Higgsfield

Executive Summary

In January 2025 Higgsfield, an AI video start-up, had spent more than $10 million of the $16 million it had raised. It had been started by Yerzat Dulat, a self-taught programmer from Kazakhstan, and Alex Mashrabov, who had run generative AI at Snap. A million people had tried their phone app, and few of them paid. Fewer than 15 people still worked there. Mashrabov, the chief executive, said later that they had one attempt left.

Less than two years later investors valued Higgsfield at $5.4 billion. The company says its sales are running at more than $1 billion a year, and it counts 32 million users.

The case follows six decisions, from that January to one still open in November 2026. Among them: what to build with the money that was left, whose technology to use, how far to go to win customers and which partners to trust. At each one you decide first, and only then read on.

One Attempt Left, 2018-2025

In May 2023 Yerzat Dulat, a programmer in Kazakhstan, pitched an idea to a Kazakh investor, Murat Abdrakhmanov. He had no product, only a plan for software on which artificial-intelligence models could be trained. Abdrakhmanov agreed to put in $250,000 [1]. Dulat hired a team in Kazakhstan, where most of the company's engineers still work, and looked for an experienced co-founder who could promote the project in the United States [2]. By his account another investor put him in touch with Alex Mashrabov, then at Snap, the maker of Snapchat. "We agreed on a 30-minute call, but ended up talking for seven hours," Dulat said later [1]. The two founded Higgsfield together.

Dulat had gone to a school for gifted mathematicians and physicists, skipped university and taught himself, first programming from online courses and then machine learning from research papers [1]. Mashrabov grew up in Chelyabinsk, in Russia, studied at the Moscow Institute of Physics and Technology and worked at Yandex, Russia's largest search engine [3][4]. In 2018 he co-founded AI Factory, whose software put a user's face into short looping videos. Snap built a Snapchat feature on it and then bought the company. TechCrunch, the technology site, reported the purchase in January 2020 and put the price at $166 million, which Snap did not confirm [5]. Mashrabov stayed at Snap for almost four years, ran its work on generative AI and left in September 2023 [6][3].

The two dropped Dulat's idea. Mashrabov judged that OpenAI and Anthropic, the makers of the leading chatbots, were too far ahead to catch, and that video was "a completely new space where we can jump in very early and ride the wave" [3]. They would build a video model of their own, meaning the trained software that turns a written prompt or a photograph into moving pictures, and put it in a phone app. On 3 April 2024 Higgsfield announced that investors led by Menlo Ventures, a Silicon Valley venture-capital firm, had put in $8 million, and it launched an app called Diffuse [6]. Further investments took the total to about $16 million, by the founders' count [7].

Diffuse put a user into a generated video from a single selfie. It ran on Higgsfield's own model, which 16 people had built in under nine months [6]. Mashrabov explained why he had chosen the phone:

By prioritizing iOS and Android apps instead of desktop workflows, we enable creators to create compelling social media content anytime and anywhere.

Alex Mashrabov, to TechCrunch, 3 April 2024 [6]

The app was free. Mashrabov expected to charge companies' marketing departments later, for extra features and for making video in bulk [6].

By August 2024 a million people had used Diffuse, by the company's count, and Higgsfield began to charge a subscription [8][9]. Its users proved unwilling to pay, and few stayed [10][11]. The company, Mashrabov said later, drifted towards apps in which "people can upload photos of people and make them kiss each other or hug each other", because such uses brought in users [3].

Meanwhile better-funded companies had entered the field. In June 2024 Kuaishou, the company behind one of China's largest short-video apps, released a model called Kling, and Runway, an American start-up, released one with settings for the camera's movement [12][13]. In December OpenAI opened its video model, Sora, to paying users of ChatGPT, and Google announced a model called Veo 2 [14][15].

Higgsfield tried other products. ReelMagic, shown in December 2024, was a website that turned a story idea into a video of up to ten minutes, using other companies' models beside Higgsfield's own. It was open only to people on a waiting list [16][17]. Hera was a streaming service for short serial dramas, made by AI to be watched on phones; the company tried it and shelved it [18]. A third idea was to supply video to the studios that make such dramas. In the autumn of 2024 Mashrabov noticed that people who made social media video as their job edited it on desktop computers, not phones [3].

By the start of 2025 the company had spent more than $10 million of its $16 million, by Mashrabov's later account, and employed fewer than 15 people. It felt, he said, that it had one attempt left [19][20]. Asked afterwards whether he had thought the company might fail, he said: "Most of the time, it felt like things are not working at all" [3].

In January 2025 Mashrabov had four ways to spend what remained. He could stay with Diffuse and look for a reason for its users to pay, in competition with OpenAI, which now included a video model in a ChatGPT subscription. He could finish ReelMagic, open so far only to a waiting list, for people who wanted to tell long stories. He could supply the short-drama studios, a handful of business customers. Or he could turn to people who make adverts and other social media video as their job, and build a web tool for them. They had budgets, but Runway already sold them tools, and Higgsfield would be starting again with less than $6 million.