AI is set to play an even larger role across industries in the coming years, driving innovation and unlocking new possibilities. Four key trends to watch include:

  • The rise of small language models (SLMs) and multimodal systems
  • Expanding applications of AI in healthcare and life sciences
  • Significant advances in AI reasoning capabilities
  • The global development of custom AI chips by leading technology companies

The rapid advancement of AI is already reshaping industries at an unprecedented pace. AI-driven solutions are being deployed for data analysis, decision support, predictive modeling, and other high-impact use cases.

As enterprises move forward, a critical question remains: what should they anticipate in the next phase of AI evolution?

Emerging AI Models: Small Language Models (SLMs) and Multimodal AI

As artificial intelligence evolves, organizations are exploring alternatives to large, general-purpose models. Two notable directions are Small Language Models (SLMs) and multimodal AI, which emphasize efficiency, adaptability, and closer alignment with specific business needs.

Small Language Models (SLMs)

Small Language Models (SLMs) are optimized for specific tasks, delivering focused performance at a lower cost. Trained on domain-specific data, these models help businesses address particular challenges more efficiently.

For example, SLMs can generate actionable insights for inventory management that would otherwise take weeks to compile manually. Because they can run on local devices, SLMs also enhance data privacy and reduce infrastructure requirements, making them an attractive option for enterprises.

In practice, more than 75% of organizations are now customizing small open-source models to meet specific requirements. This approach offers a faster, cheaper, and more manageable alternative to relying solely on larger models.

Multimodal AI

Multimodal AI refers to systems that process and integrate multiple types of data—such as text, images, audio, video, and code—within a single model. Unlike traditional AI, which typically handles only one data type, multimodal models can reason across diverse inputs to produce more nuanced outputs.

This capability makes multimodal AI more context-aware, enabling richer, more human-like interactions. It also supports greater creativity and productivity in areas such as software development, scientific research, and other complex knowledge work.

As demand for sophisticated, context-sensitive applications grows, McKinsey predicts that multimodal AI will expand significantly over the next 18–24 months, reshaping industries with more intelligent and adaptable systems.

Unsure whether a Large Language Model (LLM) or a Small Language Model (SLM) is the better fit for your business? Explore a detailed comparison in our article: Breaking Down AI: The Comparative Edge of Language Model Sizes.

More AI advancements in health & life sciences

AI is transforming how medical and life science professionals analyze large-scale datasets to uncover new insights. Its ability to process and interpret complex information has made AI an essential driver in accelerating disease discovery and advancing drug development.

AlphaFold: accelerating protein structure prediction

In 2025, one of the most influential AI applications in life sciences is AlphaFold, an AI system developed by DeepMind. For decades, scientists struggled to determine the 3D structures of proteins from their amino acid sequences—a critical step to understanding how proteins function in the body. Traditional experimental methods were slow and resource-intensive, often taking months or even years to solve the structure of a single protein.

To address this bottleneck, Demis Hassabis and John Jumper of Google DeepMind developed AlphaFold to predict the three-dimensional structure of proteins directly from their amino acid sequences with remarkable accuracy. In the 14th Critical Assessment of Protein Structure Prediction (CASP14), AlphaFold outperformed all other approaches, setting a new benchmark for the field.

This level of accuracy enables researchers to model protein structures reliably, significantly accelerating studies in molecular biology and drug discovery. DeepMind’s open release of AlphaFold’s source code and a database of predicted structures for more than 200 million proteins—including the entire human proteome—has given millions of scientists worldwide access to high-quality structural data and vastly expanded the coverage of known protein structures. In recognition of this breakthrough, the two researchers were awarded the 2024 Nobel Prize in Chemistry.

AI-guided discovery of treatment for a rare disease

Another notable breakthrough in 2025 is an AI tool developed by researchers at the University of Pennsylvania, which helped identify adalimumab—an FDA-approved monoclonal antibody—as the leading treatment candidate for idiopathic multicentric Castleman’s disease (iMCD). iMCD is associated with cytokine storms, in which the immune system releases excessive inflammatory proteins that damage tissues and organs, causing widespread inflammation and potentially life-threatening organ failure.

By applying machine learning to analyze 4,000 existing drugs, the research team found that adalimumab could effectively counteract elevated tumor necrosis factor (TNF) signaling in patients with severe iMCD. This discovery opened the door to a potential life-saving therapy. One patient, who was preparing to enter hospice care after several failed treatments, received adalimumab and has now remained in remission for nearly two years.

Together, these recent advancements have laid a strong foundation for the future of AI applications in health and life sciences, particularly in developing treatments for diseases that are currently difficult to manage. To see how AI is already reshaping clinical workflows, explore how FPT empowered patient-centric care with the AI solution AIScribe in our latest case study.

Breakthroughs in AI logical reasoning architectures

As business challenges grow increasingly complex, simply retrieving information or generating content is no longer sufficient. AI systems must be able to pause, evaluate, and make decisions in real time. Traditional pre-trained models focus on training-time computing, predicting outcomes based on vast datasets they have previously seen. While this approach is effective for straightforward or well-defined tasks, it often fails when applied to more intricate, ambiguous, or high-stakes problems.

This is where AI reasoning becomes essential. Instead of merely restating learned patterns or extrapolating from historical data, AI needs to engage in inference-time computing—actively evaluating different scenarios, weighing potential outcomes, and making decisions grounded in logic. Although this process demands more computation and effort, it produces results that are significantly more meaningful, context-aware, and impactful. By enabling AI to effectively "pause and think," these systems can move beyond surface-level responses to address complex challenges that drive real business value and innovation.

