Since the emergence of Chatgpt at the end of 2022, the technological ecosystem has seen the birth of a multitude of startups positioned on generative artificial intelligence. However, the majority of them collide with the same structural weaknesses. During the event Ai Ascentthe partners of Sequoia Capital shared a return of experience on the most common errors observed in the fund portfolio. Three points of vigilance stand out with false turnover, deceptive margins and Data Flywheels without real effect.
Turnover is not always a sign of traction
The appearance of an income line does not necessarily mean market membership. Pat Grady, partner at Sequoia, alerts to what he calls the “Vibe returned”a turnover based on the fashion effect or technological curiosity, without proof of lasting use.
Several founders struggle to distinguish a punctual essay from a lasting adoption. The distinction between simple exploration and real transformation of uses takes place by analyzing the commitment, recurrence, and the evolution of the use of the product. In the absence of this reading, some founders overestimate their traction, which can cause errors in the dimensioning of the team or positioning on the market.
The raw margins of today are not an end in itself
One of the paradoxes of the current AI is in the cost structure. While inference costs drop as the models are campaigning, the cost by Token has dropped 99 % in 18 months, de facto margins remain unstable. In question: an dependence on owners’ aptors, costs linked to real -time use, and a difficulty in changing the value proposition towards deliverables with high perceived value.
It is not so much a question of judging the current gross margin as the trajectory towards a Credible “Pricing Power”. The transition from a tool to a solution, then from a solution to a result, allows certain players to get out of the logics of “feature” to capture a more strategic budget, often in the operational cost lines rather than informatics.
The myth of the Data Flywheel
The argument is often advanced: the more users interact with the product, the more it improves. This principle of positive feedback loop, or Data Flywheel, is often invoked as a differentiation lever.
But it is still necessary that the data collected is really used to modify a metric business. In many cases, the loop remains theoretical: the data is stored, analyzed, but do not alter the performance of the model or the user experience. If this loop does not modify any strategic data (conversion rate, acquisition cost, execution time) then it has no economic value.
