AI Tokenomics: Have we let the Meter set the ambition?
For a few months now, something has been bothering me about our AI conversations.
In leadership workshops, I’ve watched discussions move quickly from choosing the right problem to explaining how token costs will be controlled. Sometimes we’re debating the meter before we’ve properly agreed what the transformation should achieve.
I’ve noticed something similar at a couple of recent AI hackathons, where the approach to tokenomics featured among the key evaluation criteria.
Frankly, I find that shift bewildering.
AI has made it to the boardroom. (Technology in a boardroom? Surely not.) The pressure to demonstrate returns on those investments is understandable. So is the need to control spending.
But I wonder whether we’ve given one slice of AI’s economics rather too much importance.
By tokenomics here, I mean the pricing and management of the tokens consumed by AI models. There’s useful work in understanding those economics. My concern is what happens when they begin to set the ambition of an AI strategy.
Have we chosen a problem worth the tokens?
Picking the wrong use case can waste months of effort, integration work and organisational attention. The tokens consumed along the way are part of that waste. Making those tokens cheaper does little to rescue the original decision.
Take a customer-service application. We could reduce the tokens used in each response and report an impressive improvement. But if the shorter answers leave customers confused, bring them back for another conversation, or send more work to a human agent, what have we actually improved?
I’d want to know whether we’re resolving more problems correctly, with less customer effort, and at what cost compared with the existing service. That includes model usage, data preparation, integration, human review, rework and ongoing operations. I’d also want to establish whether that problem was worth addressing with AI in the first place.
Token spend directly measures one component of cost. It becomes a proxy when we ask it to stand in for business value. Token spend directly measures one component of cost. AI FinOps is a much bigger, layered financial discipline, spanning infrastructure, compute, data, integration, ongoing operations and the business value those investments deliver. Getting those economics right matters. Yet tokenomics, one small piece of that picture, seems to have claimed most of the airtime. Does it simply roll off the tongue better? Somewhere along the way, the cost of consuming a model appears to have become a proxy for the value of an entire AI initiative. That’s quite a promotion.
This is where I see a Goodhart’s Law problem: optimise a convenient measure hard enough, and it can stop serving the objective we actually care about. A team rewarded for lower token consumption has a reason to produce a smaller bill. Whether the customer gets a better outcome needs a separate assessment. [1]
Reasoning models add another wrinkle. We may pay for internal reasoning as well as the answer we see. There’s a distinction here, though: hidden reasoning doesn’t necessarily mean hidden billing. A message saying “Thought for 9s” reports elapsed time; it isn’t a token-cost statement.
Even perfect visibility into consumption would leave the larger question unanswered: was the result useful?
The economics won’t stand still
There’s another reason I’m uncomfortable building a strategy around today’s token prices. The price of a given level of capability has already moved dramatically.
The economics have continued moving. In a September 2026 analysis, Epoch AI estimates that the cost of achieving a given level of benchmark performance has fallen by around 47% per quarter since 2023—roughly a thirteenfold reduction per year. That tracks the most cost-effective models available, so individual organisations won’t automatically see equivalent savings. Still, it makes me uncomfortable building a long-term AI strategy around today’s unit prices. We could spend months designing around a constraint while the economics are already moving beneath our feet. [2]
It’s a historical comparison, rather than a guarantee about the next frontier model. Still, it should make us cautious about treating today’s unit prices as permanent strategic constraints.
Storage is the analogy that comes to my mind. Falling prices expanded what we could afford to retain and do with data. Storage bills still deserve attention, especially at scale. The possibilities expanded alongside the economics.
I suspect something similar will happen with tokens. Cheaper inference may encourage so much additional use that total bills rise. Agents making repeated calls can amplify that effect. That makes the value of the work being done even more important.
Our methods are evolving too. Think of the progression from individual prompt engineering, to shared prompt libraries, to skill files that package instructions, reference material and repeatable workflows. These approaches coexist, but each offers ways to make useful knowledge easier to reuse.
