Exploring opportunities in the generative AI value chain – McKinsey

Ajouter à la liste
ClosePlease login

No account yet? Register

Synthèse de Beesnest

Les 6 éléments de la chaîne de valeur de l'Intelligence Artificielle

Alors que l’IA générative est en train de faire naitre un nouvel écosystème, et de nombreuses nouvelles opportunités business, McKinsey revient en détail sur les éléments composants sa chaîne de valeur, avec des analyses inédites. Les systèmes d’IA générative s’accompagnent d’une nouvelle chaîne de valeur, composée six catégories : matériel informatique (1), plateformes cloud (2), modèles fondamentaux (3), hubs de modèles et opérations d’apprentissage automatique (4), applications (5), services (6). 

Les bases des systèmes d’IA générative sont nettement plus complexes que la plupart des systèmes d’IA traditionnels. En conséquence, le temps, le coût et l’expertise nécessaires à leur mise en œuvre représentent des obstacles importants pour les nouveaux entrants et les petites entreprises tout au long de la chaîne de valeur. 

1- Computer hardware

  • To generate content, AI systems like OpenAI’s GPT-3 need a significant volume of knowledge, which it gains from extensive text data. This model was trained using approximately 45 terabytes of text data.
  • Traditional computer hardware cannot handle the computational load required by such generative AI systems. To process such large data volumes, you need large clusters of specialized hardware like GPUs or TPUs that have accelerator chips for parallel data processing.
  • Once the primary AI model is trained, businesses might use these hardware clusters for customization and running of these energy-intensive models. Although these subsequent stages need less computational power than the initial training stage.
  • For new entrants, the start-up costs associated with research and development in this field are high. To serve the generative AI market, traditional hardware designers need to build up specialized skills, knowledge, and computational capabilities.

2- Cloud platforms

  • A lot of the work related to building, tuning, and running large AI models takes place in the cloud. This allows companies to have easy access to the required computational power and manage their expenditure as needed.
  • Major cloud service providers hold the most comprehensive platforms for executing generative AI tasks and have preferential access to the necessary hardware and chips.

3- Fundation models

  • Central to generative AI are foundation models, large deep learning models pretrained to generate specific types of content but adaptable for various tasks. Once a foundation model is established, it can be leveraged by anyone to build applications on top of it.
  • The development of foundation models involves deep expertise in several areas such as data preparation, model architecture selection, model training, and tuning to enhance the output.
  • The current market for foundation models is dominated by tech giants and well-funded startups due to the high costs involved. However, there are ongoing efforts to create smaller, more efficient models that could broaden market accessibility. 

4- Model Hubs

  • Businesses looking to build applications on top of foundation models need storage and access to these models and specialized MLOps tools, technologies, and practices for adapting and deploying these models. 
  • Model hubs fulfill these needs, offering services varying based on whether the models are closed-source or open-source. Closed-source model developers typically act as model hubs, providing access and sometimes MLOps capabilities for customization and deployment. In contrast, for open-source models, independent hubs like Hugging Face and Amazon Web Services provide a range of services, including access to various models and end-to-end MLOps capabilities, fulfilling a growing need for companies that lack the in-house talent or infrastructure.

5- Applications

  • Applications built on foundation models enable completion of specific tasks and come in two types: those with minor customizations to foundation models, and those with fine-tuned foundation models for specific use case outputs. The latter, requiring less data and cost, are more accessible and represent the most attractive part of the value chain.
  • Application developers can amass data from industry knowledge, daily business operations, or end-user rating systems. This continuous feedback loop can create a virtuous cycle of improvement, resulting in significant competitive advantages.

6- Services

  • As the generative AI landscape evolves, specialized services are anticipated to help companies navigate technical complexities and business opportunities. Existing AI service providers will likely adapt to the generative AI market, and niche players may emerge with expertise in applying generative AI to specific functions, industries, or capabilities.

Veuillez noter qu’il ne s’agit pas d’une liste exhaustive de toutes les informations contenues dans le rapport, mais plutôt d’un résumé de certains points et chiffres clés. Pour plus d’informations, veuillez lire le rapport complet.

Informations sur l'étude

Notation de l'étude

Les informations contenues dans la synthèse d’études et de rapports produite par Beesnest sont fournies à titre informatif uniquement et ne constituent pas des conseils professionnels. Les études et les rapports utilisés pour produire cette synthèse proviennent d’autres organismes et n’ont pas été rédigés par Beesnest. L’exactitude, l’exhaustivité ou la pertinence de ces informations ne peut pas être garanties par Beesnest qui n’est pas responsable des erreurs, omissions ou imprécisions dans cette synthèse. En utilisant cette synthèse, vous reconnaissez que vous le faites à vos propres risques et que vous êtes responsable de la prise de décision qui en découle. Beesnest n’est pas responsable de tout préjudice ou dommage résultant de l’utilisation de ces informations.

Laisser un commentaire