AI’s Price Revolution

Policy for a rapidly evolving age

AI is now improving faster than Moore’s Law, the old yardstick that predicts a doubling of price-performance every 18 to 24 months. The acceleration is driven primarily by innovations in AI model architectures and training methodologies, producing more efficient and more powerful systems.

Capability density measures the ratio between a model’s effective parameter size (the minimum number of parameters needed to achieve a given performance level) and its actual parameter count. An empirical trend dubbed the “Densing Law” reveals that the maximum capability density of LLMs doubles roughly every 3.3 months. This exponential growth means models rapidly become more efficient, achieving similar or superior performance using fewer parameters and significantly reduced costs. If this trajectory continues, we could see an improvement of approximately one million times in AI price-performance by 2030.

This rapid advancement is partially due to the adoption of Mixture of Experts (MoE) architectures. MoE models incorporate multiple specialised expert sub-models, selectively activating only those needed based on the specific input during inference. This sparsely gated design lets models scale to trillions of parameters without a matching rise in compute.

DeepSeek, an emerging Chinese AI lab originally spun out of a quantitative trading group, has disrupted the AI industry by releasing its V3 family of LLMs and subsequently introducing DeepSeek R1. The R1 model, characterised as a “reasoning-first” or “reasoning-heavy” model, matches or nearly matches OpenAI’s advanced o1 model on various coding, maths, and logic benchmarks. DeepSeek did this cheaply. The final training run of V3, the base model beneath R1, reportedly cost under $6 million, puncturing the assumption that a GPT-4-level model requires tens or hundreds of millions in compute.

More notably still, DeepSeek R1 is open source, with both architecture and weights publicly available. It uses a large-scale Mixture-of-Experts configuration with over 600 billion total parameters, though only a subset is activated at any given time. The model also employs a specialised “simulated reasoning” training approach, including iterative reinforcement learning cycles focused on tasks such as maths, coding, and puzzle-solving. This process reinforces careful “chain-of-thought” reasoning.

R1’s rapid spread triggered international attention and debate. Critics, notably Microsoft and OpenAI, have alleged that DeepSeek’s “student-teacher distillation” approach could amount to unlawfully leveraging proprietary knowledge from models like GPT-4, potentially infringing upon intellectual property protections. Given OpenAI’s own practice of mining online data without explicit permission, the ethical force of that argument is debatable.

These IP disputes are entangled with broader geopolitical concerns. The Chinese government may leverage open-source AI advances like R1 to challenge American leadership in AI, lowering barriers to entry for building powerful systems globally. Because R1’s weights are openly released, the model can be run locally without the extra filtering on DeepSeek’s hosted service, which deflects on politically sensitive topics such as Tiananmen Square, though some of that reticence is trained into the model itself. That gap has intensified discussions around censorship and free speech. There is speculation about potential actions from US agencies, including partial bans or blacklisting, particularly if they view the model as violating US intellectual property rights or circumventing export controls.

Meanwhile, OpenAI’s “Operator” system illustrates a parallel development. Operator equips ChatGPT-like agents with browser automation: navigating e-commerce platforms, filling out complex forms, acting autonomously online. It extends well beyond traditional question-answer interactions, with the potential to streamline business workflows and create new digital marketplaces.

Operator is not the only browser-automation tool available, but it represents a meaningful advance in practical applicability. Early use cases range from bill payments and travel arrangements to quality assurance testing for local software environments. The open-source community is following suit, developing browser agents compatible with various GPT-like models, with specialised applications emerging in healthcare, finance, and government services. Despite early inefficiencies, the trajectory is clear: these systems will mature quickly, transforming workflows and expanding AI-driven commerce.

Together, low-cost open-source AI and integrated digital agents mark AI’s transition from research curiosity to practical automation. Jevons Paradox applies: increased efficiency stimulates greater demand rather than reducing it. Lower per-token costs drive increased use, which in turn drives greater investment in GPUs, data centres, and energy infrastructure.

The US “Stargate” programme exemplifies the dynamic: a $500 billion initiative backed by President Trump alongside SoftBank and OpenAI, focused on large-scale GPU clusters, specialised hardware, and next-generation nuclear-powered data centres. Microsoft has signed a deal to restart a reactor at Three Mile Island to power its data centres, while Amazon and Google are backing small modular reactors.

Europe faces distinct regulatory challenges. Since Brexit, British firms that sell into the EU must satisfy both the EU AI Act and the UK’s own, lighter-touch regime. The UK government has promised to expand its sovereign AI computing capacity, with designated “AI Growth Zones” and faster approval for data centres and nuclear plants. Proposed “rights reservation” frameworks aim to balance copyright protection with AI innovation.

For investors and industry leaders, the picture is one of growth rather than a race to the bottom. Open-source breakthroughs complement rather than replace proprietary solutions, enabling wider experimentation and specialised applications across sectors.

The net effect is a more democratised, yet more energy- and capital-intensive, AI ecosystem.