AI spend nearly tripled in six months. OpenAI's own data says tokens do not predict revenue.

OpenAI released a study last week covering more than 1,500 organisations and found that revenue per employee was not meaningfully associated with an employee's token output ^[1]. Over roughly the same window, spending on the thing that does not predict revenue went vertical: Ramp's AI Index, drawn from more than 70,000 firms using its corporate card and bill payment platform, puts median AI spend among the top 1% of businesses at $7,400 per employee in July 2026, against $2,590 in January ^[2][3][4].

That is a 2.9x increase in six months ^[5]. The top 10% moved from $281.23 to $650, a 2.3x increase ^[3][4][6]. The gap between the two groups widened from about 9x to about 11x, so the heaviest adopters are pulling away from a field that is already committed ^[7].

Goldman Sachs expects token consumption to multiply 24-fold to 120 quadrillion tokens a month between 2026 and 2030, while semiconductor suppliers deliver cost reductions of 60% to 70% per year per token for inference ^[8][9]. Read that alongside Ramp's numbers and the volume curve is steeper than the dollar curve: spend nearly tripling while unit prices fall by roughly two thirds a year implies consumption growth well above 2.9x ^[10].

The failure cases are already on the record. Uber's COO said in May that the company had run through its AI budget in four months ^[11]. The Financial Times reported in July that Amazon incurred $1.8 million in cost overruns after a failed Claude Sonnet deployment ^[12]. OpenAI chairman Bret Taylor has predicted that companies will stop worrying about tokens ^[13], but G2's research finds 80% of software buyers now give developers or technical teams a token or LLM usage budget ^[14]. Buyers are treating this as a permanent line item, tracked like any other.

G2 is at least measuring something other than volume. Chief innovation officer Tim Sanders told Forbes the company has spent over $1.27 million on tokens in 2026, the equivalent of 970 billion tokens ^[15], or roughly $1.31 per million tokens ^[16]. Claude Code and Cowork went from 17% of that spend in January to 79% in July, a 62 point shift in where the money goes inside a single year ^[17][18]. G2 runs an internal AI Efficiency Index, published as Tokens / Dollar * ln(1 + Tokens / 1,000,000), to identify who gets the most from each token ^[19]. Sanders compares AI to electricity and says the company looks for time savings, cost savings and the ability to increase income rather than maximum consumption ^[20], and that "we're not just measuring how much time we saved; we're also measuring our velocity in delivering outcomes against our key objectives" ^[21]. The superuser figures G2 released do not parse as published: its cofounder and CTO is listed at 9.54 million tokens at $1.09 tokens per dollar, and its VP of engineering at 9.55 million tokens at $1.34 million tokens per dollar ^[22]. Units that change inside one table are a sign of a measurement layer younger than the budget it is being used to justify.

The case for restraint comes from Zapier's AI Workflow Index, also released in July and covering more than 1,500 organisations: among leading adopters, AI appears in just 18% of workflow steps, while 82% still run on code and logic ^[23].

Two things to watch in the next budget cycle. Whether finance asks for a denominator at all, or approves growth on gross token volume, which is the one number OpenAI's dataset says carries no signal on revenue per employee ^[1]. And whether the 24-fold consumption forecast ^[8] arrives with the per-token price declines that make it affordable ^[9], because if it does not, the Uber and Amazon overruns stop being anecdotes ^[11][12].