AI and Profitability: A Crucial Challenge for 2024

AI Faces a Major Profitability Challenge

Companies are confronted with a significant obstacle: the profitability of artificial intelligence. As costs associated with tokenization soar and many projects struggle to demonstrate their added value, it becomes imperative to shift from chaotic adoption to more thoughtful budget management.

Financial departments are sounding the alarm. Budgets allocated to AI, often spent without tangible impact on profits, raise serious concerns. The idea of a miraculous, cost-free technology collapses in the face of infrastructure spending realities. To overcome this "profitability wall," it is crucial to rethink current consumption models.

From Initial Enthusiasm to the Reality of Return on Investment

Warning signs are multiplying in the sector:

• Microsoft had to lower some of its internal licenses.
• Uber consumed its annual AI budget in just four months.
• At Nvidia, the Vice President of Deep Learning acknowledged that computing costs exceeded those of his engineering team.

According to a 2024 study by the RAND Corporation, over 80% of AI projects fail to deliver significant business value. A striking example is that of a company where a single employee exhausted the budget allocated for ChatGPT Enterprise for the entire department, leading to an immediate restriction of access. Naive adoption of AI gives way to necessary optimization. Without a structured framework, widespread use of AI can lead to budgetary chaos. Results are only compelling when AI is integrated iteratively with frequent feedback.

Licenses, Usage, or Private Infrastructures: A Budgetary Dilemma

Large companies like OpenAI, Google, and Mistral primarily offer two business models: licensing and usage. Licenses, often limited, encourage their multiplication, while uncontrolled usage can generate unpredictable costs.

An alternative is to deploy one's own GPU infrastructure to run Open Source models. This allows for better control over costs and security but involves complex management of memory and context engineering. It is a job in itself.

A hybrid approach is recommended: for token-intensive tasks, internalizing on private GPUs is ideal. Tokenization imposes budgetary discipline. Using a cutting-edge model for simple tasks, like sorting an Excel file, is akin to handing a Ferrari to a novice driver: the risk of financial skid is high. The future lies in agnostic multi-model orchestration, oriented towards more economical models. A well-thought-out hybrid architecture could significantly reduce inference costs.

Frugal Innovation and Collaboration to Catch Up

Europe, lacking sovereign Cloud infrastructure, is at a disadvantage compared to the United States and China in terms of infrastructure. However, it possesses talent and ideas. At the Nexus conference in Luxembourg, local players showcased "full GPU" solutions offering supercomputers accessible to businesses, unlike in France where these resources are reserved for research. Europe is finally beginning to invest.

Frugal innovation is the card to play: it is about doing better with less, focusing on algorithmic optimization. To achieve this, it is necessary to break a French taboo: the separation between "noble" academic research and the lucrative professional world. In the United States and China, researchers and industry collaborate closely. Bridging these two worlds is essential to transform our scientific excellence into economic profitability.

The era of open bars is over. Salvation will come from strict governance and hybrid architectures. By betting on sobriety and uniting research and industry, France and Europe can turn their lag into strength. The AI of tomorrow will not be the heaviest, but the most efficient.