The AI industry may still be obsessed with which model is smartest, fastest or most powerful. But for Basil Fateen, head of startups and venture capital at Nvidia for the Middle East, Turkey and Africa, that is increasingly the wrong question.
Speaking at Communicate AI in Dubai on September 29, Fateen offered a more practical view of where businesses should focus as generative AI moves from experimentation into production: understand the cost of every AI interaction, build proprietary advantages around data and expertise, and create systems that become better as people use them.
His starting point was unusually blunt.
“I believe that makes me uniquely qualified to inform you that without a shadow of a doubt, nobody knows exactly what the hell is going on right now. Guys. Nobody. Not even me,” Fateen told the audience.
That uncertainty, he argued, is not a reason to wait. It is a reason to experiment.
“We have no frame of reference for this current phase, which means that there are no gurus and there are no playbooks to follow because the old playbooks simply don’t work anymore,” he said. “There’s only two things. There is experiments and there’s opinions.”
Fateen’s presentation, titled “Signals @ Communicate AI”, distilled his observations into three areas: tokenomics, moats and flywheels.
The new AI unit economics
His first signal was perhaps the most relevant for businesses selling AI-powered products and services: understand the economics of tokens.
“Tokens are generative AI unit economics,” Fateen said, describing them as the “base currency” of generative AI. Every prompt, response and agent action consumes tokens, making token consumption directly relevant to the cost of delivering an AI-powered service.
The presentation breaks tokenomics into six components: cost per token, tokens per task, cost per task completion, model routing and prompt optimisation, outliers and outcome-based pricing.
His warning was simple: businesses can no longer say they are “all in on AI” without defining what that means financially.
“The people who are utilizing you, they don’t care about the technology under the hood,” he said. “They care about accomplishing a certain task or solving a problem.”
That means businesses should examine how many tokens are actually needed to complete a task, rather than automatically defaulting to the most expensive frontier model. Fateen argued that companies should compare models on three variables — quality, cost and speed — and use routing systems to determine which model is appropriate for a particular task.
He also pointed to a paradox: as token costs fall, usage can rise sharply. Fateen said token costs had fallen by almost 300%, while usage had increased by roughly 300%, making optimisation increasingly important.
That optimisation can ultimately change the business model itself. Fateen cited outcome-based pricing, pointing to AI company Sierra, which he said had moved from pricing around usage to charging for measurable outcomes.
When the model stops being the moat
The second signal was the changing nature of competitive advantage.
Before generative AI, software development itself could constitute a moat. In creative industries, the ability to produce copy, imagery or other creative work could also provide a barrier to entry. Generative AI has lowered many of those barriers.
“So now in this world we’re living in, everyone needs to think about what is the new moat that gives what I’m doing value,” Fateen said.
His answer starts with proprietary data.
He cited Tajima AI, an Nvidia Inception startup, as an example. The company had spent 15 years as a translation business before becoming an AI company, giving it proprietary Middle Eastern translation data across areas including legal, HR and marketing. Fateen said that historical data and domain knowledge became a competitive advantage when the company began building AI models.
But data is only one layer. Fateen also highlighted domain expertise, evaluations, user experience, humans in the loop and orchestration as emerging moats — all of which are listed in his presentation.
The human component, he argued, is particularly important.
“It’s not about utilizing humans, it’s where they are in the loop to make sure that the solution or service you give that’s powered by AI is consistent, secure and scalable,” he said.
AI products that learn from use
Fateen’s third signal was the flywheel: systems that improve as they are used.
Exceptional AI companies, he argued, do not simply launch a product and wait for customers. They capture behavioural signals, user feedback and other forms of data that can continuously improve the product. His presentation identifies behavioural signals, lightweight prompts, power users, closed-loop feedback, synthetic data and communities as the building blocks.
“I’m still shocked at how many people and how many organizations do not log and track all of this proprietary usage data coming from their solutions,” Fateen said.
But the final flywheel is not technological.
“As technology gets more and more advanced and we’re able to automate more and more intelligent things, the value of human relationships and real communities and communication increases because you can’t automate that,” he said.
That idea ultimately brought Fateen back to his broader message: AI’s technological capabilities are advancing faster than adoption, and businesses still have to earn confidence.
“The key in this phase and the real value lies in the trust layer, not in technology,” he said.
For Nvidia, that ecosystem also extends into startup support. Its Inception programme offers startups access to more than 1,000 prebuilt SDKs, models and containers, courses, preferred pricing, cloud credits, investor connections and other benefits. The presentation says the programme covers more than 120 countries, with no fees or equity requirements.
Fateen’s final message was therefore less about predicting what AI will become than about how businesses should behave while the answer remains uncertain: experiment, measure, learn, build proprietary advantages and keep humans in the system.
“The main skills you need are not technical proficiency,” he said. “They are being humble, being scientific in your experiments, not listening to hype and sharing your insights and being as collaborative as possible.”
