Ken Griffin’s Citadel Securities built one of the most sophisticated trading operations on Wall Street. Now its own research arm warns that the hottest trade in markets faces a serious check. Surging expenses for compute, power and tokens have begun to reshape how companies deploy artificial intelligence. The shift marks a departure from earlier optimism about frictionless scaling.
Frank Flight, macro strategist at Citadel Securities, laid it out plainly in a June 2026 note titled “Tokenomics.” Even the most powerful technologies, he wrote, must pass through the prosaic discipline of cost curves, capacity constraints and marginal returns. Adoption is no longer about what frontier models can do in principle. It centers on the price and scarcity of the inputs required to run them at scale.
Compute. Power. Cooling. Memory bandwidth. Inference budgets. All real. All binding. Prices signal scarcity. They create incentives for substitution. They ration resources toward highest-value uses.
Evidence has piled up quickly. Amazon removed its internal token leaderboard. Microsoft cancelled Claude Code subscriptions. Multiple companies reported unexpectedly large token bills. These moves arrived as employees engaged in tokenmaxxing — maximizing AI tool usage to boost productivity scores and climb internal rankings. The party, as Citrini Research put it, is ending.
“Free AI is ending. Tokenomics is beginning,” Citrini Research declared in its State of the Themes: June 2026 report (Citrini Research). The firm, known for an earlier viral post on potential AI-driven job displacement, now argues that explosive token consumption has collided with rising monetization from labs and hyperscalers. Enterprises moved from tokenmaxxing to tokenpanic. Bills arrived. Behavior changed.
But Citadel takes a measured view. Flight does not forecast abandonment of advanced AI. Instead he sees bifurcation. Frontier models will concentrate among a narrower set of firms with strong balance sheets, deep research teams and operating domains where hard problems justify the expense. For the broader economy, simpler, cheaper models offer a more cost-effective path to productivity gains until physical constraints ease.
The Silicon Data LLM Expenditure Index has declined recently. The benchmark, which tracks price and mix of token usage, reflects users substituting toward efficient alternatives. As Citadel noted, the index can fall when prices drop, users shift to better models or the market diversifies away from expensive concentration (Citadel Securities). Lower unit costs in an elastic market can unlock additional volume even as spending composition changes.
This dynamic explains why falling token prices need not undermine demand for infrastructure. Volume grows. Composition shifts. Infrastructure providers still benefit. Yet the narrative has cooled. Markets once priced in ubiquitous, immediate AI. Reality demands selectivity.
Jim Esposito, president of Citadel Securities, acknowledged the tension earlier this year. “Our spend has been increasing,” he said at a Bloomberg event. “At the moment we’re getting a healthy return on that spend.” The firm’s trading business accelerated alongside AI investment. Costs rose in tandem. Returns followed. Not every organization enjoys that equation.
Citrini Research sees fresh opportunities in the reaction. The first AI trade focused on centralized compute. As those costs price in, asymmetry emerges in distributed inference, surrounding hardware and orchestration software. Edge AI — running models locally on devices — will not replace cloud systems. It complements them. “AI devices, running local models, will eventually be a thing,” Citrini stated. Nvidia has already doubled down on PC chips to support local inference (Business Insider).
But. The road remains uneven. Earlier this year Citrini’s “2028 Global Intelligence Crisis” memo roiled markets with warnings of mass white-collar displacement. Citadel pushed back hard in its February 2026 note. Unemployment stood at 4.28%. AI capex reached roughly 2% of GDP, or $650 billion. Nearly 2,800 data centers were planned in the U.S. Software engineer job postings rose 11% year-over-year. Displacement narratives overlooked these facts (Citadel Securities).
Technological diffusion follows an S-curve. Early adoption stays slow and expensive. Growth accelerates as costs fall and infrastructure builds. Saturation eventually sets in. Marginal returns diminish. Displacing white-collar work at scale would demand orders of magnitude more compute intensity than current levels. Rapid automation would raise compute demand, lift its marginal cost and create natural economic boundaries.
So companies ration. They monitor. They set budgets. They route high-value tasks to frontier models and routine work to lighter versions. Amazon, Uber and others have begun reassessing AI investments after token bills surprised (Business Insider). Microsoft’s decision to drop certain subscriptions underscored the point. Visibility into costs forces discipline. ROI tracking becomes prerequisite.
For investors the implications cut across tech giants. Microsoft and Amazon face pressure to demonstrate returns on massive AI infrastructure outlays. Their cloud businesses must convert capex into sustainable margins. Nvidia benefits from any expansion in total compute demand, even if per-token revenue compresses. Yet the market’s willingness to pay premium multiples for pure AI momentum has limits. Fewer clear winners may emerge than once assumed.
Flight remains constructive on AI’s long-term role as a productivity technology. The route to that value, however, looks more selective and cost-conscious. Asset prices will need to reconcile ambition with physical and economic constraint. Improvements in model efficiency, better cooling and power infrastructure can ease bottlenecks over time. But markets should avoid anchoring to visions of frictionless ubiquity.
Recent X discussions echo the theme. Traders and analysts highlighted Citadel’s note as institutional validation that economics now dominate capability. One post summarized it neatly: the hype centered on what AI could do. The reckoning centers on what it costs. Another noted best returns will flow to firms that lower costs and raise efficiency rather than chase the most advanced models.
Broader data supports caution. AI-adjacent commodities have climbed sharply since 2023. Electricity demand from data centers continues its steep rise. Planned facilities face regulatory, grid and supply chain hurdles. These factors reinforce Flight’s central observation. Physical limits bind. Prices matter. Substitution accelerates when marginal costs exceed marginal benefits.
Enterprises have responded with tiered access, approval gates and model-mix optimization. Frontier systems stay reserved for engineering, research and complex reasoning. Everyday tasks shift to open-source, distilled or faster variants. This bifurcation preserves high-end innovation while broadening productivity gains across the economy. It also tempers expectations for explosive, economy-wide labor displacement in the near term.
And the trading implications? Citadel itself demonstrates that targeted AI deployment can deliver healthy returns even as overall spend climbs. Its quantitative pipelines process massive data volumes. Gains in speed and insight offset infrastructure expense. Other market participants watch closely. Those who treat tokens like any scarce resource — with dashboards, budgets and clear ROI — stand to gain advantage.
The conversation has changed. Wall Street once fixated on capability roadmaps and release calendars. Attention now splits between those roadmaps and the invoices that follow. Tokenomics has arrived. Costs have surged. Adaptation follows. The AI trade endures. Its character, however, grows more disciplined by the day.
WebProNews is an iEntry Publication