The brutal math: fast-rising token costs, flat productivity
The current AI economics problem is the widening gap between rapidly compounding token costs and much slower productivity gains, which is creating a serious sustainability risk for enterprise AI adoption as pilots scale into real workloads and budgets come under scrutiny.
The cleanest snapshot of this gap comes from a software startup where venture capitalist Chamath Palihapitiya says his CTO reported that their AI token costs are doubling every forty-five days, while productivity gains sit at “maybe 5% max.” That is not hype; it is a unit economics red flag. In plain terms, engineers are coding a bit faster, but the bill is growing like a high-interest credit card. Worse, those early gains look asymptotic: the team has already captured the easy wins from AI coding assistants, and now every extra improvement demands a disproportionate number of tokens.
If token spend keeps compounding faster than output, this is not transformation; it is a slow-motion margin crush that most boards will not tolerate once the novelty fades.
Why token-based pricing hides the real AI cost
Enterprises obsess over per-token discounts, but AI cost calculation based on sticker price per token is misleading. The number that matters is cost per completed task and the value of that task, not how cheap your tokens look on paper. When Databricks benchmarked coding agents on real internal engineering tasks, it found that models with cheaper tokens often cost more overall because they used more tokens and failed more often.
One clear comparison: a frontier model variant charged more per token than its sibling, yet ended up cheaper per task. While one model was around 1.7 times cheaper per token, it consumed more tokens, completed only 81 percent of tasks versus 87 percent, and ended with a higher price per task. As the company bluntly put it, “Cheaper per-token does not imply cheaper per-task.” That should be engraved on every enterprise AI ROI spreadsheet.
If you do not measure task completion rates and downstream business outcomes, even a low-per-token model can quietly become the most expensive tool in your stack.
Open source AI pricing and the lock-in trap
There is an uncomfortable contrast between what enterprises are paying and what they could pay. Open-weight models are now reported to be roughly ten times cheaper per token than closed frontier models. Boris Renski argues that open-source models are only four months behind frontier systems in capability at a fraction of the cost, and that most enterprise buyers are paying premium prices for “commodity enterprise features” such as identity integration, connectors, and observability rather than for unique intelligence.
That premium looks worse when you run the numbers at scale. One serverless inference platform claims that a team consuming around 100 billion tokens per month would face an annual AI bill above USD 1,500,000 (approx. RM6,900,000) with frontier models but could instead pay a fixed USD 90,000 (approx. RM398,000) per year by using an optimized open-weight model, a difference of over USD 1,460,000 (approx. RM6,460,000) in savings. That is not a rounding error; that is headcount-level money.
Renski warns that multi-year contracts with frontier labs risk turning into an Oracle-style lock-in, where the fabric of the business is so tied to proprietary vendors that migration becomes almost impossible. Jonathan Bryce calls paying ten times more for a four-month capability lead “not an enterprise AI strategy” but “a very expensive form of lock-in.”
What Databricks’ benchmarks reveal about enterprise AI ROI
If Chamath’s anecdote shows the pain, Databricks’ internal benchmark shows the mechanics. The company built its own coding benchmark based on real engineering tasks from its codebase to see how AI agents perform in realistic conditions, and it encourages other companies to run similar tests over their own code. In those experiments, an open-weight model, GLM 5.2, landed in the top capability tier and was statistically tied with a frontier model on quality while costing USD 1.28 (approx. RM5.66) per task versus USD 1.94 (approx. RM8.58) per task for the frontier model.
This is the core enterprise AI ROI problem: a cheaper model per token can still be more expensive per task, and an open model can match frontier quality at meaningfully lower task cost. In response, Databricks built a tool called Omnigent, a wrapper that can combine multiple coding agents and swap them in and out. That is a quiet but important design choice: it treats models as interchangeable components rather than fixed platforms, which is the opposite of the lock-in logic that many current AI contracts encourage.
The lesson is blunt: until you benchmark against your own tasks and instrument per-task cost, your enterprise AI ROI is guesswork wrapped in hype.
The coming AI cost reckoning
These stories are not isolated. Uber’s CTO has already admitted that the company burned through its entire annual budget for tools like AI coding assistants in just four months. Chamath expects that “everybody in the next three or four years” will face a similar reckoning when soaring token spend collides with modest productivity improvements. The pattern is simple: early pilots look good, costs quietly compound, and only later do executives ask whether the gains justify the burn.
Meanwhile, one open-weight platform argues it can reduce frontier-class inference expenses by about 94 percent, flattening token-driven variability into a predictable annual rate. Combine that with warnings that enterprises may end up viewing their long LLM contracts the way they see legacy database licenses, and a clear picture emerges: this phase of AI adoption is as much about financial engineering as it is about model engineering.
The conclusion is unavoidable. If enterprises keep chasing frontier models without fixing AI cost calculation, measuring per-task economics, and pushing toward open infrastructure, they are not building an AI strategy. They are signing up for an AI tax. The winners in this wave will be the companies that treat tokens like a scarce resource, not a magic commodity, and demand that every extra dollar of AI spend earns more than five cents of productivity in return.
