Image source: Arcads official site. The page shows AI actors, generated ads, and the creation entry point.
One morning in March 2024, Romain Torres opened Slack the way he usually did. His company Arcads had just launched a product: users pasted in a script, chose an AI actor, and generated a spoken advertising video in minutes.
Normally, the team’s Slack would show 10 to 20 Stripe payment notifications overnight. That morning, Romain kept scrolling. Hundreds of new orders had arrived. A user had posted an Arcads-generated video on Twitter, and the product unexpectedly went viral.
Then he opened customer support, and the excitement changed. The traffic had arrived before the product was ready. Arcads was almost unusable under the load. The team refunded every order and rebuilt a large part of the product.
That story comes from Romain’s founder interview in November 2025. In the same interview, he showed the Stripe revenue curve and said Arcads reached $10 million ARR in 20 months with more than 6,000 paying customers and a team of eight. Those figures are founder-disclosed and not independently audited.
The important decision is still useful: the team gave back a few hundred orders in order to protect the subscription business that came later.
Before the software worked, they made ads by hand
Arcads’ founders, Dylan Fournier and Romain Torres, had previously built mobile apps together. Every new app needed attention, and that attention was expensive. They had to pay creators, produce ad variations, and test materials in crowded app markets.
AI actors offered a different cost structure. But in January 2024, Arcads did not yet have a full dashboard. Customers could not simply log in and generate videos. The founders instead contacted brands directly and offered to make UGC-style ads with AI, faster and cheaper than a traditional creator workflow.
That manual phase looks primitive, but it answered the hard questions before the software did. Would brands pay for AI-generated ads? Were they buying a cheap video, or were they buying more creative tests? If the first result was weak, would they come back for more variations? How much budget could move from shooting to AI production?
Stripe’s Arcads customer story confirms the same pattern: the team gathered feedback, tested pricing, and used Payment Links before launching three subscription tiers and enterprise plans. Manual pricing was not a detour. It was the market research system.
The viral night exposed false product-market fit
The Twitter spike pushed Arcads from roughly 5,000 euros in monthly revenue toward about 64,000 euros and near a $1 million revenue run rate. If the team had only looked at the revenue chart, it would have looked like product-market fit.
Support tickets told the truth. Customers had paid, but they could not reliably use the product. The team chose to refund and rebuild, turning the viral moment from a victory lap into a stress test: demand was real, but delivery was not yet trustworthy.
That failure mattered because Arcads had already learned what customers wanted and what price they could accept. Rebuilding did not mean guessing the market again. It meant turning a previously manual creative service into a reliable subscription workflow.
Arcads now offers more than 1,000 AI actors, actor swapping, translation, remixing, product showcases, and a workflow canvas. The product moved from “generate one talking-head video” toward a production line for continuous ad testing.
That distinction is central. A single generated video can be commoditized by new models and cheaper tools. A marketing team that needs to test actors, hooks, scripts, languages, and markets every week creates recurring demand. The budget is not for one asset. It is for creative throughput.
At $10 million ARR, pricing still changes
Stripe reports that Arcads generated $4.2 million in revenue in 2024 and expected to exceed $13 million in 2025. It also says more than 6,000 customers generated all revenue through subscriptions, and that payment acceptance improvements recovered about $124,000 in additional transactions.
Once a product reaches that scale, payment retries, checkout conversion, multiple currencies, and subscription operations become leverage. Arcads had already learned what it could deliver and what customers would buy. Stripe then made fewer payments leak out of the system.
The company did not flatten all customers into a simple self-serve plan. Romain said its largest enterprise contract had passed $100,000 per year. Self-serve customers could buy by card; large brands like Nike, Adidas, or Samsung needed sales support. In 2025, Romain spent about two months building the sales function while Dylan kept product focus.
Arcads still has unanswered questions. It has not publicly disclosed acquisition cost, retention, gross margin, or the revenue mix between self-serve and enterprise. AI actors and video generation models are also becoming easier to copy. $10 million ARR proves demand, but it does not by itself prove durable retention.
Even so, Arcads made one decision that many early AI companies avoid. When orders arrived before the product could carry them, it did not confuse payment with delivery.
The refund was painful. It also protected customer trust and forced the product to become worthy of recurring revenue.
The builder lesson
Arcads is not just a story about AI video ads. It is a story about sequencing.
First, the founders manually delivered the outcome and negotiated prices. Then they learned which workflow customers were really buying. Then a viral spike exposed that demand was bigger than the product’s reliability. Then the team refunded, rebuilt, and turned a fragile generator into a subscription production system.
For AI builders, the lesson is direct. Early revenue can be a useful signal, but only if the product can honor it. A spike that produces angry customers is not product-market fit. It is a message from the market that the job matters and the system is not ready.
Arcads kept the message and returned the money. That is why the later subscription business is more interesting than the viral night.