The scaling law contradiction
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
For four years, the AI industry lived by a single commandment: scale everything. More parameters, more data, more compute — and intelligence would follow. But in 2026, the returns from simply throwing more resources at larger models appear to be diminishing. The contradiction is not just technical — it is economic, philosophical, and existential.
The Law That Defined an Era
Scaling laws are a foundational principle in the AI industry, first articulated by OpenAI in 2020. The concept is simple: model performance improves predictably and proportionally as you increase three factors: compute, data, and parameters. For years, this held true — and the industry spent hundreds of billions of dollars acting on it.
The law itself is a specific instance of a general economic principle: the law of diminishing returns. In economics, diminishing returns describe a situation in which the increase in production, profits, or benefits resulting from something is less than the money or energy invested. The more you invest, the smaller the marginal gain.
The Wall That Wasn't Supposed to Exist
By late 2024 and into 2026, signs of a slowdown became apparent. Reports indicated that OpenAI's next-generation Orion (later released as GPT-5) had not achieved the dramatic leap seen from GPT-3 to GPT-4. The improvement was marginal relative to the enormous cost. Google faced similar issues with its Gemini updates, and Anthropic postponed the release of its Claude 3.5 Opus model.
The response from industry leaders was sharply divided:
Ilya Sutskever, a former OpenAI chief scientist and early proponent of scaling, has stated that the returns from scaling pretraining have "plateaued". He argues that the era of scaling is giving way to an era of discovery — searching for new approaches.
Sam Altman, CEO of OpenAI, has publicly declared "there is no wall". He and other Microsoft executives (Mustafa Suleyman, Kevin Scott) maintain that scaling will continue to deliver progress.
Demis Hassabis of Google DeepMind, however, has provided a critical nuance. He has noted that his team was actually the original discoverer of scaling laws (internally called "Chinchilla Scaling Laws" around 2017-2018). He maintains that scaling remains essential and has not hit a fundamental wall, but believes AGI will require "one or two significant breakthrough innovations" beyond the current Transformer paradigm to achieve AGI.
The Economic Contradiction
The investment figures underscore the contradiction. By 2026, aggregate U.S. hyperscaler AI capital expenditure for 2025-2026 approached $700 billion — a figure that has nearly doubled in a single year. Yet the returns on that investment are becoming less certain. As one analysis put it, "spending is projected to outpace earnings by year-end, pushing net free cash flow into negative territory until late 2027."
This is not merely a technical problem — it is a business model problem. If the next generation of models costs billions more but delivers only incremental improvements, the investment case becomes untenable. As an OpenAI researcher noted: "We really want to spend trillions of dollars or tens of trillions of dollars training models? At some point the scaling paradigm is going to collapse".
The Response: Scaling at Inference Time
The industry has responded by pivoting to a new axis of scaling: inference-time compute. Instead of simply training larger models, systems now scale up the compute allocated during the reasoning phase. This is the principle behind OpenAI's o1 reasoning models and DeepSeek-R1 — systems that take longer to "think" before responding, using techniques like test-time compute, reinforcement learning, and long chain-of-thought reasoning.
These methods demonstrate that even without larger models, more compute at runtime can significantly improve reasoning performance. The "new" scaling laws are about allocating compute intelligently to problems as they arise, rather than just building bigger engines. Noam Brown of OpenAI illustrated the potential: "having a bot think for just 20 seconds in a hand of poker got the same boosting performance as scaling up the model by 100,000x and training it for 100,000 times longer".
The Architectural Frontier
Beyond inference-time scaling, a range of architectural innovations are reshaping the scaling landscape:
Mixture of Experts (MoE): Instead of activating all model parameters for every input, MoE dynamically engages only the relevant parts. DeepSeek-V3, for example, requires significantly fewer GPU hours than a full dense model of similar capability.
Post-training and Reinforcement Learning: Techniques applied after the initial pretraining phase have become more important. Evidence suggests that substantial performance gains can be achieved through careful post-training optimization, rather than simply increasing the size of the model .
Precision scaling: Recent research has proposed that the precision of model weights matters more than previously thought, and that 7-8 bits may be optimal for compute efficiency — not simply 16-bit or lower.
The GFN Context
For Global Future Nexus, the scaling contradiction carries profound implications for governance and sustainability.
First, architectural diversity. If pure scaling is reaching its limits, the path to AGI will not be a single track but a branching landscape. Governance frameworks must be architecture-agnostic.
Second, sustainability. The energy demands of traditional scaling — data centres, chips, cooling — are immense. If inference-time scaling and architectural efficiency can deliver intelligence with less compute, they may also deliver intelligence with less environmental damage. GFN's Stewarded Sustainability framework must account for this trade-off.
Third, economic concentration. The $700 billion infrastructure bet is not evenly distributed. Those who control the next generation of compute infrastructure will shape the AGI era.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
