Sometime in the past year the price of a million tokens at a given level of capability fell from roughly ten dollars to about two and a half, and most operators experienced this the way you experience weather: pleasant, free, not yours to question. Per-token prices are now down something like 99.7 percent from the GPT-3 era. The reasonable reading is that the technology got radically cheaper to run. The more useful reading is that four companies decided to lose money on every token until the others stop trying.
OpenAI, Google, Anthropic, and Meta are pricing inference below what it costs them to serve, funded by venture capital and by hyperscaler cross-subsidy, in a race to own the default before the market settles. The buildout behind that race is the largest concentrated infrastructure cycle in the history of technology. Big Tech is on track to spend about US$725 billion on AI infrastructure in 2026, up from a February estimate near US$610 billion, a figure roughly one and a half times the size of Singapore’s entire economy. JPMorgan puts cumulative spend at US$5.5 trillion through 2030, most of it debt-financed, and calculates that earning a mere 10 percent return on it would require about US$650 billion in annual revenue, the equivalent of US$35 a month from every iPhone user on earth, in perpetuity.
The price you pay for a token today is a marketing budget, not a cost of goods.
So here is the proposition, stated plainly: the price you pay for a token today is a marketing budget, not a cost of goods. That gap between the bill and the revenue is the subsidy you are currently enjoying, and it is real, and it is not promised to you past the quarter in which one of the four blinks.
The honest counter-argument is strong and worth stating at full strength. The productivity gains are not hypothetical; operators are clearing real work with these tools now. The people calling a bubble have been calling it for two years and have been wrong every quarter, while compute demand kept outrunning supply. And sitting out is its own expensive mistake: the firm that waits for the economics to “make sense” cedes two years of compounding workflow advantage to the firm that did not wait. Cheap capital paid for the railways, and the railways stayed after the speculators left.
The warning lands on a narrow target: build a business whose survival depends on the cheap thing staying cheap, and you have written the subsidy into your margin. Alibaba’s Joe Tsai, no bear on AI, told the HSBC summit in Hong Kong last spring that data centres were being built on spec, with no tenant lined up, which is the supply side of the same imbalance. The price normalisation, when it comes, comes upward, and the operator exposed is the one who priced a product, signed a customer, and staked a margin on API rates that were always a customer-acquisition cost wearing the costume of a unit cost.
For a cè reader the local tell arrived in June, quietly, in the Monetary Authority’s survey of professional forecasters. The economists trimmed Singapore’s 2026 growth forecast to 3.5 percent and named, for the first time in plain language, “a bursting of the AI bubble” or a slowdown in AI capital expenditure as a top downside risk, alongside the Middle East. They also named a sustained AI upcycle as the chief reason to be optimistic. Both can be true. Singapore’s good year is levered to AI-hardware flow; the same lever that lifts the city is the one the forecasters are now watching for slack.
Use the cheap thing hard, wire it into a workflow you own, and assume the price you pay is the lowest it will ever be.
The posture that survives both outcomes is unglamorous. Use the cheap thing hard, wire it into a workflow you own, and assume the price you pay is the lowest it will ever be. Own the process, the data, and the customer relationship, the parts no vendor can reprice on you, and treat the model underneath as a commodity you will swap the morning the bill arrives. The subsidy is a gift. A gift is not a plan.