A newcomer can begin with an established method. A well-designed skill can also load relevant material as needed, rather than putting everything into the model’s context upfront. [3] When those practices reduce corrections and repeated attempts, they can improve efficiency alongside quality. The actual savings still need measuring; a markdown file alone guarantees neither expertise nor fewer tokens.
That interests me: better ways of doing the work can improve its economics as they mature.
The intelligence is moving closer
My first PC, roughly twenty years ago, had 256 MB of RAM. My current MacBook has 48 GB. RAM alone tells us very little about model performance, but the comparison does remind me how dramatically the devices around us have changed.
And then there’s the computing power and collection of sensors squeezed into a smartphone.
Smaller models and increasingly capable devices are expanding the work AI can do locally. This is already happening.
For autonomous equipment, the possibilities become particularly interesting. Sense at the edge. Decide at the edge. Where local processing can meet the required performance and reliability, avoiding a network round trip can matter enormously. The value of a timely decision could extend well beyond the economics of generating a text response.
Sovereign AI adds a different consideration: control. With geopolitical uncertainty and valuable organisational knowledge in the picture, I expect enterprises to think harder about where their AI runs, who can access it, and how their business context is protected.
That context may become a source of differentiation. Fine-tuning is one route; connecting models to trusted organisational information and workflows offers others.
The deployment choices deserve care. Sovereignty can be supported by managed cloud services as well as infrastructure an organisation owns. Location, operational control, capacity and cost involve trade-offs; sovereignty alone promises neither lower latency nor a shift to capital expenditure. [4]
I expect a mixture of approaches. Suitable workloads run locally, while more demanding work goes to cloud resources where privacy rules and connectivity allow it. That requires deliberate routing and integration. It’s more involved than assuming every device can automatically burst into the cloud.
For organisations buying infrastructure, more expenditure may move into upfront investment. There will still be power, maintenance, engineering and hardware renewal to consider. Each additional prompt consumes energy and finite capacity. Once that capacity fills, expansion has a price.
So I’d be careful with the phrase “zero marginal cost”. On existing hardware with spare capacity, the extra cash expense of another prompt may be very small. The psychological difference could still be significant: a team exploring what it can do with capacity it already has, without a separate API charge attached to every request.
The relevant economics increasingly include utilisation, reliability and the total cost of delivering a useful result.
My dishwasher has a small part in this
Of course, plenty of enterprise tasks can tolerate a wait. They don’t need the responsiveness of autonomous equipment.
There are practical choices here too. Anthropic’s Message Batches API, for example, offers a 50% discount against standard API prices for asynchronous processing. [5]
I’ve done enough rough household electricity calculations to appreciate the attraction of running the washing machine and dishwasher during cheaper off-peak periods. The lighting and air conditioning still run when they’re needed.
The AI version means understanding which work needs an immediate response and which can wait. The provider handles the batch-processing schedule; we’re accepting delay for a lower price, rather than choosing a household-style night tariff.
OK… I’ve oversimplified it a bit. Data handling, completion deadlines and reliability need more attention than that analogy gives them.
But the principle feels sensible: choose the processing approach around the needs of the work. I can see a useful engineering decision there. I’m less convinced it deserves to define the ambition of the transformation.
The mindset may outlive the price
This is the part that concerns me most.
Technology can move quickly. Organisational habits tend to take longer. If we teach people to approach AI primarily through the fear of consuming too much, that instinct could remain long after the economics have changed.
We could spend years designing around a constraint that has become much less important.
There’s something odd about shackling our legs before we’ve worked out the jump we want to make.
Tokenomics will still have work to do. As the centrepiece of AI strategy, though, I think its time in the limelight should be over.
[1] David Manheim and Scott Garrabrant, [Categorizing Variants of Goodhart’s Law](https://arxiv.org/abs/1803.04585), 2018.
[2] Luke Emberson and David Roodman, The plunging price of thought, Epoch AI, 22 September 2026.
[3] Anthropic, [Agent Skills overview](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview).
[4] AWS, [Evaluate sovereign solutions using a data-driven approach](https://docs.aws.amazon.com/wellarchitected/latest/digital-sovereignty-lens/dsperf01-bp01.html).
[5] Anthropic, [Batch processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing).
Comments