Why Expanding AI Infrastructure Increases Danger Instead of Performance
Massive AI expansion consumes trillions in power resources while magnifying fundamental flaws. Alternative approaches using neurosymbolic frameworks and distributed cognitive architectures offer dependable intelligence minus the hazards.
Opinion by: Mohammed Marikar, co-founder at Neem Capital
The trajectory of artificial intelligence development has been characterized primarily by magnitude — larger neural networks, accelerated computational power, proliferating data infrastructure. The prevailing belief, drawn from conventional technological evolution patterns, held that expansion would continuously enhance capabilities while eventually reducing expenses and broadening availability.
This foundational premise is currently collapsing. AI development does not follow the same trajectory as traditional software systems. Rather, it demands enormous capital investment, faces concrete physical constraints, and encounters declining marginal returns much sooner than industry experts anticipated.
Statistical evidence demonstrates this reality unmistakably. Power consumption from worldwide data infrastructure is expected to increase more than twofold by 2030 — consumption rates previously linked to complete industrial economies. Within the United States specifically, data center electricity requirements are forecast to surge beyond 100 percent growth prior to the decade's conclusion. Such massive expansion necessitates investments reaching into the trillions of dollars combined with substantial enhancements to electrical grid infrastructure.
At the same time, AI technologies are becoming integrated into legal systems, financial operations, regulatory compliance, market trading and risk assessment functions, where mistakes cascade rapidly while trustworthiness remains absolutely essential. In June 2025, the UK High Court issued warnings to legal professionals demanding immediate cessation of filing submissions containing fabricated legal precedents generated through AI applications.
The scaling AI debate
Once an AI platform can fabricate a legal precedent with no factual basis, and qualified professionals depend on such output, conversations about expansion transform into legitimate concerns regarding societal confidence. The scaling process is magnifying AI's fundamental vulnerabilities instead of resolving them.
A significant portion of the challenge resides in understanding what expansion genuinely enhances. Large language models (LLMs) demonstrate increasing fluency precisely because linguistic communication follows recognizable patterns. The greater quantity of examples an LLM processes showing how actual humans compose text, condense information and perform translations, the more rapidly its performance advances.
True cognitive capacity — logical reasoning — does not follow identical scaling dynamics. Next-generation AI platforms must comprehend causality and relationships, recognize when responses contain uncertainty or gaps, and provide explanations for how conclusions were derived rather than merely generating responses that appear authoritative. These capabilities do not consistently strengthen through additional parameters or increased computational resources.
The result is an expanding verification requirement. Human operators must dedicate increasing time to validating machine-generated results rather than implementing them directly, and this overhead intensifies as these systems achieve broader deployment.
The cost of training AI models
Developing cutting-edge AI models has evolved into an exceptionally costly endeavor, with reliable tracking indicating expenses have been multiplying annually, alongside forecasts suggesting individual training cycles may shortly surpass $1 billion. Training represents merely the initial investment.
The more substantial ongoing expense involves inference: operating these models perpetually, at significant scale, with genuine latency constraints, availability guarantees and verification obligations. Each individual query demands energy consumption. Each implementation necessitates supporting infrastructure. As adoption expands, power requirements and financial costs multiply exponentially.
Regarding financial markets and cryptocurrency ecosystems, AI platforms are progressively deployed to supervise blockchain activity, evaluate market sentiment, produce code for smart contract development, identify potentially fraudulent transactions and execute automated decision-making.
Within such rapidly evolving, highly competitive environments, linguistically sophisticated yet fundamentally unreliable AI propagates mistakes at alarming speed; inaccurate signals redirect investment capital, while manufactured justifications and hallucinatory outputs erode foundational trust. A clear illustration of this phenomenon involves false positive results produced through automated Anti-Money Laundering (AML) detection systems, a widespread problem that squanders valuable time and resources investigating legitimate trading operations.
Time to improve reasoning
Expanding AI infrastructure without strengthening fundamental reasoning capabilities magnifies hazards, particularly in applications where automation and credibility are both critical and inherently interconnected.
Guaranteeing AI remains economically sustainable and delivers societal benefit requires moving beyond reliance on pure scaling. The prevailing methodology currently emphasizes expanding computational power and dataset size while maintaining the core reasoning mechanisms essentially unmodified, a tactical approach that grows increasingly expensive without proportional improvements in safety or reliability.
The solution lies in fundamental architectural innovation. Platforms must accomplish more than predicting sequential tokens. They require the ability to model relationships, implement logical rules, validate their own procedural steps and enable transparent visibility into how determinations were achieved.
This represents the domain where cognitive or neurosymbolic architectures demonstrate their value. Through organizing information into interconnected conceptual frameworks, rather than depending exclusively on computational brute-force pattern recognition, these platforms can achieve superior reasoning performance while demanding dramatically reduced energy consumption and infrastructure investment.
New "cognitive AI" frameworks are proving how structured reasoning architectures can function on local server infrastructure or edge computing devices, permitting users to maintain governance over their proprietary knowledge rather than delegating cognitive processing to remote centralized infrastructure.
Cognitive AI architectures present greater design complexity and may exhibit reduced performance on unstructured exploratory tasks, but when reasoning capabilities become reusable in this manner instead of being regenerated from initial states through massive computational effort, operating costs decline substantially and verification processes become practically manageable.
Authority over AI construction methodologies matters equally as much as reasoning design itself. Communities require systems they can customize, examine thoroughly and implement independently without awaiting authorization from centralized platform controllers.
Certain platforms are pioneering this territory by leveraging blockchain technology to enable both individual participants and corporate entities to provide data resources, model architectures and computational capacity. Through decentralizing the AI development process itself, these methodologies diminish concentration vulnerabilities and synchronize deployment with localized requirements rather than globally centralized objectives.
AI confronts a critical inflection point. When reasoning capabilities can be recycled rather than reconstructed through extensive pattern matching operations, platforms demand reduced computation per individual decision and create smaller verification obligations for human operators. This transformation fundamentally alters the economic equation. Experimentation becomes more affordable, inference operations become increasingly predictable. Expansion no longer requires exponential growth in underlying infrastructure.
The scaling approach has already accomplished what it could feasibly achieve. What it has simultaneously revealed, with equal clarity, are the inherent limitations of depending on magnitude alone. The critical question facing the industry now is whether continued investment pursues ever-larger scale or redirects toward architectures that establish intelligence reliability before pursuing dimensional growth.
Opinion by: Mohammed Marikar, co-founder at Neem Capital.