The double-duty idea in that essay works as a financing mechanism before it works as a product insight. Untangling the company's own data can be justified on internal savings alone, which is what clears the cash bar [4][5], and the system that results becomes the candidate product for merchants whose data sprawl resembles the company's own, with fewer resources to fix it [13]. The customer-facing bet is optioned rather than funded, which is how a business with limited adjusted EBITDA [2] can hold a hard hurdle rate and still build something speculative.
The hurdle is worth sizing. A $100 million free cash flow goal against revenue approaching a billion implies a margin near 10 percent [1], and on the stated sequence, a target set a year after an early-2024 return with three years to reach it, the clock runs to roughly early 2028 [2]. Set that against the cost line he singles out: multiyear contracts for coding tools such as Codex, Cursor and Claude that can lock in tens of millions [6]. Ten million dollars of committed spend is a tenth of the whole target [3], and Uber, according to the essay, went through its entire AI coding budget in four months, with its COO questioning whether the return justified it [8]. Nor does the unit cost trend rescue you, since larger models can cost more to run rather than less as the technology matures [7]. That is why the test shows up in procurement, with OpenCode, open-source and model-agnostic, keeping the exit cheap when a retail-data job does not need a frontier model [9].
A cash hurdle like this can look like the way incumbents underinvest during a technology transition, since the returns on learning often arrive after year three. The essay's implicit answer is the cost of the opposite discipline: a third of US hiring managers who cut a role because of AI have since rehired for that role or a similar one [12], so spending that has to be repeated does not count as a saving. It is also worth noting what the record does not include. This is one contributor's account, published as opinion [15], and it states the rule without publishing the ledger, giving no total AI spend and no progress figure against the target [4].
The tradeoff left implicit is a classification problem. If a three-year cash number is the gate, work whose payback lands in year four has to enter the budget as cost reduction and earn product status later, so the decision that matters is who gets to call a project internal cleanup rather than product development. A founder-CEO a year into a turnaround mandate can arbitrate that himself [1]; a divisional owner copying the move inherits the discipline without the authority to reclassify. The pricing choice compounds it. Having declined to reprice around AI consumption because transaction fees already tie revenue to merchant success [10], at a moment when software firms were being pushed toward consumption models [11], the company recovers its internal AI build only when merchants transact more, not when they consume more tokens. Whether the proving ground pays off will show up in adoption figures, not usage figures.
