AI and Energy Consumption

Written by Ben Esplin

The Resource Paradox: How AI's Greatest Inefficiency Reveals Its True Purpose

Construction scaffolding represents an expensive paradox: temporary metal frameworks that consume materials and labor, only to be dismantled once the permanent building stands complete. Today's large language models present a strikingly similar contradiction—computationally profligate systems whose greatest value may lie in designing their own elegant replacements.

The numbers are staggering. Training GPT-4 required approximately $78 million in computational resources, while cumulative inference costs by end of 2024 reached an estimated $2.3 billion—roughly fifteen times its training expense. Data centers supporting AI workloads are projected to consume between 4.6% and 9.1% of total U.S. electricity demand by 2030, up from approximately 4% in 2023. An August 2025 Goldman Sachs analysis forecasts that approximately 60% of increasing data center electricity demands will be met by burning fossil fuels.

From a conservation perspective, these projections represent a fundamental question about prudent stewardship of finite energy resources. We are consuming vast quantities to produce systems that remain fundamentally inefficient at their core tasks. Yet this very inefficiency may point toward their most valuable application: serving as expensive scaffolding from which to design fundamentally better approaches to machine intelligence.

Architectural Waste by Design

The transformer architecture underlying most contemporary LLMs requires O(n²) attention complexity—computational requirements grow quadratically with sequence length. Every token must attend to every other token, which becomes computationally expensive for long sequences. Research demonstrates that proper application of efficiency optimizations can reduce total energy use by up to 73% from unoptimized baselines—a remarkable figure that simultaneously highlights both the potential for improvement and the baseline wastefulness of current implementations.

The 73% reduction figure reveals how much waste is baked into the current paradigm. We are squandering resources not because the computational task inherently demands such consumption, but because the architectural approach we have chosen is fundamentally inefficient.

The models themselves exhibit frustrating limitations despite their resource demands. LLMs frequently hallucinate, lack genuine long-term memory, and struggle with tasks that adult humans find straightforward. These limitations exist not at the margins but at the core of the technology's fundamental architecture.

The Expensive Ladder

Why might these inefficient systems be valuable precisely because they enable discovery of efficient ones? The answer lies in understanding technological development as iterative exploration. We often must build wasteful prototypes to understand the problem space well enough to create efficient solutions.

Historical precedent supports this pattern. The iPod revolutionized music consumption, storing around 1,000 songs without physical media—only to be replaced by iPhones that consolidated all media access into a single device. Pagers were essential for emergency communications until mobile phones rendered them obsolete. Each represented a necessary step in understanding what users needed and how technology could meet those needs.

Large language models may represent precisely this kind of expensive ladder. Their resource intensity and architectural limitations have enabled rapid exploration of what machine intelligence systems can achieve, what they cannot achieve, and what users need from such systems. This exploration, in turn, is revealing what more efficient architectures might look like.

Conservation as Intellectual Property

From an intellectual property perspective, this technological inflection point creates complex strategic considerations. As jurisdictions increasingly mandate energy accounting, companies developing energy-efficient AI systems may gain regulatory advantages alongside their technical edge. Resource conservation itself may become a valuable form of technical achievement worthy of intellectual property protection.

Trade secret protection may prove particularly relevant for resource-efficient AI architectures. Unlike patents, which require public disclosure, trade secrets protect information that derives competitive advantage from remaining confidential. The specific architectural choices, training procedures, and optimization techniques that enable dramatically more efficient AI systems could constitute valuable trade secrets—particularly if the approaches are not readily reverse-engineerable from model outputs.

A series of blockbuster trade secret cases in 2024 demonstrated the explosive value of such protection. In Propel Fuels v. Phillips 66, a jury found the oil giant liable for willful misappropriation of 77 trade secrets involving proprietary algorithms, awarding Propel $604.9 million in unjust enrichment damages. This case demonstrates that trade secrets—especially those involving cutting-edge technology—can be among a company's most valuable assets.

Current USPTO guidance clarifies that AI systems cannot be named as inventors, but AI-assisted inventions remain patentable if natural persons significantly contribute to the claimed invention. This framework creates interesting tensions when LLMs generate novel architectures specifically designed to conserve computational resources. Who owns the intellectual property in a neural architecture discovered entirely through automated search—an architecture that might use a fraction of the energy of conventional approaches?

The Race to Obsolescence

The present moment in AI development resembles using an extraordinarily expensive, energy-intensive tool to design the simple, elegant, resource-conserving tools we should have built first. This is neither criticism nor failure—it represents a necessary stage in technological evolution. But from a conservationist perspective, the urgency lies in accelerating this transition before we exhaust precious resources on the scaffold itself.

The killer application for large language models may ultimately be their role in accelerating the discovery of post-LLM approaches to intelligence: architectures that achieve comparable or superior capabilities while consuming a fraction of the resources. Once those paths become clear, the ladder itself may become largely obsolete.

The true measure of success for contemporary AI may be how rapidly it designs its successors and how gracefully it steps aside, minimizing the resources squandered in the process. This represents not sustainability through perpetual optimization, but genuine conservation: achieving more while consuming less, and ultimately preserving finite resources for purposes beyond computational excess.

The scaffolding metaphor reminds us that temporary, expensive structures have value when they enable permanent, elegant solutions. The challenge lies in ensuring we dismantle the scaffolding before it collapses under its own weight—and in protecting the intellectual property that makes the permanent structure possible.

If you made it this far, and would like support for any of the facts stated above, please reach out to me directly.