A concrete example is OpenAI’s newly released o1 model, which introduces chain-of-thought (CoT) reasoning with significant implications for improving AI accuracy and safety. This architecture enables the model to perform real-time double-checking throughout the decision-making process, helping to reduce risks associated with AI hallucinations, biases, and harmful content generation. By embedding CoT reasoning into the core operations of o1, each step of the model’s reasoning can be reviewed for ethical and safety concerns as it processes information.

AI trends powered by AI chips

AI chips in health and life sciences and new model architectures

As AI continues to evolve, its progress increasingly depends on a critical foundation: AI chips. In the health and life sciences industry, AI chips enable the high-throughput processing required for genomic data, medical imaging, and real-time diagnostics. Accelerators such as GPUs and TPUs help drive breakthroughs in drug discovery and personalized medicine by supporting large, complex models and massive datasets. At the same time, custom AI chips embedded in wearables and medical devices support continuous monitoring and on-device inference.

In parallel, small language models (SLMs) are designed to operate with limited computational resources, making them well suited to low-power edge devices where energy efficiency and rapid response times are essential. Custom AI chips make this possible by optimizing for fast inference and low latency while consuming minimal power. By contrast, multimodal models require AI chips with high memory bandwidth and extensive parallel processing capabilities. These chips can manage simultaneous data streams and intensive computations to deliver real-time, context-aware outputs, which are crucial for seamless multimodal experiences.

Custom AI chips and the new semiconductor race

The rise of AI chips signals a new era in semiconductor innovation and intensifies competition among technology leaders such as Meta, Google, and Intel. These companies are developing custom chips to meet rapidly growing AI demands, motivated largely by the need to reduce the high costs associated with relying on general-purpose processors and external suppliers. Custom AI chips offer businesses enhanced operational efficiency, faster processing speeds, and the flexibility to tailor hardware to specific AI applications.

Owning proprietary chip technology also allows companies to exert greater control over their infrastructure, improve data privacy, and secure competitive advantages by aligning hardware design with their strategic priorities. Meta, for example, has begun testing its in-house AI training chip as part of a broader strategy to cut infrastructure costs and strengthen AI capabilities, with plans to expand usage by 2026. The latest Meta Training and Inference Accelerator (MTIA) chip significantly improves compute performance and memory bandwidth over its predecessor, enabling support for complex AI models that power recommendation systems and generative AI products.

Vietnam: An emerging AI hub

Vietnam is emerging as a significant hub for AI chip innovation with the launch of the FPT AI Factory, equipped with thousands of NVIDIA H100 GPUs, considered among the world's most advanced AI superchips. Scheduled to begin services in January 2025, this facility will provide massive computing power capable of billions of calculations per second, accelerating AI model training and optimization by up to 1,000 times. The factory integrates NVIDIA's AI Enterprise software stack and supports a comprehensive ecosystem of generative AI technologies, allowing businesses to rapidly build intelligent AI applications and enhance creativity.

NVIDIA founder Jensen Huang has emphasized the importance of local AI development for Vietnam's industries and citizens, underscoring the factory's role in advancing sovereign AI capabilities. Through partnerships with global industry leaders, FPT AI Factory will offer early access programs, cloud credits, and expert support that together can accelerate AI adoption across the region.

The Next Big Things with AI

In the coming years, the impact of AI will continue to expand, opening up greater opportunities across virtually every industry. To fully unlock the potential of this cutting-edge technology, leading nations, including Vietnam, are prioritizing AI education for domestic talent, both in universities and within the workplace.

As AI keeps transforming business models and operations, navigating this rapidly evolving landscape can be challenging. Therefore, enterprises need a trusted partner to help ensure the safe and ethical use of AI while driving successful digital transformation. FPT can play that role by accompanying organizations on their AI journey and supporting them in adopting responsible, high-impact AI solutions.

Find more about FPT’s AI service offerings here: https://fpt-aicenter.com/en.

Frequently Asked Questions

What are the most important AI trends businesses should watch in the next few years? Four trends stand out: smaller and multimodal AI models, rapid AI adoption in healthcare and life sciences, major improvements in AI reasoning, and custom AI chips. Together they enable faster, cheaper, and more capable AI, reshaping data analysis, decision-making, and innovation across industries.

What are small language models and multimodal AI, and when should a business use them? Small language models are compact, task-focused models that run cheaply, often on local or edge devices. Multimodal models combine text, images, audio, or video to give richer, context-aware results. Businesses use SLMs for efficient, private, domain tasks and multimodal AI for complex, real-world interactions.

How is AI transforming healthcare and life sciences in practice today? AI now analyzes huge clinical and molecular datasets to uncover patterns, speed disease discovery, and guide drug repurposing or design. AlphaFold’s protein structure predictions and AI tools that scan thousands of drugs for new uses are shortening research cycles and enabling treatments for hard-to-treat diseases.

What do new AI reasoning architectures change compared with traditional pattern-matching models? New reasoning architectures let AI pause and evaluate options at inference time, not just replay patterns from training. Techniques like chain-of-thought guide stepwise reasoning and self-checking, reducing hallucinations and bias. This makes AI more reliable for complex decisions, risk-sensitive tasks, and high-value workflows.

Why are AI chips like GPUs, TPUs, and custom accelerators so critical to future AI applications? AI chips deliver the compute, memory bandwidth, and energy efficiency needed for training and running modern models. GPUs, TPUs, and custom accelerators power genomics, medical imaging, SLMs on edge devices, and multimodal systems. They lower cost, increase speed, and let firms tailor hardware to strategic AI workloads.