Key Takeaways
- • Token prices have reportedly fallen by around 80% a year since 2021, turning AI tokens into what the article calls a new industrial commodity, according to Epoch AI data cited in the piece.
- • The article argues that gains from large-scale pre-training have flattened, shifting AI development toward post-training, reasoning, harnesses, and autonomous agents, each with its own risks.
- • China's installed generating capacity is cited as 3.4 times that of the US, which the article says may tilt the emerging inference economy toward China after pre-training favoured the US.
- • The article warns that a costly financial reckoning across real estate, government, credit and technology is possible, and that dollar-based payment alternatives combined with crypto-linked dollar supply could create a feedback loop from a dollar crisis to a bond crisis.
Summary
The article, based on a talk by MIT Institute for the Digital Economy research fellow Paul Kedrosky at the Man Alternative Investing Symposium, examines the changing economics of AI tokens and the risks AI poses for geopolitics and finance. It notes that AI systems process roughly 10 billion tokens in the time it takes to say "10 billion," and characterises AI as fundamentally a statistical engine in which hallucinations reflect variance.
On economics, the article states that the era of ever-larger pre-training runs may already have passed. The first generation of progress relied on pre-training that ran for 12 to 18 months, but differences between frontier models have narrowed and performance gains have begun to flatten. Token prices have fallen by around 80% a year since 2021, according to Epoch AI, and the article says frontier providers therefore work hard to persuade customers to adopt newer offerings to sustain premium pricing. It also distinguishes expansive uses such as coding, which can produce far more output than input, from compressive white-collar tasks such as summarisation, and warns that infrastructure built on legacy software-engineering assumptions may misread future demand.
The article identifies new risks as development shifts to post-training with human feedback, chains of thought, harnesses and agents. It flags recursive self-improvement, in which models contribute to their own training, as a possible move into an accelerating cost regime, and suggests that a consequential accident may be a more likely danger than deliberate harm as autonomy increases and verification becomes harder.
On geopolitics, the article argues that inference, which is constrained by memory and power, may favour China, whose installed generating capacity is cited as 3.4 times that of the US according to National Energy Administration and EIA data. It also says that several individually unlikely events could combine into a much larger catastrophic risk.
Financially, the article says AI may be widely useful yet still coincide with a costly reckoning across real estate, government, credit and technology. It highlights efforts by several countries to build alternatives to existing cross-border payment infrastructure, which could erode the US dollar's privileged position, and warns that using crypto to increase dollar supply and channel it into Treasuries could turn a dollar crisis into a bond crisis. The article concludes that AI retains value, but that cheapening tokens may make its industrial, geopolitical and financial consequences harder to manage than the technology itself.
On economics, the article states that the era of ever-larger pre-training runs may already have passed. The first generation of progress relied on pre-training that ran for 12 to 18 months, but differences between frontier models have narrowed and performance gains have begun to flatten. Token prices have fallen by around 80% a year since 2021, according to Epoch AI, and the article says frontier providers therefore work hard to persuade customers to adopt newer offerings to sustain premium pricing. It also distinguishes expansive uses such as coding, which can produce far more output than input, from compressive white-collar tasks such as summarisation, and warns that infrastructure built on legacy software-engineering assumptions may misread future demand.
The article identifies new risks as development shifts to post-training with human feedback, chains of thought, harnesses and agents. It flags recursive self-improvement, in which models contribute to their own training, as a possible move into an accelerating cost regime, and suggests that a consequential accident may be a more likely danger than deliberate harm as autonomy increases and verification becomes harder.
On geopolitics, the article argues that inference, which is constrained by memory and power, may favour China, whose installed generating capacity is cited as 3.4 times that of the US according to National Energy Administration and EIA data. It also says that several individually unlikely events could combine into a much larger catastrophic risk.
Financially, the article says AI may be widely useful yet still coincide with a costly reckoning across real estate, government, credit and technology. It highlights efforts by several countries to build alternatives to existing cross-border payment infrastructure, which could erode the US dollar's privileged position, and warns that using crypto to increase dollar supply and channel it into Treasuries could turn a dollar crisis into a bond crisis. The article concludes that AI retains value, but that cheapening tokens may make its industrial, geopolitical and financial consequences harder to manage than the technology itself.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.