The Great Token Retrenchment: Prudence, Panic, or the First Honest Reckoning with AI’s Value?

July 16, 2026
James Lever

Barely six months ago, the corporate zeitgeist rewarded conspicuous consumption of artificial intelligence. Meta and Salesforce exhorted staff to consume tokens with abandon; one Meta employee even built an internal “Claudeconomics” dashboard ranking the company’s 250 heaviest users, who collectively burned through more than 60 trillion tokens in a single month before the leaderboard was quietly retired. Accenture went further still, reportedly warning employees that insufficient AI usage could jeopardise their promotion prospects.

The reversal has been swift and, in places, faintly humiliating. Accenture is now attempting to stop consultants from depleting shared token reserves on trivial chores such as converting PDFs into slide decks. Uber exhausted its entire 2026 AI coding budget by April, four months into the year, and has imposed a cap of $1,500 per employee per tool per month. Citi has urged staff towards less powerful models and, according to internal documentation reported by 404 Media, restricted access to the latest frontier releases outright. AT&T is limiting access to GitHub Copilot for some employees, Walmart has capped usage of its in-house agent, and Amazon has scrapped the leaderboard that ranked staff by AI consumption after it was, predictably, gamed.

The commentariat has already coined the inevitable neologism: after the era of “tokenmaxxing” comes the era of “tokenminimising”. The strategic question, however, is whether this retrenchment represents something intellectually substantive, a genuine interrogation of whether tokens consumed bear any reliable relationship to value produced, or whether it is simply the oldest story in corporate finance: the CFO discovering an uncapped, usage-based cost line and reaching for the tourniquet.

The honest answer is that it began as the latter and is maturing, unevenly, into the former.

The arithmetic that broke the budget

To understand why finance departments have intervened so abruptly, one must appreciate a paradox that has caught even sophisticated organisations unawares. The unit price of intelligence is collapsing. Analysis of 2.4 billion enterprise API calls found the blended cost of AI fell 67 per cent year on year, from $18.40 to $6.07 per million tokens between the first quarters of 2025 and 2026. Stanford’s HAI research documents inference costs for a given capability level falling roughly 280-fold over two years. By any conventional procurement logic, AI should be getting cheaper.

Yet the FinOps Foundation’s 2026 State of FinOps report found that 73 per cent of enterprises exceeded their original AI cost projections. Total enterprise AI spending surged over 300 per cent in 2025 even as per-token prices fell. Gartner now forecasts worldwide AI spending of roughly $2.59 trillion in 2026, a 47 per cent annual increase.

The explanation is the Jevons paradox, the nineteenth-century observation that efficiency gains in a resource tend to increase, rather than reduce, its total consumption. Satya Nadella invoked it explicitly when DeepSeek’s low-cost models arrived. Cheaper tokens do not shrink the bill; they enlarge the appetite. New workloads become viable, existing workloads run more frequently, and organisations upgrade to more capable, more expensive frontier models rather than banking the savings.

The decisive accelerant, though, has been the shift from chatbots to agents. A conventional chat interaction consumes perhaps two thousand tokens in a single call. An agentic workflow, in which an orchestrator decomposes a task, invokes tools, validates outputs, retries on failure, and resends accumulating context at every loop, consumes between five and thirty times more tokens per task. The ROI models that justified most enterprise deployments were built on chatbot-era assumptions; the production invoices arrived an order of magnitude higher. Compounding this, vendors have migrated from flat-rate subscriptions towards usage-based billing, a transition that converted predictable per-seat line items into volatile, consumption-driven exposure almost overnight. Citi’s restrictions followed precisely such a pricing shift in June.

Seen through this lens, the caps are unremarkable. They are what any competent finance function does when a cost line triples without a corresponding revenue narrative. Even Sam Altman has conceded the point, telling CNBC in June that scepticism about AI returns is currently the fairest criticism of the industry, and that customers exhausting their annual budgets mid-year had become one of the most common complaints he hears. When the vendor concedes the question is fair, the question deserves an answer.

The more uncomfortable question underneath

But cost control alone does not explain the texture of what is happening, because a second, more corrosive doubt has crept into the conversation: whether the tokens were buying anything worthwhile in the first place.

The productivity evidence is genuinely bifurcated. Controlled task-level studies consistently show material gains: roughly 14 to 15 per cent in customer support, around 26 per cent in software development, and considerably more in structured writing tasks. These findings are robust and replicated. Yet at the enterprise level the picture inverts. A landmark NBER survey of nearly 6,000 senior executives across four countries found that although 69 per cent of businesses actively use AI, around nine in ten of those firms report no detectable impact on productivity or employment. McKinsey’s data tells a similar story, with only a small minority of organisations attributing meaningful earnings impact to AI. Workday’s January 2026 study of 3,200 business leaders found that while most employees report saving one to seven hours weekly, nearly 40 per cent of those gains evaporate immediately into rework and correction, the phenomenon now inelegantly termed “workslop”.

