Technology history will likely remember 2024 as the year of the great realization: Artificial Intelligence (AI) is not merely software, but a resource-heavy and capital-intensive infrastructure that demands a fundamental rethink of business models. As we navigate the summer of 2026, the promise of immediate profitability from Large Language Models (LLMs) remains trapped in what 36Kr analysts famously termed a "muddled mess." Silicon Valley and Chinese giants alike face the paradox of an "oversupply of intelligence" coupled with an "undersupply of value."
The Cost of Intelligence and the Revenue Trap
The core issue remains CAPEX (Capital Expenditure). Microsoft, Google, and Meta have funneled hundreds of billions of dollars into acquiring Nvidia’s H100 and Blackwell GPUs, constructing data centers with energy demands equivalent to small nations. However, the revenue side of the equation has not followed a similarly exponential trajectory. The $20-a-month subscription model, pioneered by ChatGPT, is proving insufficient to cover the inference costs of power users, while the integration of AI into existing products (like Microsoft 365 Copilot) faces friction from enterprise clients who have yet to see a commensurate boost in measurable productivity.
According to the 36Kr analysis, companies are locked in an arms race where innovation is occurring faster than the market's capacity to absorb it. Every time a model becomes more efficient, the price per token drops, leading to a deflationary cycle that benefits end-users but squeezes the profit margins of model providers. The "mess" is essentially a lack of pricing power in a world where everyone is offering similar capabilities.
The Clash of Models: Open Source vs. Proprietary
A primary driver of the commercialization deadlock is the rise of high-quality open-source models, led by Meta’s Llama series. When a GPT-4 class model is available for free download and local execution, the ability of OpenAI or Google to charge premium API fees is severely eroded. Enterprises are increasingly pivoting toward training smaller, specialized models (SLMs) on their own proprietary data, rather than relying on generalized models that are prone to hallucinations and raise significant data privacy concerns.
- The Shift to AI Agents: The market is moving from simple "chat" interfaces to autonomous execution of complex tasks.
- The GPU Utilization Problem: Many firms over-provisioned hardware and are now struggling to monetize that capacity.
- The Geopolitical Divide: While US firms lead in raw compute, Chinese giants like Baidu and Alibaba are excelling in vertical market applications.
The Chinese Perspective and the Price Wars
In China, the situation is even more acute. The "war of a hundred models" led to a brutal price war, with Alibaba and ByteDance offering model access at near-zero costs to capture the developer ecosystem. This "scorched earth" strategy aims for platform dominance but delays the emergence of a healthy AI software market. In 2026, Chinese firms are focusing less on LLM benchmarks and more on "AI-native" applications, desperately searching for the next TikTok-level success in the generative AI space.
"AI is currently in the same phase electricity was in the early 20th century: everyone knows it's essential, but nobody has figured out how to price it correctly without going broke," notes a Shanghai-based tech executive.
Conclusion: Survival of the Useful
To move past the current muddled state, tech giants must transition from demonstrations of raw power to the provision of seamless solutions. Commercialization will not stem from how intelligent a model is, but from how invisible and integrated it becomes within daily workflows. The "muddled mess" of the 2024-2026 period is a necessary correction—a period of adaptation before AI becomes a true economic engine, leaving behind those who invested only in the hype rather than the utility.
