From $1M AI Bills to Profitability: One Founder's Story

Nobody complains about AI costs until the bill gets too high.

That's what Ben Cera, founder of Polsia, says at the start of his story — and it cuts right to the heart of what's quietly killing a lot of promising AI startups right now. The product works. Users love it. The numbers are going up. And then the invoice arrives.

Ben's journey is one of the most honest founder stories I've come across lately. It's not a highlight reel. It's a real account of building something cool, watching it grow faster than expected, and then staring down a million-dollar monthly AI bill that was on track to bankrupt the whole thing.

The Beginning: Build What You Actually Want

Polsia started in Los Angeles in October 2025. Ben had been living there for seven years, working with Travis Kalanick — yes, the co-founder of Uber. During this time he was experimenting with building AI-native products on the side. He tried a tool called Decision AI. He tried something called VibeCoder. His own words: "honestly none of them really took off."

That failure taught him something important. He stopped trying to guess what the market might want and started building something he actually found useful and cool for himself. That shift in mindset — from market-first to problem-first — is what gave birth to Polsia. It's a lesson that keeps showing up in the best founder stories. The products that resonate most deeply tend to start as something the founder genuinely needed.

Paris: Validate First, Optimize Later

After conceiving the idea in LA, Ben moved back to his apartment in Paris in January 2026 to lock in and build. This phase was all about product development and customer validation. The AI bill was still manageable — starting at $240, then climbing to $25,000 — but he was already aware that leaning on Anthropic and OpenAI for compute wasn't going to be free.

His philosophy at this stage was deliberate: validate first, fix costs later. There's no point engineering a cost-efficient system for a product nobody wants. So he focused on building the core experience, embedding what he describes as Paris's culture of beauty and community into the product's ethos, and figuring out whether customers actually wanted what he was making.

They did.

San Francisco: Explosive Growth and a $1,000,000 Bill

His pre-seed investors at True Ventures made the case that San Francisco was where he needed to be. He moved, and the growth exploded. In March alone, Polsia went from 500 to 5,000 paying users. That's the kind of traction that gets investors excited and gets founders on stage at conferences.

But scaling fast with an AI product that relies on expensive closed models has a brutal financial consequence. The AI bill went from $100,000 to $500,000, and then to $1,000,000 in a single month. Not annually. Monthly. His user base had grown from 20,000 to 100,000 users, and the infrastructure he had built was starting to crumble under the load. The business model that worked at 20,000 users was completely broken at 100,000.

This is the part of the story most founders don't talk about publicly. You can go viral, get press, hit your growth milestones — and still be on a direct path to bankruptcy because the unit economics don't work at scale.

The Pivot: Own Your Infrastructure

Ben's investor Ilan from Sapiom was the one who put it plainly: this cost trajectory is unsustainable. Something had to change.

The solution Ben landed on was to stop renting intelligence from OpenAI and Anthropic and start renting raw compute instead. As he explains it: "What we could do is we could rent GPUs and we could start using open-source models." This meant visiting the Sciforium Data Center in California and seriously exploring what it would look like to stand up their own AI infrastructure.

The CEO of Sciforium, Hassan, framed the open-source model landscape in a way that made it click: the race to build better open-source models is a "never-ending race" that keeps improving. The gap between open-source and the top closed models has been closing fast, and for many production use cases, open-source is already more than good enough — especially when you're running it on your own hardware and controlling your own costs.

This isn't a trivial switch. It took two months of work to configure the new system properly. But the payoff was dramatic.

The Result: 90% Cost Reduction

By June, the monthly AI bill had dropped from $1,000,000 to $100,000. A 90% reduction. Ben calls it a "game changer," and honestly, that's an understatement. That's the difference between a company that's bleeding out and a company that's profitable.

The downstream effects of that cost reduction ripple through everything. Better margins mean better pricing flexibility. Better pricing means you can offer a free tier. A free tier means you can bring in users who couldn't justify paying before, which accelerates growth. The whole flywheel starts turning differently when the infrastructure isn't eating you alive.

The Bigger Vision

What Ben is building with Polsia isn't just another AI chatbot. His explicit vision is that Polsia becomes synonymous with AI that *takes action* — not just AI that *provides information*. He draws a clear distinction: ChatGPT gives you an answer. Polsia does the thing.

That shift from informational AI to agentic AI is where a huge chunk of the industry is heading, and Ben wants Polsia to be the iconic name attached to that category. Used by billions. Accessible to everyone. That last part is only possible if the cost structure actually works — which is exactly why the infrastructure overhaul wasn't just a financial fix, it was a strategic necessity.

When he meets with his early customers at the end of the video, what they say is simple: the product is just easy to use. That's the validation. Not the growth metrics, not the press coverage — just real people saying it works and it doesn't feel intimidating. That's what he set out to build.

What Every AI Founder Should Take From This

Ben's story is a compressed masterclass in what building an AI startup actually looks like in 2025 and 2026. A few things stand out:

• Build for yourself first. Decision AI and VibeCoder didn't go anywhere. Polsia did, because it started from a place of genuine personal need.
• Validate before you optimize. The Paris phase was expensive relative to zero, but it was the right call. Don't engineer cost efficiency into a product before you know people want it.
• Closed models are a starting point, not a forever strategy. OpenAI and Anthropic are incredible for prototyping and early validation. But if your product scales, their pricing will eventually become your biggest existential threat.
• Open-source plus owned infrastructure is the endgame for serious AI products. A 90% cost reduction isn't a rounding error — it's the difference between a viable business and a cautionary tale.
• Affordable AI is the prerequisite for mass adoption. If your cost structure forces you to charge $50/month, you're capping your own ceiling. Fixing the infrastructure unlocked the free tier, which unlocks the next hundred thousand users.

The bill almost won. It didn't. And the story of how Ben fought back is one worth paying attention to — especially if you're building something in this space right now.