In the time it takes to say “10 billion”, AI systems use roughly that many tokens, the units of text that a large language model like Claude or ChatGPT reads and writes.¹ The number is hard to picture. Humans are poor at reasoning about large numbers and probabilities, yet both are essential to understanding AI.

Birthday coincidences offer a familiar illustration as human intuition struggles with how often birthdays are shared within a growing group. AI presents a greater challenge of scale. It is fundamentally a statistical engine, and what is commonly referred to as a hallucination is, in fact, an expression of variance. To plan for what comes next, we need to be able to better grasp the numbers involved.

Token economics

One common perception about AI is that models will keep improving through ever-larger rounds of pre-training, but that era may already have passed. The first generation of progress was driven by expensive, time-consuming pre-training that ran flat out for 12 to 18 months. Since then, differences between frontier models have narrowed. Practices are widely shared, models assimilate advances across the industry and performance gains have begun to flatten.

The economics are changing just as quickly. Tokens have become a new industrial commodity, with the price falling by around 80% a year since 2021.² That decline reflects both technical and non-technical factors, and may continue. It also helps explain why frontier model providers work so hard to persuade customers to adopt their latest offerings. If they cannot sustain a premium for newer models, they face a difficult cost recovery problem.

Who uses tokens, and for what purpose, matters too. Coding has been one of AI’s most consequential applications because it is expansive. A short instruction, such as “write me a fitness app”, can produce a much longer output (high number of tokens). Other forms of white-collar work are compressive, such as asking a model to summarise material, which consumes input but produces a shorter response (fewer tokens). Building infrastructure around assumptions drawn from legacy software engineering may therefore misread future demand.

New methods bring new risks

As pre-training has become less productive, attention has shifted to other ways of improving AI: post-training with human feedback, reasoning through chains of thought, harnesses that sit around models, and agents. Each has had its own difficulties. While pre-training ran short of data, post-training encountered sycophancy, and effective harnesses raised the question of why underlying models needed further training. Agents bring a more serious concern that systems operating with greater autonomy may create risks that are harder to contain.

That concern is sharpened by the prospect of recursive self-improvement, which describes the process whereby models contribute to their own training and use large amounts of compute in the process. This could mark a shift into an accelerating cost regime.

More inference, describing the work of running a trained model to produce answers, could also mean more accidents. Notably, a consequential accident, rather than a deliberate attempt to cause harm, may be the more likely danger as verification becomes harder and autonomy increases.

The geopolitics of model development

Pre-training favoured the US, but the emerging industrial inference economy may favour China. Inference is a different production problem from training, with memory a particular constraint and power a central requirement. China's inference-manufacturing capacity has grown quickly, and its installed generating capacity is 3.4 times that of the US.³ These are material factors in how global compute capacity may be distributed.

The same shift could lower the cost of actions that were once difficult. Biological warfare is one concern alongside cyber risk, but attention should not remain fixed on threats we already understand.

“P(doom)”, used to refer to the probability of catastrophic outcomes, is not just a question of one dramatic event. Several individually unlikely events could combine into a much larger risk.

A wider financial reckoning

AI may prove widely useful, but that does not rule out a costly financial reckoning across real estate, government, credit and technology. The longer-term concern extends to the global payments system. A number of countries have been developing alternatives to the existing cross-border payment infrastructure, attracting cross-border inflows from sovereign and institutional sources. A further shift could erode the US dollar’s privileged position. An attempt to increase dollar supply through crypto and channel it into Treasuries could create a dangerous feedback loop whereby a dollar crisis becomes a bond crisis.

The central message is not that AI lacks value. It is that rapidly cheapening tokens are changing what can be produced, who can produce it and at what scale. The industrial, geopolitical and financial consequences may prove harder to manage than the technology itself.

1. Source: tokensperday
2. Source: The plunging price of thought | Epoch AI
3. Sources: China - National Energy Administration (NEA), National Power Industry Statistics, January–May 2026; United States - Energy Information Administration (EIA), Electricity Generation, Capacity, and Sales and 2026 Generating Capacity Additions Forecast.