TL;DR:
- Training frontier models has become more expensive, keeping frontier token prices high.
- Hardware improvements and smaller models matching older capabilities will drive token prices down.
- Cheaper tokens will still increase overall AI spending.
- More of the value shifts to the companies that apply tokens well.
Training frontier models requires large investments
As AI models have grown ever more capable, training them has become more expensive. To make models more capable, AI labs typically make them bigger, which requires more resources. This led to large investments in infrastructure for AI such as data centers that cause a ripple effect across the economy. In The infrastructure of AI, I discussed those effects in more detail.
High infrastructure costs lead to expensive tokens
This led to increasing costs per token for the latest, most capable models. Before it was taken down, Anthropic’s Fable 5 cost $50 per million tokens. Because these models are optimized for long-lasting agentic workloads, a single session can easily consume millions of tokens. This threatens to make the most capable models available only to people with capital, hindering democratization of this technology.
Do costs continue to rise indefinitely, or will falling prices threaten the companies building these models?
AI tokens will become a commodity like electricity or data bandwidth. Historically, frontier capabilities have trickled down into smaller, cheaper models in a matter of months. Inference costs drop steadily due to hardware improvements and algorithmic efficiency. The added value of additional model capabilities will get closer and closer until it gets too hard to accept the heavily increased costs. For most use cases, the market will settle for models that are good enough and comparably cheap to run.
This dynamic will force token prices down to marginal cost and threaten the valuations of frontier labs. Even when taking structural advantages beyond raw model size into consideration (e.g., Claude Code, Claude Cowork), the valuations face corrections within the next couple of years.
As discussed in Pushing AI integration, the systems built around AI (“AI harness”) will dictate the actual value created. The hunger for AI tokens is so huge that cheaper token prices will be overcompensated by increased consumption. This will increase revenues of AI model companies in the long term, but at lower margins.
AI becomes a general purpose technology
The companies at the core of technological revolutions almost never capture the majority of the economic value they create. The cycle typically goes like this: reach a technological breakthrough, deploy massive amounts of capital to make the technology generally available, then enter price wars as the marginal benefit of new models decreases and competitors offer similar products. At the same time, the ecosystem around the new technology benefits extremely. New companies built around the technology create products that have not been possible before.
One example of this is railroads. Track owners enabled an economic boom, but much of the lasting value accrued to the businesses that used rail transport to scale distribution and optimize their supply chains. The internet followed a similar pattern. Building networks and laying fiber required enormous capital, but the biggest businesses of the next era were companies built on top of that connectivity: search engines, marketplaces, and social networks.
AI will likely follow the same path: lower token prices will spread the technology further, while more of the value shifts to the companies that apply it well.