Nila Kartika Wati
The rapid ascent of artificial intelligence, particularly large language models (LLMs), has been largely predicated on an unwavering belief in scale: bigger models, faster processing, and ever-expanding data centers. This paradigm, echoing traditional technology cycles, assumed that continuous scaling would inherently lead to improved performance, eventual cost reductions, and broader accessibility. However, this foundational assumption is now facing significant challenges, as AI’s development trajectory diverges sharply from conventional software, proving to be intensely capital-intensive, constrained by physical limits, and encountering diminishing returns far earlier than anticipated. This shift necessitates a critical re-evaluation of AI’s future, pushing the industry towards architectural innovation rather than merely exponential growth in compute power.
Mohammed Marikar, co-founder at Neem Capital, highlights this pivotal moment, arguing that the current scaling strategy is amplifying AI’s inherent weaknesses rather than resolving them. While larger models undeniably enhance fluency due to language’s pattern-based nature – allowing LLMs to excel at mimicking human writing, summarization, and translation – they struggle profoundly with deeper intelligence, particularly reasoning. The ability to understand cause and effect, acknowledge uncertainty, and provide transparent explanations for conclusions does not reliably improve with an increase in parameters or raw computational power. This fundamental limitation has profound implications, eroding public trust and creating a growing verification burden on human operators.
The Unsustainable Costs of AI’s Expansion
The financial and environmental footprint of AI’s current scaling trajectory is becoming increasingly stark. Data from the International Energy Agency (IEA) projects that global data center electricity demand will more than double by 2030, reaching levels comparable to entire industrial sectors. In the United States alone, data center power consumption is anticipated to surge by well over 100 percent before the decade’s end, driven predominantly by the insatiable demands of AI workloads. This explosive growth necessitates trillions of dollars in new investment, not only in specialized hardware like GPUs but also in the underlying energy infrastructure, including significant expansions in grid capacity.
The economic implications extend beyond operational costs. Training frontier AI models has become astronomically expensive, with credible tracking suggesting costs have been multiplying year over year. Projections indicate that single training runs could soon exceed $1 billion. This figure, however, represents only the entry cost. The far greater expense lies in inference – the continuous running of these models at scale, demanding low latency, high uptime, and rigorous verification. Every query consumes energy, and every deployment requires robust infrastructure, leading to a compounding effect on energy use and operational costs as adoption proliferates.
Erosion of Trust: AI’s Reasoning Deficit in Critical Sectors
Beyond the economic and environmental concerns, the reliability and trustworthiness of AI systems are coming under intense scrutiny, particularly as they are embedded into critical sectors such as law, finance, compliance, trading, and risk management. In these domains, errors propagate quickly, and credibility is non-negotiable. A significant incident in June 2025 saw the UK High Court issue a stark warning to lawyers, urging them to immediately cease submitting filings that cited fabricated case law generated by AI tools. This event underscored a critical flaw: when an AI system invents a precedent that never existed, and a professional relies on it, the debate about scaling transcends mere technical performance and becomes a serious question of public trust and professional ethics.
This issue of "hallucination" – where AI models confidently present factually incorrect or entirely fabricated information – is a direct consequence of the current scaling approach prioritizing fluency over reasoning. In fast-moving, competitive environments like crypto and traditional finance, fluent but unreliable AI can quickly propagate errors. False signals can move capital, and fabricated explanations can undermine market integrity and investor confidence. A common example is the generation of false positives in automated Anti-Money Laundering (AML) flagging systems. While intended to enhance security, these false alarms waste significant time and resources investigating innocent trading activity, highlighting the practical inefficiencies introduced by AI’s reasoning limitations. The verification burden on humans increases disproportionately with wider deployment, turning what should be an efficiency gain into a new layer of operational overhead.
