The AI Video Startup That Wins by Using Everyone Else’s Models.
By Creatives Takeover Editorial Team · October 26, 2026
From a GitHub Username to a $5.4 Billion Company.
Higgsfield launched with its own AI video model. A few weeks later, its founders realized that was the wrong bet.
The company's CEO, Alex Mashrabov, has described what they noticed. A new video model was landing almost every week, from labs around the world, and no single lab was going to be best at every kind of video. Building and defending one model of their own meant fighting that race alone. So the founders changed course. In August Higgsfield was valued at $5.4 billion, and the product that earned that valuation is built mostly from models made by other companies.
The Race Nobody Can Win for Long
To understand why that choice worked, picture how AI video has developed. A handful of well funded labs, including Google, OpenAI, ByteDance, Kuaishou, Alibaba and MiniMax, keep releasing new models. Each new release is better than the last at something, and the title of "best model" changes hands constantly.
They also differ in what they are good at. According to one review of Higgsfield, Kling leads on human movement, Google's Veo on realistic physics and sound, Sora on story sequences and ByteDance's Seedance on product footage for ads. Nobody has built one model that wins every job. For a small company that is a hard thing to compete against. Training a top video model takes huge amounts of computing power, and the lead you earn can be gone within weeks.
Higgsfield's founders saw that and made a different bet. Instead of trying to win each round, they would become the place where the winners all show up.
The Two Founders Behind the Bet
The decision to give up on building one model fits the people who made it. Mashrabov is a veteran of the field. He sold his earlier computer vision company, AI Factory, to Snap for $166 million and then led generative AI work at Snap before starting Higgsfield in 2023 with Yerzat Dulat and a third co-founder, Mahi de Silva.
Dulat, the chief technology officer, is a Kazakh engineer with an unusual path. By his own account he started in physics olympiads, studied computer science for two or three years, got bored, and taught himself machine learning at a time when Kazakhstan had very few specialists. In the developer community he is known by the handle "higgsfield," which the company also carries. He says OpenAI's co-founder emailed him in 2018 after one of his research papers and he chose to keep working on his own projects, though investor accounts of why differ. A founder who has built models himself is well placed to see how fast they are being overtaken, which may be why the company moved on from its own so quickly.
What Higgsfield Actually Sells
It helps to think of it as a shop that tests every camera brand so the customer does not have to. Higgsfield offers around 35 models for images, video, audio and text, by one count, all under a single account and a single balance of credits. A customer who wants a cinematic shot can pick a model, or let the platform pick one.
What the company really sells is everything around those models. Its early hit was "Click to Video," where a user picks a curated style preset and gets a cinematic shot without writing a long, careful prompt. It has built studios for film-style scenes and for marketing and ad content. Its character system, Soul ID, learns a person's face from five or more photos so the same person looks consistent from shot to shot, which is a well known weakness of AI video.
Speed is part of the product too. Mashrabov has said Higgsfield tries to add each new leading model within about 24 hours of its release. He describes the company's long term value as orchestration: choosing and combining the right models to finish a whole creative job, not simply reselling access.
To picture how that looks in practice, imagine a marketing manager who needs a short video ad for a new sneaker. This is an illustration, not a real customer. She writes one description and runs it through three different models side by side. One gives the best movement, another the best lighting, a third the cleanest product shot. She picks the winner, fixes a detail, and exports it, all inside one account. Without a platform like this she would need three subscriptions and a lot of time to compare.
The Day a Model Disappeared
The strategy faced a real test this year. On March 24, 2026, OpenAI announced it was discontinuing Sora, the AI video model it had launched with great fanfare. The Sora app went dark on April 26, and the developer interface was switched off on September 24, according to OpenAI's own help page. OpenAI never gave one official reason, but reports pointed to high computing costs, falling usage and a shift of resources toward other products.
For a company built around a single model, that would have been a disaster. One report says Higgsfield was Sora's largest customer by spend and usage before the shutdown, which is the kind of dependency that can kill a business. Because Higgsfield already offered many models, it lost one option among dozens and customers moved to the others. Sora's end made the company's argument for it: any single model can disappear, and a platform that spans many survives.
