Just under a year ago, the AI sector began to recognize mounting concerns regarding the possibility of a bubble. Sam Altman of OpenAI remarked that “intelligent people sometimes get overly enthusiastic about a hint of truth.” Mark Zuckerberg acknowledged the empirical possibility of a bubble, while Demis Hassabis from Google noted that “certain sections of the AI industry might indeed be experiencing a bubble.” Jeff Bezos pointed out indicators of an “industrial bubble,” despite his belief that AI would revolutionize various industries.
A few months later, discussions about the bubble subsided with the introduction of a new wave of AI coding tools. These tools proved to be significantly more effective for complex, real-world programming tasks, gaining widespread adoption within the tech sector. This shift had several implications for AI companies. It revitalized the narrative of constant progress, though the industry's trajectory was still unexpected. Insider sentiment was once again focused on acceleration, as AI experts transitioned from describing the majority of AI-generated code as “sloppy” to entrusting a considerable amount of their workload to a new tool named Claude Code. This transition not only reset industry narratives but also provided AI firms a compelling product to offer other businesses, sparking immediate interest and leading to a feasible business model that extended beyond merely acquiring users. However, the advanced AI coding tools were notably compute-intensive, requiring vast amounts of tokens for operation, which surprised customers and prompted concerns about sustainability on the provider side. Yet, this scenario also alleviated, at least momentarily, fears regarding the overexpansion of data centers. By the first half of 2026, AI capacity was in short supply, with major companies maximizing access sales, and ambitious infrastructure investments regained investor confidence.
However, this optimistic outlook was short-lived. Growing anxieties by late 2025 were partly due to concerns that progress in AI models might be slowing. The introduction of agentic AI coding brought some relief; nonetheless, the emergence of cost-effective Chinese open-weight models raised further worries. These models, which could run on customers’ own hardware, had capabilities comparable to their American counterparts but were priced much lower. A model known as DeepSeek R1, released in mid-2025, made headlines by promising functionality akin to OpenAI’s GPT-4 while citing significantly lower training costs, which briefly impacted American tech stock values. Now, in this new phase of AI coding, Chinese firms have resurfaced. Over the past six months, both Anthropic and OpenAI crossed a significant milestone in AI software development. Simultaneously, several competitors, including Moonshot AI with its Kimi model lineup and Z.ai with its updated GLM model, are also achieving breakthroughs.
Nathan Lambert, an AI researcher, noted that the reception of Z.ai's GLM-5.2 mirrors the excitement previously seen with the launch of DeepSeek R1. Many respected voices within the AI community have praised the model after firsthand experience. Compared to Kimi K2, GLM-5.2 is being hailed as a significant leap forward for open-weight models. It is notable for its effectiveness in coding as a general agent, making it the most capable affordable AI tool in comparison to Claude Code’s more recent Anthropic models, if not better than Google’s offerings — an outcome that would have seemed unlikely just a few months prior.
The broader impact of this democratization of agentic coding tools as of early 2026 remains uncertain. While frontier labs have continued their forward momentum, with Anthropic recently unveiling (and subsequently recalling) its Mythos and Fable models and introducing novel capabilities in cybersecurity, it seems that high-end models still retain unique and valuable offerings for clients. OpenAI has recently claimed to match this new capability at a lower cost, while Z.ai intends to provide similar offerings by year’s end. The advent of more affordable AI coding models could challenge the leading American firms, particularly as they have significantly out-funded their Chinese competitors. Alternatively, these economical options may stimulate demand and broaden access to advanced coding tools without significantly undermining the premium offerings of leading firms.
Complicating matters further is the hint of another potential breakthrough. Industry leaders are increasingly discussing “loops,” suggesting that their models could soon achieve “recursive self-improvement,” which might solidify their competitive edges in crucial ways. Although it may sound fantastical, given the financial resources and research dedicated to innovation in AI, dismissing the possibility that the next generation of models could provide immense economic value beyond coding would be unwise. The industry is eagerly searching for its next major narrative.
Looking at the landscape from a distance, the overarching dynamic presents considerable risks for the industry. With each significant advancement in AI, affordable and adaptable models closely follow, turning previously expensive deployments into affordable solutions. American AI firms have expressed concerns about this phenomenon, framing it as a threat to national security, intellectual property, and a source of uncontrolled AI risks, complicating their emerging business models. While discussions of a bubble haven’t resumed in full force, early indications—seen in the sentiment among tech leaders and a sudden slump in AI and chip stock valuations—suggest that perceptions may be shifting once more, much like they have multiple times over the past three years.
