The Information reported on Sunday that Higgsfield AI, a San Francisco-based video generation platform founded by Alex Mashrabov and Yersat Durat, is in talks with investors to raise between $300 million and $500 million at a pre-money valuation of $5 billion. That’s more than a fourfold increase from its $1.3 billion valuation just six months ago. DST Global, a company founded by early Facebook backer Yuri Milner, is also in talks to join. The round is not over yet.
News of the funding came along with revenue numbers that explained investor appetite. Higgsfield’s annual revenue exceeded $500 million this month, up from $200 million at the end of 2025. A platform that didn’t exist before March 2025 has generated $500 million in annualized sales in about 15 months. This trajectory, along with the simultaneous exit of two major rivals from the commercial video market, is the story behind the $5 billion figure.
What went wrong for Mr. Higgsfield?
OpenAI ended its service on April 26, 2026 after the Sora consumer app reportedly cost approximately $1 million per day to run while generating just over $2 million in total revenue. The Sora API is scheduled to be deprecated on September 24, 2026. Before it was discontinued, Higgsfield was Sora’s largest customer in terms of spending and usage. This was a critical dependency for a single model platform. For multi-model aggregators like Higgsfield, Sora’s closure was contentious. Single models may die, platforms that abstract away much of them will survive, but single model apps will not.
Runway, Higgsfield’s most direct competitor among dedicated AI video platforms, closed a $315 million Series E in February 2026 at a valuation of $5.3 billion. But Runway clearly signaled a pivot away from commercial video production to “world models,” AI systems that simulate physical environments for gaming, robotics, and medical use cases. The company’s 4.5-generation model currently leads the independent text-to-video benchmark, but the company’s strategic energies have shifted elsewhere. This has left a significant part of the professional video production market less competitive than it was at the beginning of the year.
How the Higgsfield engine actually works
What distinguishes Higgsfield architecturally from simple model marketplaces is its cinematic logic layer, an inference engine that sits between the user’s creative intent and the underlying generational model. Even if the creators say “dramatically”, that’s not a description that the video model can live up to. The film logic layer translates vague creative direction into a structured video plan, including timing rules, camera motion specifications, and visual priorities that models can actually use. The translation step is essentially invisible to the user, as simpler deterministic tasks are routed through a faster language model, and more ambiguous interpretations of intent are routed through a more competent language model.
The platform’s current flagship interface, Cinema Studio 3.0, adds a physically aware generation engine with native audio synthesis and allows users to direct shots using over 50 named camera presets (Dolly-in, Crash Zoom, 360 Orbit, Dutch Angle, FPV Drone) rather than typing motion descriptions into text boxes. Proprietary models handle specific tasks. While the DoP I2V-01 model conveys cinematic grammar directly with diffuse weights, WAN 2.5 handles up to 10 seconds of 1080p clips with consistent characters, intentional camera movement, and synchronized audio in a single render.
Foundational models available at Higgsfield include Veo 3.1, Kling 3.0, Seedance 2.0, WAN 2.7, and our own in-house model, giving you over 15 options in one subscription. Veo 3.1 delivers the best results for atmospheric and outdoor scenes. Kling 3.0 is a cost-effective option for simple clips. Seedance 2.0 processes multi-shot sequences with synchronized audio. The platform’s pitch is that marketing teams don’t need five separate accounts and five credit systems to access all three.
The economics of realizing free tiers
Running AI video generation at scale requires massive GPU computing. This produces approximately an order of magnitude more output per minute than a standard language model query. The economics behind Higgsfield’s ability to offer a free tier and competitive pricing is the move to GMI Cloud, a specialized GPU-only infrastructure provider that has reduced the platform’s compute costs by 45%. On hyperscale cloud providers like AWS and Google Cloud, NVIDIA H100 GPUs cost between $7 and $13 per hour. GMI Cloud offers equivalent H100 access starting at $2 per hour, with a purpose-built inference engine that avoids the 10-15% virtualization overhead typical of broader cloud platforms.
This cost gap has allowed Higgsfield to price his video production at about $30 per clip. The work so far cost $10,000 and would have taken weeks of traditional production. A credit-based consumption model scales revenue directly to usage rather than seats, with different models and quality levels earning different amounts of credits. Corporate clients who spend $200,000 or more a year are a completely different demographic. Get shared workspaces, prioritized computing, and premium content safety tools.
How businesses moved from viral tools to enterprise platforms
In the first few months after its launch in March 2025, Higgsfield’s growth was driven by aggressive viral marketing, some of which turned out to be debt. In February 2026, Forbes documented that the company’s marketing team was distributing AI-generated racist content to influencers as promotional materials, including cartoon characters from children’s movies containing racial slurs. Higgsfield’s X account was suspended that week. CEO Mashrabov publicly acknowledged that “our internal processes and external communications are not always aligned with our core values.” The company then launched a similarity scoring tool in March 2026. The tool alerts you to potential matches to known celebrity likenesses or brand logos before delivering content, and reported a video detection accuracy of 86.6% versus 48.5% for leading third-party alternatives.
