Run the arithmetic on the threshold and it stops reading like a stretch goal. Huang's figure is half the engineer's own cost [4], and spread over twelve months it comes to roughly $20,800 of tokens per head per month [1]. The Databricks engineer whose $7,000 bill Ali Ghodsi held up as a marker of the new way of working is running at about a third of that pace [3][2]. Put that target on a leaderboard [4] and it starts selecting for behaviour nobody would defend in a design review.
Token spend is one of the easiest numbers in the stack to inflate without doing more work. TNW puts Opus 5 at five times Haiku's per-million-token price on July 2026 Claude pricing, with top-tier models generally running 5x to 10x their optimized counterparts [12][11]. The same piece recommends Haiku for quick lookups and subagent tasks, and Opus for complex coding sessions [15]. An engineer who ignores that and routes everything through the frontier model quintuples their score on identical output; the one who follows it takes an 80 percent cut to their ranking [5]. Context is the second dial: doubling the window can quadruple the computation required [13], so whoever never prunes a session outranks whoever does.
The conflict of interest is flagged in the source itself. TNW notes that critics can point out that Huang's employer books revenue directly when consumption rises [2]. That does not make the number wrong, but it does make it a supplier's metric, and the buyer is the one who has to defend it internally.
The bill has already landed at the firms that adopted first. Uber's CTO Praveen Neppalli Naga told The Information the company burned its annual AI budget in four months, about three times the planned rate [7][3]. Meta has added policies to cut wasteful usage and Adam Mosseri expects per-employee token limits [8]. Microsoft cancelled Claude Code licences and consolidated everyone under Copilot [10]. The AI spend leaderboard that made news earlier in the year has been shut down [9]. Harvey's spend is up roughly 12x to 12 trillion tokens a month [5]; the piece reports the multiple and not what the additional tokens produced, which is the gap the metric was supposed to close.
The dashboards built during 2026 to encourage employees to use more AI [14] are now the instruments used to ration it. Same tooling, opposite sign. What it measures in either direction is spend per engineer, a figure the supplier can already read off its own revenue line.