The Imperative for Architectural Innovation: Beyond Brute Force
The current dominant approach to AI development, which largely focuses on increasing compute and data while leaving the underlying reasoning machinery unchanged, is proving economically unsustainable and proportionally less safe. Marikar argues that scaling AI systems without a commensurate improvement in their reasoning capabilities amplifies risk, particularly in use cases where automation and credibility are tightly coupled.
Ensuring AI is economically viable and socially valuable necessitates a departure from the sole reliance on scaling. The alternative lies in architectural innovation. Instead of merely predicting the next word or pattern, future AI systems must be designed to represent relationships, apply logical rules, check their own steps, and transparently explain how conclusions were reached. This is the realm of cognitive or neurosymbolic systems.
Cognitive AI platforms aim to organize knowledge into interrelated concepts, moving beyond brute-force pattern matching. By doing so, these systems can deliver high reasoning capability with significantly lower energy and infrastructure demands. Unlike massive LLMs that require immense data centers, emerging "cognitive AI" platforms are demonstrating the feasibility of structured reasoning systems operating on local servers or even edge devices. This shift not only reduces the carbon footprint but also allows users to maintain greater control over their own knowledge and data, rather than outsourcing cognition to distant, centralized infrastructure.
While cognitive AI systems can be more challenging to design and may initially underperform on highly open-ended tasks compared to their LLM counterparts, their advantages in specific, critical domains are profound. When reasoning is reusable – meaning a system can apply learned logical structures across different problems rather than re-deriving solutions from scratch through massive compute – costs fall dramatically, and the process of verification becomes far more tractable. This fundamental shift means that experimentation becomes cheaper, inference more predictable, and scaling no longer depends on exponential increases in physical infrastructure.
Decentralizing Intelligence: A New Paradigm for Development and Control
Beyond architectural shifts, the conversation around AI’s future also encompasses control and governance. The increasing centralization of AI development and deployment by a few tech giants raises concerns about concentration risk and the alignment of AI with global, rather than local, needs. Marikar emphasizes that control over how AI is built matters as much as how it reasons. Communities require systems they can shape, audit, and deploy without needing permission from centralized platform owners.
This growing need for autonomy and transparency has led to exploration of decentralized AI development models, often leveraging blockchain technology. Some platforms are actively exploring how blockchain can enable both individuals and corporations to contribute data, models, and computing resources in a distributed manner. By decentralizing the AI development process itself, these approaches aim to mitigate concentration risk, foster greater community involvement, and ensure that AI deployments are more closely aligned with local needs and ethical considerations. Such initiatives promise a future where AI’s evolution is not solely dictated by a few powerful entities but is a collaborative, transparent, and distributed endeavor.
An Inflection Point: Redefining AI’s Path Forward
Artificial intelligence stands at a critical inflection point. For years, the industry has chased the mirage of infinite scale, assuming that sheer size would eventually solve all problems. However, this strategy has exposed profound limitations, revealing that while scale can enhance fluency, it does not reliably produce deeper intelligence, reasoning, or inherent trustworthiness. The current path is economically unsustainable, environmentally detrimental, and increasingly problematic for public trust in critical applications.
The question facing the AI industry today is not whether to scale, but how. The dominant approach has already pushed the limits of what can be achieved by simply making systems bigger. What this journey has exposed, with startling clarity, are the inherent limitations of relying on size alone. The alternative, championed by experts like Mohammed Marikar, is to invest strategically in architectures that prioritize reliability, reasoning, and transparency before chasing further increases in magnitude.
By shifting focus from brute-force pattern matching to intelligent, cognitive architectures and embracing decentralized development models, AI can evolve into systems that require less compute per decision, impose a smaller verification burden on humans, and fundamentally alter its economic and ethical landscape. This paradigm shift promises a future where AI is not just bigger, but genuinely smarter, more sustainable, and ultimately, more trustworthy. The decision now rests with researchers, developers, investors, and policymakers: to continue down a path of diminishing returns, or to embark on a new era of architectural innovation that redefines what intelligence truly means in the age of machines.