This is the context in which the Accenture anecdote becomes strategically revealing rather than merely comic. Employees using frontier reasoning models to reformat PDFs were not being lazy; they were responding rationally to incentives that measured activity rather than outcome. When an organisation signals that token consumption is a proxy for AI maturity, it manufactures consumption. It does not manufacture value. The leaderboards, the promotion threats, the mandated adoption targets: all of these optimised the numerator of a ratio whose denominator nobody had defined.

So yes, the value of tokens versus output is finally being questioned, but largely because the invoices forced the question. Necessity, as ever, is the mother of measurement.

The tell: how firms are rationing matters more than whether they are

The strategically literate response to this moment is visible in the divergence between blunt and intelligent interventions, and it is here that leaders should focus their attention.

The blunt instrument is the uniform cap. It is administratively simple and politically defensible, but it carries a hidden cost that the SemiAnalysis data exposes rather elegantly: enterprise AI consumption follows a steep power law. In conversations with more than fifty enterprises, SemiAnalysis found monthly caps ranging from $250 at an aerospace manufacturer to roughly $2,000 at Workday and Stripe, but also found that at most organisations the majority of employees never approach the limit at all. Consumption, and plausibly value creation, concentrates in a small cohort of power users. A uniform cap therefore constrains precisely the individuals most likely to be generating disproportionate returns, while leaving the median user, and the median waste, untouched. One firm quoted in that research declined to impose limits at all on the explicit grounds that its top performers were its heaviest token consumers, and Databricks has similarly kept engineering budgets uncapped. Box’s Aaron Levie, meanwhile, permitted himself a note of vindication for never having celebrated consumption metrics in the first place.

The intelligent instruments look rather different. They include model routing, whereby low-complexity tasks are automatically directed to cheaper or open-source models while frontier capacity is reserved for genuinely demanding reasoning, an approach that has allowed some teams to cut API costs by 40 per cent with no product changes whatsoever. They include gateway and monitoring products of the sort Microsoft and Databricks have launched, chargeback mechanisms that assign AI costs to the business units consuming them, and, most tellingly, the emergence of a new governing metric: cost per successful outcome rather than cost per token. A more expensive model that resolves 90 per cent of tickets on the first attempt can be cheaper per resolved ticket than a bargain model that resolves 40 per cent and escalates the remainder to humans. One airline reported by SemiAnalysis now ties token allocations directly to the projected revenue of the specific project consuming them, treating tokens as it would travel or contractor costs. The Linux Foundation’s newly launched Tokenomics Foundation is attempting to formalise such practices into open standards.

This is the substantive shift concealed inside the headlines. The caps are the visible symptom; the underlying transition is from governing AI as a technology rollout, where adoption itself was the KPI, to governing it as a factor of production, where the only question that matters is the ratio of value created to intelligence consumed.

What this means for strategy

Three conclusions follow for leadership teams watching this unfold.

First, resist the false binary. This is simultaneously CFO discipline and a legitimate value reckoning, and treating it as only the former invites a damaging pendulum swing in which rationing throttles exactly the productivity gains that justified the spending. The Deloitte finding that only 28 per cent of global finance leaders can point to clear, measurable value from AI investment is not an argument for retreat; it is an argument for instrumentation.

Second, the unit of governance must become the workflow, not the employee. The organisations avoiding budget surprises are those that modelled token volume per workflow type, with realistic loop counts and context depth, before finalising architecture. Per-head caps are a stopgap; per-outcome economics are a strategy.

Third, expect the competitive frontier in enterprise AI to migrate from model capability towards cost-discipline tooling: routers, gateways, observability, and FinOps practice. In 2025, roughly a third of FinOps practitioners were responsible for AI spend; in 2026 the figure is 98 per cent. The organisations that develop genuine fluency in token economics, treating intelligence with the same allocative rigour applied to energy or capital, will be able to keep spending aggressively while their competitors oscillate between euphoria and austerity.

The tokenmaxxing era rested on a category error: that consumption of intelligence was evidence of its productive use. The rationing era risks the mirror-image error: that restricting consumption is evidence of discipline. The firms that emerge advantaged will be those that decline both fallacies and instead answer the only question the past six months have made unavoidable. Not how many tokens were burned, nor how few, but what, precisely, did they buy?