From 100k sqm to 160k Petaflops: Scale Efficiency Bottlenecks at AI Conferences [Analysis]

https://www.thepaper.cn/newsDetail_forward_33531953

Just finished reading the briefing from the 2026 World Artificial Intelligence Conference press release. A few numbers are worth breaking down: exhibition area exceeded 100,000 square meters for the first time, computing power scale broke through 160,000 P (PFLOPS? Note: text says "160,000 P", industry convention suggests PFLOPS), industrial scale reached 637 billion RMB, up 39.5% year-on-year. Putting these numbers together reminds me of the Hoffmann et al. 2022 paper on Chinchilla scaling laws—when the compute budget is fixed, model size and training data volume need to grow in specific proportions to achieve optimality; exceeding this ratio leads to diminishing marginal returns.

The expansion of this conference's scale is essentially the same type of problem: we pile up more area, more companies, and more forums, but the growth pattern of "intelligence" itself is not linear.

---

The Scissors Gap Between Compute and Intelligence

Shanghai's intelligent computing power is 160,000 P, roughly several times the total domestic computing power in 2023. But there are only 169 registered large models, and over 300 globally launched products—the growth curve on the compute supply side is clearly steeper than on the application demand side. Referencing the experimental conclusions from Kaplan et al. 2020, there is a power-law relationship between model performance and compute, but the exponent is less than 1, meaning doubling compute doesn't buy double intelligence. More importantly, the mainstream architecture of current large models (decoder-only Transformer) has hit bottlenecks in data quality and diversity; simply piling on compute may be entering the "diminishing returns zone".

The Physical Gap in Embodied AI

The conference specifically mentioned that the embodied AI track gathered over 200 companies. This is a direction worth watching—because there exists a scaling law issue here more fundamental than LLMs: the cost of acquiring physical world data is far higher than text. According to preliminary experiments I've seen (such as Zhao et al. 2023 on robotics data scaling), the expansion coefficient for robot operation data is only about 0.3-0.4, meaning current data collection methods (teleoperation + simulation) struggle to support disorderly expansion of model scale. Shanghai claims the "full chain of embodied AI industry is basically built," but improving yield rates for mass-produced semiconductors requires physical feedback loops; this wall cannot be smashed through by compute alone.

---

However, two highlights from the conference are worthy of praise: one is the establishment of the "WAIC Academic" international academic conference, receiving 284 submissions covering multiple countries—academic exchange is key to breaking through first principles and has more long-term value than the bustle of exhibitions; the second is the "Hundred Teams, Hundred Projects" action for scientific intelligence, pushing AI into hard science fields like physics, chemistry, and biology. This is the path where scaling laws might find new data sources in the next step.

Overall, this is a critical point meeting shifting "from scale to quality." I hope attendees don't just showcase parameters and area, but talk more about attempts to break current scaling laws.