mashrabov says the goal is for the best commercial video content to be made on Higgsfield and to show the workflows behind it. The creative team makes product launch videos and tutorials. It also made an AI-generated movie and open-sourced all of it. He says this owned content is what drives most of the revenue.
“And we have an in -house team of over 150 creative professionals.”
Outsourcing Higgsfield's influencer program to an agency went badly, and Mashrabov's lesson is that distribution has to be owned. Listen
Asked about controversy over the company's influencer marketing, Mashrabov says Higgsfield had only two people on the creator and customer success side and outsourced the work to an agency. He calls it not a good experience and says distribution matters more than ever. He describes typical influencer deals as a video production fee plus cost-per-click attribution, which he says works well on YouTube.
“We had just a team of two people on creator and customer success sides. And we just did outsource to the agency. And that was not a good experience.”
Listen to the episode Pipeline & demand generation Link to this
Producing 90 minutes of TV-quality AI video took over 100 hours of generated footage, so human creative selection still matters. Listen
Mashrabov cites Higgsfield's open-sourced AI-generated movie project. In it, average prompt length was over 3,000 words and each scene used at least 10 image references to define characters, backgrounds and positioning. He compares video models to a modern rendering engine like Unreal or Unity, which he says cannot be directed through text alone.
“for 90 minutes of, let's say, TV quality content, it was over 100 hours of AI-generated content. So creative decisioning, picking the right piece, is still very important.”
Large AI labs game benchmarks, and that video text-to-video benchmarks don't reflect real workflows. Listen
He says researchers at larger labs have told him that test data gets put into training and other tricks are used to hit benchmarks for quarterly bonuses. He attributes this to incentives at big companies. As evidence he cites OpenRouter data showing Google as the only relevant US incumbent, while in China, where he says benchmark obsession is probably lower, Tencent, Xiaomi and Alibaba are relevant. He admits he himself once wrongly chased benchmarks.
“what happens is that they start to put test data into the training.”
Higgsfield builds its own models only for specific customer-requested use cases, after Mashrabov concluded that chasing benchmarks with in-house models was a mistake. Listen
Mashrabov calls the earlier focus on benchmarks his mistake. Higgsfield still builds models where customers want them, such as its image model for aesthetic photoshoots and product consistency, which he says took revenue from $20M to $100M ARR.
“So it's all driven based on the customer feedback, not just by ambition to conquer the world and build the best model in the world.”
Most companies that claim to build their own models are post-training open-weight models, and the most valuable post-training uses customer decision sequences. Listen
He says the most valuable form is reinforcement learning on sequences of customer decisions, teaching the model to compress something like 10 steps into one. He expects more companies will have to do this. He cites OpenRouter data that the share of open-source models went from below 30% to over 60% this year. He still expects OpenAI and Anthropic to hold more than 50% of the market by dollars, partly because coders keep jumping to the newest models.
“you can teach the model to actually take, like, learn how to compress these 10 steps into one step. Like this type of reinforcement learning is the most valuable.”
Harry Stebbings is skeptical that OpenAI can win legal with a general model, because big law firms need deep, practice-specific functionality and buy through multi-year sales cycles. Listen
Responding to OpenAI's announced product for law, Stebbings calls it 'complete bullshit', citing the very deep functionality needed to serve the largest law firms. He also points to multi-year sales cycles with partnerships. Mashrabov replies that general tools for internal teams outside law firms are definitely happening, and both cite Solve Intelligence as a specialized winner.
“if you want to sell into these law firms, it's a multi -year sales cycle”
Whoever builds an AI-native system of record will win, and for Higgsfield that means a searchable, brand-aware asset library. Listen
Using Solve Intelligence's patent-workflow system of record as an example, he says creative assets today are scattered across Dropbox, Google Drive, Miro and Frame.io. Marketers want to search content and check it against brand guidelines through natural interfaces, which semantic understanding now makes possible. He says Adobe and Canva built for a 'pixel first era'. Higgsfield invests in a harness that learns a customer's visual style over time, which he says Claude and OpenAI cannot necessarily do.
“So whoever can create a AI native system of records is going to win.”
The only moats today are delivering an outcome and network effects. Listen
For Higgsfield, the outcome is helping businesses sell more through AI ads. On network effects, he says AI does not replace them and he doubts agent swarms will in the next five years. Higgsfield encourages its community to open-source projects others can fork, as on GitHub; he says these grew from about 10 seeded projects eight weeks ago to over 10,000, and he believes this can become a moat over time.
“We do believe that there are only two ways of modern value creation or moats today. First is when you deliver the outcome.”
Harry Stebbings largely thinks moats are bullshit, and that what matters is speed of decision-making, product execution and building value quickly. Listen
Stebbings says he was mocked for investing in Lovable because it was 'a wrapper', and concedes it was one. He argues that wrapper companies win by building very valuable features very quickly and compounding that over time.
