John Zito, co-president of Apollo Asset Management, delivered a blunt message at the Morgan Stanley US Financials Conference this week. The metric driving Wall Street’s AI valuation debate is the wrong one. Speaking to Bloomberg on Wednesday, Zito said the industry has fixated on token prices while ignoring a far more important measure: the cost of intelligence per unit of output.
The setup
Zito’s criticism lands at a moment when AI infrastructure spending has become one of the most closely watched numbers in markets. Companies across the S&P 500 have ramped capital expenditure on data centers, GPUs, and cloud computing. Investors have used token consumption — the volume of AI queries and responses flowing through large language models — as a proxy for growth.
Zito called that approach flawed. “I think tokenmaxxing and token talk is — it’s a lot of BS, honestly,” he said. “If you look at per unit of knowledge and cost per unit of knowledge, prices are collapsing. Prices are collapsing per unit of IQ, if you did it that way.”
The analogy he offered was straightforward. A laptop in 2026 costs roughly what one did in 2010. It is roughly fifty times more capable. Applied to AI, the cost of delivering a fixed amount of intelligence is falling rapidly. Total spending still rises because usage is exploding.
Goldman Sachs strategist Rich Privorotsky made a similar argument five days earlier. He wrote that the relevant economic metric is not token volume but “useful task completion per watt and per dollar.” When Apollo and Goldman independently arrive at the same framework within a week, institutional consensus is shifting in real time.
Key numbers
| Metric | Detail |
|---|---|
| AI private credit at Apollo | ~$250 billion in financing discussed |
| Enterprise AI cost concern | 60% of IT executives per UBS Evidence Lab |
| Uber engineer token cap | ~$1,500 per month |
| Silicon Data LLM Token Expenditure Index | 6 consecutive down days as of early June |
| GitHub Copilot billing change | Usage-based pricing effective June 1, 2026 |
| AI dollar efficiency | ~18 cents of every dollar reaches stable product per data spanning 2,444 firms |
The spending problem is real. The metric describing it was wrong. That tension is where the AI trade heads next.
What to watch
Enterprise behavior is splitting into two tracks. High-value users — quantitative trading firms like Citadel and Jane Street — continue paying premium prices for frontier models because their return on investment is substantial. Everyone else is downshifting.
Uber capped AI tool usage after burning its full 2026 budget by April. Walmart restricted an internal AI agent that helps employees with spreadsheets. Multiple enterprises told UBS they are cutting other IT spending — cloud services, external contractors, even headcount growth — to fund AI rather than throttling usage.
The pattern is not retreat. It is cost containment. Companies are adding guardrails: alerts, pooled tokens, and model downshifting. Open-source and Chinese models now deliver near-frontier capability at ten to twenty-five times lower cost. Cursor’s latest model matches frontier coding performance at one-tenth the price per task.
Gartner research adds a counterpoint worth noting. Even a 90% collapse in inference costs may not reduce enterprise AI spending overall. Agents consume tokens faster than prices fall. A falling unit cost multiplied by exploding volume still produces a larger bill. That is the lived experience in the UBS checks.
Bottom line
For investors, the takeaway is nuanced. The semiconductor and infrastructure complex has been priced for token demand climbing indefinitely at the frontier. If Zito’s “use-case economy” materializes, a meaningful share of that volume shifts to commodity inference. Cheaper chips, open models, and local hardware take share. Volume can keep growing while the dollars, and the margins, pool somewhere else.
The Silicon Data Token Expenditure Index has printed its longest losing streak since January. Citadel’s own desk attributes part of the drop to adoption shifting toward cheaper models rather than falling usage. The numerator and denominator are colliding in real time.
Volume and value have decoupled. The repricing phase is the part that comes next. Investors positioning for AI’s next chapter should watch not how many tokens are burned. They should watch which companies can prove durable enterprise value per dollar spent.
Get the week’s most important financial stories, market analysis, and investment insights delivered straight to your inbox. No fluff. Just what matters. Subscribe free →