This controversy did not have a material impact on earnings. By the time Forbes published its findings, Higgsfield’s ARR had already exceeded $300 million. The more significant changes in business were structural. Social media marketers currently account for about 85% of platform usage, and about 80% of them produce commercial content (ad campaigns, e-commerce product videos, UGC-style content for brands) rather than personal creative experiments. Approximately 70% of the platform’s user base is made up of business customers, with some accounts spending more than $200,000 annually. Collectively, the company serves 25 million users and generates approximately 6 million video and image productions (including 2 million videos) per day.
Team and Enterprise plans, launched in November 2025, formalized the upmarket movement with shared workspaces for agencies, brand studios, and in-house creative operations teams. Higgsfield Ads products allow you to create commercials from your product images in one click. UGC Factory targets e-commerce creative teams who produce large amounts of product content. Supercomputer Agent, released this month, automates a broader range of marketing workflows.
Is $1 billion ARR achievable by the end of the year?
Higgsfield said the company is targeting $1 billion in annual sales by the end of 2026, which would roughly double from now in about six months. The revenue curve clearly does not exclude that possibility. The company went from $100 million to $200 million in ARR in about two months in the second half of 2025, and from $200 million to $400 million in the next five months, according to data tracked by Sacra. It reportedly jumped to $500 million in June 2026 alone.
The structural risk to that trajectory is the aggregator model itself. The underlying model developers, Google, Kuaishou (which produces China’s Kling, which is required by the National Intelligence Act to collaborate with government intelligence agencies on demand), and ByteDance, could increase API access, build competing workflow platforms, or lower the price of Higgsfield as the market matures. Several platforms with similar multi-model positioning have already appeared. And while Higgsfield’s Cinematic Logic Layer and Cinema Studio currently offer true differentiation, those benefits are lost if model providers decide to build the same orchestration layer on their own.
With a pre-valuation of $5 billion and annual revenue of approximately $500 million, Higgsfield’s earnings will be approximately 10 times higher. That’s high for a software company, but in line with what the market is currently allocating to fast-growing AI application companies. Investors’ central argument is whether commercial AI video has found a lasting business model, or whether Sora’s failure (huge computing expenditures for meager revenue) is the case in this category, and Higgsfield has simply delayed the same reckoning by targeting commercial buyers rather than consumer enthusiasts.
Evidence so far suggests that the latter is unlikely. A marketing team that produces 50 product videos in two hours has a concrete ROI calculation that an enthusiast experimenting with AI videos doesn’t. Buyers for your business are hard to come by and churn is slow. Whether that durability can sustain the $1 billion run-rate threshold, or whether the funnel narrows significantly beyond $500 million, are the key questions to answer with the new funding.
FAQ
What is Higgsfield AI and how does it work?
Higgsfield is a multi-model AI video generation platform that provides subscribers with access to over 15 underlying generative models, including Veo 3.1, Kling 3.0, and Seedance 2.0, from a single subscription and credit system. Its key architectural innovation is the movie logic layer, an inference engine that translates vague creative intent into structured video plans before the model generates frames. Its flagship interface, Cinema Studio 3.0, allows users to direct shots using over 50 named camera presets and a physical recognition generation engine, rather than relying solely on interpreting text prompts.
How would you compare Higgsfield to a professional video work runway?
Runway trains its own model, and its Gen 4.5 currently leads independent text-to-video benchmarks. We also have a history of working with Hollywood studios and advertising agencies. Higgsfield does not train the underlying generative models (although it does have its own in-house models for specific tasks). Aggregate the best available models and add workflows and enterprise infrastructure on top. Runway is clearly focused on robotics and simulation world models, while Higgsfield is also focused on commercial short video production. For the best output quality of a single model, Runway remains the benchmark. Higgsfield is a practical alternative for teams that need flexibility across multiple model types without managing multiple subscriptions.
Is it safe to use Higgsfield for professional or corporate work?
In February 2026, Forbes documented that the company’s marketing team distributed racist promotional content generated by AI, leading to a platform suspension. The company publicly acknowledged this failure and in March 2026 launched a similarity scoring tool that flags potential celebrity likenesses and brand logos before distributing content. Higgsfield’s Section 6.1 clause states that users have no expectation of privacy regarding the content they submit through the platform, which companies may use for moderation and model improvement. Corporate buyers who manage confidential customer assets or unreleased products should review these data terms carefully before uploading any sensitive material.
What happened to OpenAI Sora? Is Higgsfield a viable alternative?
OpenAI shut down Sora’s consumer app and website on April 26, 2026. This is because the product had an operating cost of about $1 million per day, but only generated about $2 million in total revenue. The Sora API is scheduled to be deprecated on September 24, 2026. Higgsfield offers access to Sora 2 as one of its integrated models, alongside Veo 3.1, Kling 3.0, and more. This means creators who relied on Sora will continue to have access to their models through Higgsfield until the API is shut down, while also gaining access to competing models not available in Sora’s standalone app.
