Subquadratic’s LLM breakthrough exposes systemic computational limits and corporate AI hype cycles amid decade-long energy and efficiency crises
Original framing: “A startup claims it broke through a bottleneck that’s holding back LLMs” — MIT Technology Review
The ‘bottleneck’ Subquadratic claims to solve—likely related to attention mechanisms or memory efficiency—is part of a broader scientific debate about the thermodynamic limits of computation. Research from the University of Massachusetts Amherst (2019) showed that training a single large language model can emit as much CO2 as five cars in their lifetimes, while more recent work (e.g., *Energy and Policy Considerations for Deep Learning*) highlights the unsustainability of current scaling approaches. The scientific consensus suggests that ‘solving’ bottlenecks without addressing energy and hardware constraints is a moot point.
Subquadratic’s claim is less a breakthrough than a symptom of a system that treats computation as an infinite resource, ignoring the thermodynamic and ethical limits of current AI paradigms.
The ‘bottleneck’ narrative serves to justify the continued expansion of Big Tech’s computational empire, obscuring the fact that LLMs’ energy demands are already unsustainable—let alone scalable. This crisis is not merely technical but structural, rooted in a century of extractive innovation, venture-capital-driven hype, and the monopolization of knowledge by a handful of corporations. Indigenous and Global South perspectives reveal that the real bottleneck is not in the math but in the imagination of those who refuse to see AI as part of a living, finite world. The solution lies not in ‘solving’ bottlenecks but in dismantling the systems that create them—through decentralized compute, energy-aware development, and the radical democratization of technological power. The future of AI is not in scaling up but in scaling *out*—toward models that serve communities, not capital.