Scaling next generation AI is making it riskier, not better

The traditional assumption that Artificial Intelligence (AI) performance scales with size, leading to improved performance and reduced costs over time, is breaking down. AI is proving to be capital-intensive, constrained by physical limits, and encountering diminishing returns earlier than anticipated. This contrasts with how other software technologies have scaled.

Escalating Costs and Infrastructure Demands

The electricity demand from global data centers is projected to more than double by 2030, with US data center power demand expected to rise over 100 percent before the decade ends. This expansion requires trillions of dollars in new investment and significant grid capacity increases. Furthermore, training advanced AI models is becoming extraordinarily expensive, with costs for single training runs potentially exceeding $1 billion. The ongoing expense of inference—running models continuously—also compounds energy use and costs as usage grows.

Weaknesses Amplified by Scale

While AI systems like large language models (LLMs) improve in fluency by learning patterns in language, deeper intelligence such as reasoning does not scale in the same way. AI systems are beginning to exhibit weaknesses, including generating fabricated information, as seen when a UK High Court warned lawyers against submitting filings with AI-generated case law. This amplification of AI's weaknesses rather than their resolution creates a growing verification burden, requiring humans to spend more time checking machine output.

AI in Markets and Crypto

In financial markets and the crypto space, AI is used for on-chain activity monitoring, sentiment analysis, smart contract code generation, and transaction flagging. However, fluent but unreliable AI can propagate errors quickly, leading to false signals that move capital and undermining trust through fabricated explanations. A common issue includes false positives in automated Anti-Money Laundering (AML) flagging, which wastes resources on investigating legitimate activity.

The Case for Cognitive AI

To ensure AI is economically viable and socially valuable, a shift from prioritizing compute and data to improving underlying reasoning is necessary. Architectural changes are needed, moving beyond simple pattern matching to systems that represent relationships, apply rules, and verify their own steps. Cognitive or neuro-symbolic systems offer a solution by organizing knowledge into interrelated concepts. These systems can deliver high reasoning capability with significantly lower energy and infrastructure demands, potentially operating on local servers or edge devices.

Decentralization and the Future of AI

Control over AI development is also crucial. Decentralized approaches, potentially utilizing blockchain, allow individuals and corporations to contribute data, models, and computing resources. This reduces concentration risk and aligns AI deployment with local needs. AI is at an inflection point where reusable reasoning, rather than massive pattern matching, reduces compute needs per decision and lowers the human verification burden. This shift makes experimentation cheaper and inference more predictable, moving scaling away from exponential infrastructure increases and towards more reliable intelligence.