With the risk of a speculative bubble? “The valuation of OpenAI is a little off the ground, but it can be explained both by what has been achieved and by the hope of future value creationsays Olivier Martret, associated with the Serena Capital fund. Its revenues have seen phenomenal growth, from $500 million in 2022 to an expected $4 billion this year. Whereas it started with a consumer product sold for $20 per month and is only beginning to pivot towards a platform of models dedicated to businesses, which will be the most profitable.
Investments not profitable before 15 years
If these start-ups need so much money, it is because the investments to be made are colossal. “There are barriers to entry for creating foundational modelsrecalls Olivier Martret. This requires hundreds of billions of dollars of investment, particularly in computing power which must be reserved one to two years in advance. Without forgetting the salaries to attract talent, when only 300 to 400 people in the world are capable of creating such a model.
In its AI Index Report 2024, released in April, Stanford University noted that “the costs of developing cutting-edge AI models have reached unprecedented levels.” According to the American university, $78 million was spent on training OpenAI’s GPT-4, while Google’s Gemini Ultra cost $191 million. The staggering costs affect the entire value chain, and in particular cloud providers, forced to boost data center capacity. AWS (Amazon) announced $150 billion in investment in AI infrastructure over the next fifteen years. Microsoft is also increasing its investment announcements for AI and the cloud. Amounts spent now, but which will not be profitable for fifteen years, according to the own admission of Amy Hood, its financial director.
New business models
It is this gap that worries and raises fears of overinvestment, as during the Internet bubble in the 2000s. “Technology giants are expected to invest more than 1,000 billion dollars in AI in the years to come, with so far few results to show. Will these significant expenses one day pay off?”, asks the bank Goldman Sachs in a note published in June. This is “the $600 billion question”, estimates David Cahn, associated with the Sequoia Partners fund, on his blog. According to him, it is this amount that the annual revenues generated by AI should reach so that all players in the ecosystem succeed in making their investments profitable.
After the euphoria aroused by its magical side, the time has now come for the monetization of generative AI. “After a price war to ensure mass adoption, we are entering a phase where we must demonstrate value creation”notes Michael Mansard, director of strategy at Zuora, a company specializing in monetization. If the handful of start-ups developing foundational models and cloud providers should remain in a race for volumes, he observes the emergence of a new business model among AI software publishers dedicated to a business problem : selling based on results, and no longer on the user.
Being able to measure ROI
Like Intercom, which sells its AI dedicated to customer service based on the number of incidents resolved by the software. “This allows them to evaluate the value created and take into account the specificities of AI, the costs of which increase proportionally to the number of users, since the more requests I have, the more I pay for calculation”analysis Michel Mansard.
LightOn, a Parisian start-up which develops small LLMs specialized in use cases, confirms this new concern for value creation. “With our clients, we only choose use cases for which we can have metrics and a return on investment”confides Laurent Daudet, its founder. The nugget managed to be profitable in 2023, in particular thanks to the sale of LLMs tailor-made for client companies.
She recognizes “especially target major accounts in sectors where information systems are complex and where a lot of data is available”. A reminder that the greatest returns on investment require certain digital maturity. A transformation that will take a few more years.
