Gorkem Yurtseven is the CTO and co-founder of FAL, an $8B generative media platform that hosts open and closed source image and video models and exposes them as APIs through its own inference engine. He joined us at the last SaaStr AI Annual, and here’s what he’s learned building one of the fastest-growing AI infrastructure businesses out there.
More like this at 2026 SaaStr AI Annual, May 12-14 in SF Bay!!
AI Gross Margins Are Lower. And They’re Staying That Way.
Everyone assumed the model cost curve would save them. It hasn’t.
The old SaaS logic was simple: once you built the product, selling to the next customer cost almost nothing. Marginal cost approached zero. That’s where the 80-90% gross margins came from.
AI broke that equation. Every new user, every new query, costs real money to serve. GPUs are expensive. Inference is expensive. And here’s the part that keeps founders up at night: as models get better, costs don’t go down — they go up.
FAL saw this firsthand. Image models from a year ago are now cheap to run. But customers don’t want to run those. They want the new video models, which are dramatically more expensive. So even as the old cost curve came down, the new cost curve went up. Net result: lower margins than before.
The hope was that better hardware would bail everyone out. But software advances faster than hardware. New model capabilities outrun chip improvements. The best model you can run today is meaningfully more expensive than the best model you could run twelve months ago. That gap is not closing.
What does this mean practically? You cannot assume your margins will improve just by waiting. You need to build pricing that reflects real cost to serve from day one.
Your Best Customers Can Become Your Worst
This is the uncomfortable math nobody talks about at dinner.
In traditional SaaS, your highest-usage customers were your best customers. More usage meant more value delivered, more expansion, more NRR. Usage and margin worked in the same direction.
In AI, they can work against each other. High-usage customers are also high-cost customers. They’re consuming inference at scale. And if your pricing doesn’t account for that, you’re subsidizing your biggest accounts.
FAL’s answer was to track wallet share — the share of a customer’s total generative media spend that flows through the platform. That framing changes how you think about expansion. It’s not just about growing accounts. It’s about growing the right accounts in the right way.
Annual Quotas Are Useless When You’re Growing 50% Per Quarter
FAL tried to hire a head of sales. The candidate, reasonably, wanted to negotiate a quota and commission structure. The team sat down to figure out what doubling in a year would look like as an OTE target. Then, during the interview process itself, they grew roughly 50% of that annual target.
The quota became meaningless before the hire was made.
At growth rates like this, annual targets are guesswork. FAL’s current thinking: shorter-term quotas, monthly or quarterly, where you can course-correct when the business moves. For now, the team runs on target earnings without hard quotas attached. It is not a permanent solution. It is an honest response to an unpredictable environment.
Hire Researchers With a Grants Program
FAL’s research team is small and strong. Most of the people on it didn’t come through a traditional recruiting process.
The model: open invitation, anyone can email with a research project idea. FAL cares about efficient AI, specifically efficient finetuning techniques and efficient inference. If the idea is good, they give you compute for a couple of weeks to build something. No strings attached. No expectations.
They hired four people this way.
The insight is obvious once you see it: if you want to find researchers who can do the work, the best signal is watching them actually do the work. Resumes and interviews can’t replicate that. A research grants program essentially turns your recruiting process into a paid audition that also advances your technical roadmap.
Positioning as a Category Is a Real Moat
When FAL started, the conventional wisdom was that all AI models would end up in the same market, bought by the same buyers, served by the same companies. FAL looked at who was actually buying image models versus language models and saw something different: the buyers are completely distinct, the use cases are completely distinct, and therefore the market is completely distinct.
They positioned the company as a “generative media platform” and committed to it. It shaped marketing, shaped recruiting, shaped which large customers they went after.
That positioning is now working. The brand is becoming synonymous with generative media specifically. In a market where everyone is claiming to do everything, being the clearest answer to one specific thing is worth more than most founders realize until they’ve tried the alternative.
The Metrics That Actually Matter for an AI Infrastructure Business
FAL watches three things closely:
- Logo quality and diversity. They want big, recognizable names on the platform. But they also want no single company representing too much of revenue. The target is at least 30-35 paying companies before anything looks like concentration risk.
- Churn. Not as an annual number but as a constant operational priority. If a customer is spending on generative media somewhere else, FAL wants to understand why and bring that spend back.
- Wallet share. What percentage of a customer’s total generative media budget is flowing through FAL? Expansion is not just adding new logos. It’s deepening what the existing logos spend.
One interesting observation from the data: AI-native companies are often spending more aggressively than large enterprises. Bigger companies are cautious about putting AI into production at scale. Newer companies are building on it from day one. If you are building AI infrastructure, the AI-native segment is often your best near-term growth lever.
The playbook Gorkem described is not about finding shortcuts. It is about being honest that this is a different business than what came before, with different cost structures, different hiring approaches, different metrics. The founders who are winning right now treated that as a starting point, not a problem to explain away.
And for 200+ convos like this, and 100s more workshops and 1-on-1s, come to 2026 SaaStr AI Annual! May 12-14 in SF Bay!!
Grab tix here!
