Has the progress of AI "slowed down," should the market be worried?

The breakthrough progress of AI technology is showing signs of slowing down, but for many companies seeking to leverage this technology, it may not be a bad thing. Analysts point out that current AI tools are powerful and practical enough, while most companies have yet to fully tap into the potential of existing AI technologies. The temporary pause in technological development may actually provide companies with more time to adapt and integrate AI systems.

On August 24, media reports indicated that this summer, Meta postponed the release of its flagship AI model Llama 4 Behemoth due to engineers struggling to achieve significant improvements. OpenAI's latest model GPT-5 also faced delays, and its performance post-release did not meet expectations. These signs suggest that the development of large language models may be leveling off.

However, analysts believe that the slowdown in AI development may actually provide opportunities for companies to adopt it. Current AI technology is powerful and practical enough to handle business needs such as text summarization, programming assistance, and email writing. Companies need more time to adapt to and integrate existing AI technologies rather than chasing constantly upgraded models.

Meanwhile, although the adjustment in the speed of AI technology development has begun to affect market sentiment, with tech giants like NVIDIA, Microsoft, and Meta experiencing sell-offs last week, the difficulty in improving AI model performance may actually extend the prosperity period for certain companies, especially those like NVIDIA that "sell shovels," as AI giants will invest more resources to overcome technical challenges.

Leading AI Models Encounter Development Bottlenecks

Leading companies in the AI field are facing unprecedented technical challenges.

Meta's originally planned launch of Llama 4 Behemoth has been forced to delay because the engineering team could not achieve the expected performance improvements. This delay reflects that even resource-rich tech giants encounter technical ceilings when pushing for breakthroughs in AI model performance.

The situation for OpenAI is similarly grim. The release of GPT-5 has not only been delayed but its performance has also failed to meet market expectations. Altman recently displayed a rare sense of realism at a media dinner, acknowledging that investors' expectations for AI technology may be too high.

These signs indicate that the rapid iteration cycle of large language models may be slowing down, and the development of AI technology is shifting from exponential growth to a more incremental improvement model.

Enterprise Applications Still in the Early Stages

Although AI development may be slowing, this is not a major issue for companies trying to integrate AI technology; it may even be good news. Generative AI is already powerful and practical enough in the business field to handle tasks such as large text summarization, assisting employees with programming, or writing emails.

However, most companies have yet to fully explore the potential of current AI technology applications. While some companies have rapidly deployed AI technology, many others are moving slowly. Corporate technology leaders are concerned about the potential leakage of sensitive data through chatbot conversations and are cautious about allowing AI to handle key decisions that impact finances, employees, and customers A recent study by MIT found that companies have largely accepted the off-the-shelf generative AI tools from OpenAI and Microsoft. However, the failure rate of pilot projects for building custom AI software to streamline operations—applications most likely to yield real commercial returns—can be as high as 95%.

The authors of the study pointed out that corporate AI users are generally skeptical of custom or vendor-pitched AI tools, viewing them as fragile, over-engineered, or misaligned with actual workflows.

Reports indicate that the business sector clearly needs more time to understand AI technology, and the work of adapting large language models to everyday tasks is still in its early stages. Michael Chui, a senior researcher at McKinsey's AI division, stated:

"If I want to accelerate innovation, reduce inventory safety stock, and improve connections with millions of consumers, you have to change far more than 'here's a tool for a few employees to use.' All these changes are very difficult."

This difficulty represents both a management challenge and a technical challenge, indicating that the adoption of enterprise AI will be a multi-decade effort. It is worth noting that while the internet transformed people's lives and business practices, the time required far exceeded the expectations of enthusiastic supporters in the 1990s.

According to data from the Pew Research Center, it took a decade for U.S. household broadband to rise from nearly zero penetration in 2000 to over 60% of adults subscribing.

Companies like NVIDIA that 'sell shovels' may face a longer period of prosperity

AI development will not stagnate; even if providers of the most advanced models encounter bottlenecks, people will still seek ways to improve, and the perception of a slowdown in AI development may give companies more confidence to invest time and money, viewing it as a more stable target.

In the short term, the recognition that the rise of AI may not be as rapid as expected has led to volatility in tech stocks. Leading AI stocks like NVIDIA, Microsoft, Amazon, and Meta experienced sell-offs last week until comments from Federal Reserve Chairman Jerome Powell hinting at interest rate cuts triggered a rebound.

However, as it becomes more challenging to improve AI model performance, this may actually extend the prosperity period for certain companies, especially those like NVIDIA that 'sell shovels.' Major AI investors like Altman and Meta CEO Mark Zuckerberg may invest more funds to overcome recent challenges.

Altman recently suggested that the solution to OpenAI's recent difficulties is to invest several trillion dollars in AI chips, even as the process of adapting models to real business tasks requires more incremental computing power