Higgsfield’s Growth: What Cheap AI Video Means for Ad Testing

Higgsfield, the AI video platform, said on September 29, 2026, that it had passed $1 billion in annualized revenue run rate. It calculates that figure from its latest four weeks of revenue multiplied by 13, and says it “does not represent revenue already earned over the previous year.” No independent source has confirmed it. In August the company raised a $400 million Series B at a $5.4 billion valuation, led by DST Global, and reported $700 million in annualized revenue and more than 30 million users, TechCrunch reported.

That is a long way from January, when Higgsfield announced a $200 million annual run rate, reached “in under nine months” after its platform launched in April 2025, and more than $130 million in Series A funding at a valuation above $1.3 billion. Most of the growth since has come from businesses. Business customers made up under a quarter of revenue in January and most of it by August, The Next Web reported.

The growth hasn’t been clean. In February Forbes reported racist and obscene videos made by the company’s marketing team and outside creators, plans sold as “unlimited” that users said were later throttled, and late creator payouts; a cofounder said “it was a mistake.” For advertisers, though, the bigger question is what cheap video does to the way ads are tested and bought.

What the numbers say about demand for AI video

Higgsfield doesn’t depend on one model. Its January release listed OpenAI’s Sora, Google’s Veo, Kuaishou’s Kling, ByteDance’s Seedance, Alibaba’s Wan and MiniMax alongside its own models. Its revenue measures demand for making images and video with AI, whichever model wins. Other companies show the same demand. Kuaishou said Kling’s annualized run rate was about $500 million in March 2026, and HeyGen said in June that it had passed $200 million in annual recurring revenue.

The ad platforms are building the same tools into their own products. Meta said the combined revenue run rate of its video generation tools hit $10 billion in the fourth quarter of 2025, a figure that reflects ad spending through tools it offers free, and by mid-2026 nine million small businesses were using at least one of its AI ad creative tools. Google added Veo to Asset Studio in Google Ads in March 2026, turning up to three images into a video of up to 10 seconds. Video ad buyers told the IAB that about one-third of their ad assets would use generative AI in 2026, up from a quarter in 2025.

So the cost of a usable video ad is falling fast. The cost of finding out which video works hasn’t fallen at the same pace, because testing still takes media spend and conversions.

More variants only help if the account can learn

Meta’s ad system was rebuilt to handle more creative. Its Andromeda retrieval engine narrows “tens of millions of ad candidates” to a few thousand for each person, and Meta’s engineers wrote that the number of creatives in its systems “is expected to grow significantly,” in the post announcing it. Practitioners read Andromeda as rewarding creative that differs in substance, such as the visual, format, angle or message, rather than small edits to one idea. Our pieces on Andromeda and on what Meta’s automation needs from your creative go into the detail.

Three limits decide whether more variants pay off:

• Delivery. An ad set needs about 50 optimization events within seven days to leave Meta’s learning phase, and adding a new ad counts as a significant edit. Meta’s guidance on ad volume warns that when an advertiser runs too many ads at once, each ad delivers less often, as Jon Loomer has documented.
• Statistics. At a 2% conversion rate, detecting a 20% lift between two variants with 95% confidence and 80% power takes about 21,000 visitors per variant, roughly 420 conversions in the control and 500 in the variant, by a standard sample-size calculation. Google’s Demand Gen experiments need at least 50 conversions per arm to show results, and Google notes that a test with no winner “might mean the creative variation you are testing is not a substantial one.”
• Hit rate. Across 578,750 Meta creatives, about 5% became winners, defined as ads that spent at least 10 times their account’s median, according to Motion’s 2026 creative benchmarks. Accounts spending under $10,000 a month launched about three creatives a week, against about 19 for accounts above $1 million.

Quality still matters more than quantity. Among ads that used generative AI, Kantar found those where the AI went unnoticed did better than those where it was obvious, and ads using generative AI scored lower on branding on average than ads without it. “The key isn’t whether you use GenAI, but how you use it,” Kantar wrote. Nielsen Catalina Solutions’ 2017 analysis of nearly 500 consumer-goods campaigns put creative at 47% of advertising’s sales effect, more than reach or targeting.

How to use cheap video without wasting budget

• Test concepts, not tweaks. Spend generation time on different hooks, angles, formats and offers. Ten versions of one idea teach the account little.
• Size the test to your conversions. If an ad set gets 100 conversions a week, it can judge a handful of concepts at a time, not dozens. Add new ones as old ones prove themselves or fail.
• Keep a control and one success metric. Use the platform’s experiment tools, run tests for at least a week, and decide in advance what result means a winner.
• Keep your brand codes. Put your logo, colors, product and recurring characters in every AI video. Kantar traced its branding gap mainly to obvious AI made with models that didn’t use the brand’s tone of voice or distinctive assets, and our piece on why human copy wins back ground makes the same point about text.
• Track cost per winning concept. The useful number is what it costs to find an ad that scales, including the media spent testing losers, not the cost of one video.
• Check rights and safety. Read the tool’s terms on commercial use and likeness, and review every output before it runs.

Mondelez showed where this leads at scale: it put $40 million into AI production to cut costs and still gave its biggest brand to a creative agency. Generating volume is easy to buy. Deciding what to make and reading the results is the work, which is where motion graphics and animation and performance marketing meet.

Frequently Asked Questions

What is Higgsfield’s revenue?

Higgsfield doesn’t publish audited revenue. It reported a $200 million annual run rate in January 2026, and TechCrunch reported $700 million in annualized revenue when it raised a $400 million Series B in August. On September 29, 2026, the company said it had passed a $1 billion annualized run rate, based on its latest four weeks of revenue multiplied by 13.

Is Higgsfield’s run rate the same as ARR?

Not exactly. Annual recurring revenue usually counts contracted subscription revenue. Higgsfield’s figure takes its most recent four weeks of revenue and multiplies by 13, so it includes one-off and usage-based spending and assumes that pace continues. The company says the figure does not represent revenue already earned over the previous year. It also says subscriptions fell from more than 90% of its revenue to a little over 60%.

Who founded Higgsfield?

Higgsfield was co-founded by its chief executive, Alex Mashrabov, Snap’s former head of generative AI, and its technology chief, Yerzat Dulat. Mashrabov earlier co-founded AI Factory, which Snap bought in 2020. The company was set up in 2023, released its first app, Diffuse, in 2024, and launched its current video platform in April 2025. It works with models from OpenAI, Google, Kuaishou, ByteDance and others, as well as its own.

How many ad variations should I test on Meta?

As many distinct concepts as your budget can judge. An ad set needs about 50 optimization events a week to exit the learning phase, and each variant needs enough conversions to compare fairly, often hundreds for small differences. Most accounts do better with a few meaningfully different concepts at a time than with many near-identical versions that compete for the same small budget.

Do AI-generated video ads perform as well as human-made ads?

They can. Among ads using generative AI, Kantar found that ads where the AI went unnoticed performed better than those where it was obvious, though ads using it scored lower on branding on average. Nielsen Catalina Solutions put creative at nearly half of advertising’s sales effect, so execution and branding matter more than whether a person or a model made the ad.