When we think about the transformative power of technology, it’s easy to focus on the inventions themselves: the steam engine, the personal computer, the internet, or the recent surge in artificial intelligence (AI). But beneath these revolutions lies a recurring story of commoditisation, the rapid, exponential decline in the cost of a critical resource that makes the technology scalable, accessible, and ultimately, revolutionary.
The industrial revolution was arguably the first great modern technological paradigm shift. At its heart was the steam engine, which powered factories, trains, and ships, enabling unprecedented production and transportation. But what made the revolution scalable was coal. The cost of extracting, transporting, and utilising coal fell dramatically over decades, making it the key commodity that powered industrialisation.
Coal’s declining cost fuelled exponential growth in energy availability, which in turn drove massive productivity gains across industries. Factories could produce goods faster and cheaper, while railroads and steamships connected markets across the globe. The ripple effects transformed economies, societies, and geopolitics.
Fast forward to the 20th century, and another revolution was underway: the rise of the personal computer. Here, the key commodity wasn’t a physical resource like coal but a technological one: transistor density, driven by Dennard scaling.
Dennard scaling, named after IBM researcher Robert Dennard, describes how as transistors shrink in size, their power consumption and cost decline proportionally. This principle underpinned Moore’s Law, which saw the number of transistors on a chip double roughly every two years.
The result? Computing power became exponentially cheaper and more efficient. This commoditisation of computation enabled the development of affordable personal computers, making them accessible to businesses, schools, and homes. The ripple effects included the rise of software industries, productivity tools, and eventually the internet itself.
The dotcom revolution of the 1990s and early 2000s was similarly fuelled by a commoditised resource: data transmission bandwidth. This was governed by Edholm’s Law, which predicted exponential increases in bandwidth over time, thanks to advancements in fibre optics, wireless technologies, and compression algorithms.
As the cost of transmitting data plummeted, the internet became a global phenomenon. Websites, search engines, e-commerce, and online communication tools flourished. Companies like Amazon, Google, and Facebook emerged, reshaping the global economy. The rapid decline in bandwidth costs made it possible to connect people and businesses at a scale never before imagined.
Today, we’re witnessing another paradigm shift: the AI revolution, driven by large language models (LLMs). These models, like GPT-4 and Meta’s LLaMA, are transforming industries from customer service to content creation, programming, and beyond.
But what’s the commodity fuelling this revolution? It’s the cost of training and utilising LLMs, which has been falling at an extraordinary pace. Industry estimates suggest a 10x annual decline in the cost of training models with a minimum score on the MMLU (Massive Multitask Language Understanding benchmark). This means that tasks requiring cutting-edge AI capabilities are becoming dramatically cheaper every year.
For example, training a state-of-the-art LLM in 2020 cost tens of millions of dollars. Today, similar capabilities can be achieved for a fraction of that cost, and open-source models are emerging that rival proprietary ones. This rapid commoditisation is making LLMs accessible to more organisations and individuals, democratising AI capabilities.
But, we’re now encountering a new challenge. Historically, increasing compute power, adding more GPUs, servers, and cloud resources, has scaled AI intelligence. But we’ve hit a wall: data scarcity.
LLMs require vast amounts of high-quality data to train. The internet, while massive, has finite high-value training data. As we run out of new data to train on, simply scaling compute no longer yields proportional gains in model performance. In this context, test-time compute is emerging as the next frontier.
Test-time compute refers to the resources a model uses during inference (when it’s generating outputs or predictions). Instead of relying solely on pre-trained knowledge, models can perform additional calculations, retrieve external information, or iterate on their outputs in real-time.
This shift is particularly impactful for open-source LLMs, which are becoming cheaper to develop and use. With test-time compute, even relatively small models can achieve high performance by dynamically adapting to tasks or integrating external data.
The commoditisation of LLMs is challenging the traditional SaaS (Software as a Service) model. Proprietary AI platforms like OpenAI’s GPT-4, which rely on subscription fees, face growing competition from open-source alternatives. As the cost of deploying and fine-tuning these models drops, businesses may opt to run their own models or leverage cheaper third-party solutions.
For hyperscalers like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud, the story here is a good one. These companies benefit from the commoditisation of LLMs because it drives demand for their cloud infrastructure. Training and deploying AI models require massive compute resources, and hyperscalers are well-positioned to provide them.
Hyperscalers have to invest heavily in data centres, GPUs, and specialised hardware. The falling cost of LLMs increases the return on these capital expenditures by boosting demand for AI workloads. For every dollar invested in infrastructure, hyperscalers can capture more revenue as customers scale their AI applications.
As the largest hyperscaler, Amazon stands to gain significantly. AWS already offers tools for deploying LLMs, including integrations with Meta’s LLaMA. By providing scalable, cost-effective infrastructure for open-source models, AWS can attract businesses looking to leverage AI without the high costs of proprietary solutions.
Meta’s open-source approach to LLMs, exemplified by LLaMA, positions it as a key player in the AI ecosystem. By making its models freely available, Meta encourages adoption and innovation while benefiting indirectly through ecosystem growth.
Today, the rapid decline in LLM costs is democratising AI, challenging incumbents in areas like SaaS, and creating new opportunities for hyperscalers and innovators. As businesses and individuals, understanding this dynamic is key to navigating the AI revolution and seizing its opportunities.
Disclaimer: The views expressed in this article are those of the author at the date of publication and not necessarily those of Dominion Capital Strategies Limited or its related companies. The content of this article is not intended as investment advice and will not be updated after publication. Images, video, quotations from literature and any such material which may be subject to copyright is reproduced in whole or in part in this article on the basis of Fair use as applied to news reporting and journalistic comment on events.
