The founder and Chief Executive Officer of Alluxio, Haoyuan (HY) Li, has shared his predictions for technology developments in 2025.
According to Li, multi-modal training is expected to become a more mainstream approach in 2025, facilitating the integration of diverse data types such as text, images, audio, and video in model training. This approach is anticipated to enhance the capacity of AI systems to understand and process complex real-world data, thus fostering more context-aware applications. For instance, in autonomous driving, the requirement to understand visual, auditory, and textual data is critical. The increased use of multi-modal models is also likely to drive the demand for advanced hardware and storage solutions due to the growing complexity of training environments.
Li predicts that pre-training will become a critical differentiator for organisations developing large language models (LLMs) by 2025. The evolution of the AI landscape will see access to large datasets, particularly industry-specific ones, become a major competitive advantage. Companies using extensive data infrastructure for leveraging these large-scale datasets will likely succeed in fine-tuning models to deliver specialised solutions. However, challenges will arise in managing data preparation, cleaning, and transformation, which will become key factors for success in developing robust and relevant LLMs.
Addressing data access challenges is expected to become increasingly important for AI success as workloads grow more demanding and distributed. The proliferation of data across multiple clouds and storage systems presents bottlenecks in data availability and movement, especially for AI training. Organisations will need to efficiently manage data access across distributed environments to minimise data movement and duplication. The ability to provide rapid, concurrent access to data while maintaining its locality will distinguish successful organisations in scaling AI initiatives.
The evolution of AI-driven cloud economics will reshape infrastructure decisions by shifting focus from traditional cloud cost optimisation to AI-specific ROI optimisation. Organisations will elaborate sophisticated models to predict AI workload costs across various infrastructure options, leading to nuanced hybrid deployment strategies. This involves balancing cost-performance trade-offs for training and inference workloads between cloud and on-premises resources.
In response to the exponential growth of AI model training datasets, Li foresees that maximising GPU utilisation will become a focal point in modern datacenters in 2025. The pressure to optimise costly GPU infrastructure investment is expected to drive advancements in hardware and software design to ensure sustained massive read bandwidths necessary for training, while minimising checkpoint-saving times. Successful organisations will be those that keep GPU resources continually productive while managing larger model checkpoints and increasing data requirements.
The evolution from traditional MLOps to comprehensive AIOps platforms is anticipated, managing the whole lifecycle of AI systems. These platforms are expected to integrate advanced monitoring and automation capabilities for models and infrastructure, facilitating predictive maintenance and AI system optimisation. Organisations are likely to adopt practices that consider AI models as dynamic systems with continuous learning and adaptation embedded within deployment pipelines. New tools and practices for version control, testing, and deployment are expected in order to handle the complex nature of multi-modal models and distributed training settings.
"2025 will be a pivotal year for AI infrastructure innovation, with multi-modal AI becoming mainstream, demand for compute and data infrastructure will continue to soar, redefining how we train and deploy models, and transforming how organisations manage the entire AI lifecycle," said Haoyuan (HY) Li, Founder and CEO of Alluxio. "As data access challenges grow and AI-specific cloud economics take centre stage, businesses that embrace these innovations will lead the way in driving AI-driven innovation and efficiency."
