Spending $100,000 a year on tokens now is a bet on living two years ahead of the market

The argument Lenny Rachitsky is making is simple enough to state in one sentence and consequential enough to deserve more scrutiny than it usually gets. If you are willing to spend $100,000 a year on tokens right now, he says, “you are living the way somebody in 2028 is going to live, because by then it’ll be really cheap.” That is a dated, falsifiable claim. It either holds or it does not, and there is a roughly two-year window in which to find out.

The structure of the bet is temporal arbitrage. Token costs have been falling fast enough that what requires meaningful capital expenditure today is expected to be commodity spending within a short horizon. Anyone willing to pay the current price accesses the future workflow before that workflow becomes affordable to everyone. The edge is not the capability itself. The edge is the lead time: the compounding fluency, the adapted processes, the institutional knowledge built up during the period when most organizations are still watching from the sideline because the price point looks prohibitive.

What the argument requires to hold is a continued, steep decline in inference costs over the next two years. That trajectory has been consistent enough to invite confidence, but it is not guaranteed. Model improvements can temporarily increase cost per useful output even as raw token prices fall, if the tasks being attempted grow proportionally harder. Workflow complexity that makes sense at $100,000 a year can become a liability if pricing dynamics shift unexpectedly. Rachitsky is not predicting a specific price. He is predicting a relationship between today’s frontier cost and tomorrow’s baseline, and that relationship depends on variables no single observer controls.

If you're willing to spend $100,000 a year right now in tokens, you are living the way somebody in 2028 is going to live, because by then it'll be really cheap. Lenny Rachitsky

There is also the question of what “living in 2028” actually means in practice. The claim implies more than a higher token budget. It implies that exposure to current frontier capabilities, at scale and over time, builds the kind of workflow fluency that will define competitive operations once those capabilities are cheap. That is a plausible mechanism, but it is not an automatic one. Spending at that level produces real fluency only if the spend is directed, instrumented, and learned from. High token spend that is diffuse or undirected does not create the compounding advantage the arbitrage argument assumes. The alpha is in the workflow adaptation, not the line item.

That distinction matters because it changes what the experiment actually requires. An organization that routes $100,000 a year through AI tooling without changing how work is structured, how outputs are evaluated, or how teams are organized around new capabilities is not living in 2028. It is paying 2026 prices for 2024 habits. The bet Rachitsky describes is not purely financial. It is operational, and the operational half is harder to buy than the token budget.

There is also a selection question the call does not fully resolve. The teams most likely to spend at that threshold are also the teams most likely to have the technical and organizational capacity to learn from the spend. The arbitrage opportunity, if it exists, may be less available to the organizations that need it most and more available to those already positioned to exploit it. That does not invalidate the call. It narrows the population for whom it holds.

What makes the 2028 comparison point useful is precisely that it is checkable. Rachitsky is not gesturing at a vague future in which AI changes everything. He is making a specific claim about the relationship between today’s cost curve and a two-year-out baseline. That specificity is what separates a forecast from an observation dressed as a forecast. Whether a team spending at his threshold in the next twelve months actually internalizes the future workflow, or simply incurs the cost without the organizational learning, will not be visible from the outside until the window closes. Anyone willing to run the experiment now should be tracking not just what they spend, but what they learn, so that 2028 gives them an honest answer.