Haoyuan Li, Founder and CEO, Alluxio, offers his tech predictions for the coming year.
Multi-Modal training will become more mainstream: In 2025, multi-modal training, which integrates different types of data—such as text, images, audio, and video—will become a more dominant approach in model training. This shift is driven by the need for AI systems to better understand and process the complexity of real-world data, allowing for richer and more context-aware applications. For example, multi-modal models can improve use cases like autonomous driving, where understanding visual, auditory, and textual information is critical. The rise of these models will also spur demand for more advanced hardware and storage solutions, as the complexity of training environments continues to grow.
Pre-Training Will Become a Key Differentiator for Organizations Adopting LLMs: By 2025, pre-training will emerge as a crucial differentiator among organizations developing large language models (LLMs). As the AI landscape evolves, access to vast amounts of high-quality data – especially industry-specific data – will become a major competitive advantage. Companies that can effectively harness big data infrastructure to leverage their large-scale datasets will be better positioned to fine-tune their models and deliver more effective, specialized solutions. However, this also introduces a significant bottleneck. Preparing and curating the right data for pre-training is increasingly complex and companies without robust big data infrastructure will struggle to keep up. Efficiently handling this data preparation, cleaning, and transformation process will become a critical challenge in the race to develop more powerful and relevant LLMs.
Overcoming Data Access Challenges Becomes Critical for AI Success: In 2025, organizations will face increasing pressure to solve data access challenges as AI workloads become more demanding and distributed. The explosion of data across multiple clouds, regions and storage systems has created significant bottlenecks in data availability and movement, particularly for compute-intensive AI training. Organizations will need to efficiently manage data access across their distributed environments while minimizing data movement and duplication. We’ll see an increased focus on technologies that can provide fast, concurrent access to data regardless of its location while maintaining data locality for performance. The ability to overcome these data access challenges will become a key differentiator for organizations scaling their AI initiatives.
AI-Driven Cloud Economics Reshape Infrastructure Decisions: In 2025, organizations will fundamentally reshape their cloud strategies around AI economics. The focus will shift from traditional cloud cost optimization to AI-specific ROI optimization. Organizations will develop sophisticated modeling capabilities to understand and predict AI workload costs across different infrastructure options. This will lead to more nuanced hybrid deployment strategies where companies carefully balance the cost-performance trade-offs of training and inference workloads across cloud providers and on-premises infrastructure.
Maximizing GPU utilization becomes the new standard: In 2025, as the size of AI model training datasets continue to grow exponentially, maximizing GPU utilization will become the primary design goal for modern data centers. Organizations will face mounting pressure to optimize their expensive GPU infrastructure investments. This shift will drive innovations in hardware and software design to sustain the massive read bandwidths necessary for training and minimize checkpoint-saving times that cause training pauses. Success will be measured by how effectively data centers can keep their GPU resources busy while managing larger model checkpoints and growing data requirements.
MLOps Evolution to AIOps: In 2025, we’ll see the evolution from traditional MLOps to comprehensive AIOps platforms that manage the entire AI system lifecycle. These platforms will integrate sophisticated monitoring and automation capabilities for both models and infrastructure, enabling predictive maintenance and automatic optimization of AI systems. Teams will adopt practices that treat AI models as living systems rather than static deployments, with continuous learning and adaptation capabilities built into the deployment pipeline. This shift will require new tools and practices for version control, testing and deployment that can handle the complexity of multi-modal models and distributed training environments.