Following the Best Model Wherever It Comes From
The strategy has a global side that is easy to miss. According to one report in the Chinese technology press, Chinese-made models dominate usage on Higgsfield, with ByteDance's Seedance in the lead, followed by Kling, MiniMax and Alibaba's Wan. The same report says Higgsfield has secured exclusive versions of some models, including an enhanced fast version of Seedance 2.0.
That reflects the company's position. It has no loyalty to a flag or a lab. It follows whichever model is winning, whether it comes from California, Beijing or Hangzhou. Its own engineers are largely in Kazakhstan, and the company announced new investment from Nvidia and the growth fund TCV in October, during Nvidia chief executive Jensen Huang's visit to Astana.
From Creators to Companies
The speed of the rise is part of the story. Higgsfield launched its platform publicly in March 2025. A $50 million round that September valued it at $1 billion and made it Kazakhstan's first unicorn. In January 2026 an $80 million extension lifted the valuation to $1.3 billion, and in August a $400 million round raised it to $5.4 billion. Two months after that came the Nvidia and TCV investment.
Higgsfield's first customers were individual creators and social media marketers. The company has since moved toward businesses. Enterprise customers now account for most of its revenue, up from less than a quarter at the start of the year, according to reports on its latest funding round. The company says it works with 390 of the Fortune 500, a figure that comes from its own announcement.
The business argument for companies is different from the one for creators. Large teams care about security, collaboration and the ability to produce video in bulk, and those needs are largely about software around the model, not the model itself. Mashrabov has said the company's average contract size was nearly doubling every quarter as it moved up market.
The numbers have moved quickly. Higgsfield says its annualized revenue was about $200 million at the end of 2025. It then reported $500 million in June and about $700 million at its August funding round, though one outlet put the August figure at $500 million. Mashrabov says it is on track for $1 billion. He has also said the company is cash flow positive. Higgsfield raised $400 million at a $5.4 billion valuation in August, led by DST Global, which values it above its rival Runway, which raised at $5.3 billion in February.
Where the Model Is Weakest
The approach has real weak points, and the company's critics name them.
The first is that its suppliers can compete with it. Google, ByteDance and Kuaishou all sell access to their models directly, and reviewers note that a customer who only needs one model may pay less by subscribing directly. One report calls the aggregator structure itself the main risk to Higgsfield's trajectory. If a lab decides to restrict or reprice access, the company has little leverage.
The second is margins. A Chinese outlet claims Higgsfield sells models at close to cost and makes its profit in the layers above them. That has not been confirmed, but if it is right, the business depends on the workflow tools, not the models.
The third is trust. Regional press has reported user complaints about subscriptions and refunds, and reviewers have questioned the way its discounts are presented. For a company that now sells to large businesses, that reputation matters.
The fourth is geopolitics. If the strongest models increasingly come from Chinese labs, rules about access, exports or data could change what Higgsfield can offer in some markets.
Why "They Just Resell Other People's Models" Is the Wrong Lesson
It is easy to dismiss the company as a reseller. That misses where the value sits. Anyone can sign up for the same models. What is hard is knowing which one to use for a given shot, adding each new one in a day, keeping a character looking the same across scenes, and giving a corporate team a secure place to work. Mashrabov has argued that value in AI will accrue to the applications built on models, not only to the models themselves.
Think of a camera shop that tests every brand, knows which lens suits which job, and teaches you to use it. The shop does not make the cameras, yet customers keep returning because it saves them from being experts in all of them. Higgsfield is betting that in AI video, being that shop is worth more than being one of the manufacturers.
What a Founder Can Take From This
Do not build what you can rent when it changes every week. If the best tool in your market is replaced every few weeks, owning one of them can be a trap. Higgsfield tried it first and left within weeks.
Build what is hard to rent. The things that were not available to every customer were the workflow, the speed of adopting new tools, the character consistency and the enterprise features. Those are the parts customers pay for.
Plan for your supplier to vanish. The Sora shutdown showed that even the best funded AI lab can end a product on a few weeks' notice, so a business should be able to swap a key supplier without rebuilding.
Keep the customer relationship. Higgsfield's customers use its interface, its credits and its studios, so they stay when the models underneath change.
The next test is whether the labs it depends on keep selling to it. For now, new models keep arriving nearly every week, and Higgsfield's job is to be first in line when they do.