Andrew Ng Warns of AI Training Layer Bubble; Agentic Systems to Drive Near-Term Value

The race to build better AI models has evolved into a fierce competition for compute resources, talent, and control. Foundation models-large-scale systems trained on vast datasets-now form the backbone of everything from enterprise software and cloud infrastructure to national digital strategies. However, despite the ambitious language surrounding these advancements, the economic returns remain uneven.

The Economic Reality of AI Investment

According to PwC’s 2026 Global CEO Survey, which polled 4,454 CEOs across 95 countries, only 12% reported both increased revenue and reduced costs from AI investments over the past 12 months. A staggering 56% saw no improvements in either metric. Despite this, 51% of CEOs plan to continue investing in AI, albeit with declining confidence in revenue growth. This discrepancy highlights a growing gap between engineering reality, commercial hype, and public expectations.

Andrew Ng’s Perspective on AGI and Agentic Systems

Few voices carry as much authority in the AI community as Andrew Ng. As the founder of DeepLearning.AI, executive chairman of Landing AI, and former lead of Google Brain, Ng has been at the forefront of nearly every major phase in AI development. In 2024, he introduced the concept of agentic AI-multistep, tool-using systems capable of executing workflows-which he believes will deliver more near-term economic value than simply scaling larger models.

Key Points from Andrew Ng:

• AGI Is Decades Away: Ng asserts that achieving artificial general intelligence (AGI), or human-level cognitive abilities in machines, is still decades away. The focus should instead be on practical applications.

• Training Layer Bubble Risk: He warns of a potential bubble in the training layer of AI systems. This layer involves the compute resources and data required to train models, which are becoming increasingly expensive and resource-intensive.

• Agentic Systems as the Future: Ng argues that agentic systems-AI agents capable of performing complex workflows-will be the driving force behind the next phase of AI adoption in enterprises. These systems can automate tasks more efficiently than large language models (LLMs) alone, leading to tangible business benefits.

The Challenges and Opportunities

• Compute Resources: The demand for compute resources is skyrocketing as companies race to train larger and more sophisticated models. This has led to significant costs, with some estimates suggesting that training a single state-of-the-art model can cost millions of dollars.

• Talent Acquisition: Attracting and retaining top AI talent remains a challenge. Companies are competing not only for data scientists and engineers but also for domain experts who can integrate AI into business processes.

• Economic Value: While large models have demonstrated impressive capabilities, translating these into economic value is complex. Agentic systems, which can automate workflows and integrate with existing tools, offer a more direct path to ROI.

Implementation Details

• Architecture of Agentic Systems:

  • Modular Design: These systems are often designed as modular components that can be easily integrated into existing enterprise architectures.
  • Tool Integration: They leverage APIs and other integration points to interact with various tools and services, enhancing their versatility.
  • Workflow Automation: By breaking down tasks into smaller, manageable steps, agentic systems can automate complex workflows more effectively than monolithic models.

• Benchmarks:

  • Efficiency: Agentic systems have shown higher efficiency in resource utilization compared to large LLMs, as they focus on specific tasks rather than general capabilities.
  • Scalability: They are designed to scale horizontally, allowing for the addition of more modules and tools as needed.

Conclusion

While the pursuit of AGI continues to captivate the imagination of researchers and investors, Andrew Ng’s pragmatic approach emphasizes the importance of practical applications. By focusing on agentic systems that can automate workflows and deliver tangible business value, enterprises can navigate the current AI landscape more effectively. The real risk lies in overinvesting in compute resources and talent without a clear path to economic returns.