“it's about speed of decision -making, product execution, and building value over time very, very fast.”
Investors have shaken hands on a price with him and then rallied other investors to come in at a 30% lower valuation the next day. Listen
He says this is why you can never be sure a deal is done until it is, though it did not happen with Yuri Milner. He also observes that Silicon Valley investors are extremely consensus-driven.
“people really shook hands and we do at this price. And next day, what I learned is that they called other investors and they pulled the syndicate to invest in 30 % lower valuation compared to what we discussed.”
Asian DTC e-commerce companies are rebuilding their go-to-market around AI, producing hundreds or thousands of ads a week to A/B test. Listen
Mashrabov says the trends driving Higgsfield's largest customers mostly come from Asia. He adds that short-form drama, which he calls a $10BN+ industry mostly owned by Chinese companies, now makes most new shows with AI end to end. He says Asian companies lean into tools like video AI because they want a direct relationship with customers rather than going through resale platforms.
“we're seeing a lot of direct -to -consumer e -commerce companies rebuilding their whole go -to market to be AI native, where they just make hundreds of ads, if not thousands a week, where they can A -B test what performs well.”
Listen to the episode Pipeline & demand generation Link to this
Silicon Valley's habit of changing jobs every two years makes Europe competitive on talent loyalty. Listen
Higgsfield has roughly 50 people in California, about 50 remote and over 300 in Kazakhstan. Its core team is people who won international math and physics competitions, and he notes Kazakhstan's 15% personal income tax as a draw. He rejects framing the Kazakhstan base as labor arbitrage, citing talent density, and hopes to create more dollar millionaires in Central Asia than any other company.
“in Silicon Valley, unfortunately, what I'm seeing is that people just jump between jobs every two years.”
Mashrabov's core management principle is to hire the best people, empower them and retain them, and he treats most other management theory as secondary. Listen
He says the CEOs he learns from, Jensen, Elon and Nik, abandoned conventional management principles such as one-on-ones and soft feedback. Instead they get down to the point and know the details, which corporate America might call micromanagement. He says AI makes access to raw signals and adoption data especially important.
“It's really as simple as hire the best people, empower them to do the best work, and just figure out how to establish a relationship and retain them.”
Mashrabov changed his mind on HubSpot: he no longer expects companies to build their own CRMs, because B2B go-to-market teams value a familiar interface. Listen
Hiring go-to-market talent showed him that people good at understanding and talking to customers may not readily accept a new interface. He also values HubSpot as a system of record that makes it easy to trace where a data flow went wrong. Harry Stebbings disagrees and still thinks HubSpot will be hurt.
“Oh, I was thinking that HubSpot is gonna get obsolete. Everyone is gonna build their own CRM. But when, especially when there's a higher end -scale B2B go -to -market team, just having familiar interface matters a lot.”
Listen to the episode Sales team, hiring & comp Link to this
A GTM-first culture can sacrifice product advancement, citing Snowflake versus Databricks. Listen
Mashrabov names Frank Slootman as the person he would most like on his board, for his no-bullshit culture. Stebbings counters that Slootman built a GTM machine at Snowflake, but Databricks 'wiped the floor' by prioritizing product over GTM.
“he built a GTM machine at Snowflake, but Databricks wiped the floor because they moved product as the priority, not GTM.”
LLMs amplify fraud against AI companies, including bots that consume credits and then auto-request refunds. Listen
He says new kinds of LLM-driven attacks and fraud are among the most urgent fires at Higgsfield. Because of his machine learning and data science background, he still gets personally involved in the statistics and data side of fighting them.
“basically bots using credits and then doing auto refunds”
Mashrabov's lesson from Snap is that momentum doesn't last, so a company should capitalize on it, including through fundraising, while it has it. Listen
He recalls Snapchat being worth $80BN when its gap with Meta was under 10x, and says its market cap is now below $15BN. He partly attributes this to public companies failing to tell their AI story. With Higgsfield's momentum positive, he says the company does not take it for granted.
“while we do have the positive momentum, we do not take this for granted.”
Higgsfield's teams moved to Claude between March and June, then its coders switched to Codex in mid-June, and Mashrabov sees tool preference as cyclical. Listen
Everyone, including the creative team, moved to Claude from March to June, which is when creatives started vibe coding. Coders moved to Codex as of mid-June, and top creatives followed over time. Mashrabov calls these shifts 'so cyclical'.
“But then we started to see that all the coders quickly moved from Claude to Codex as of mid June.”
Harry Stebbings personally spends two hours a day on Instagram short-form, writing and recording scripts as a new push for 20VC. Listen
Stebbings says the team decided Instagram short-form would be a big new push. He writes the scripts and records them himself, with two people handling production. Mashrabov calls it smart and says short-form clipping can build a top of funnel of hundreds of millions of views, as long as there is downstream monetization.
“I spend two hours a day just doing Instagram now.”
Listen to the episode Pipeline & demand generation Link to this