SecondBrain
Ask the Brain
Index/Sourceupdated Sat Jun 13 2026 08:00:00 GMT+0800 (Philippine Standard Time)

The New Dumbest Chart in AI (AI Daily Brief)

youtubeai-daily-brieftoken-economicsai-capextoken-scarcityipodata-centers

The New Dumbest Chart in AI

Nathaniel Whittemore on The AI Daily Brief dismantles the Wall Street panic over Citadel Securities' "tokconomics" note and the Silicon Data LLM token expenditure index. His core argument: the scary downward line everyone is screenshotting measures a usage-weighted average token price, not total token demand, total token volume, or total token expenditure — and even that signal is exaggerated because the underlying data comes only from third-party token routers whose entire job is to find cheaper tokens. The real story, he says, is the well-underway shift from a "token subsidy" era to a token scarcity era, where markets ration the most expensive frontier tokens to the firms that can use them best — a market rationalizing, not an AI Bubble popping. He pairs this with a headlines run: a record SpaceX IPO, Bezos-backed Prometheus, the KKR–Nvidia Helix Venture, and Goldman Sachs' ~$1trn+ AI capex forecast.

Summary

The episode opens on the speed of Wall Street's swing "from token maxing to token panic" and argues the panic rests on a misread chart. The Silicon Data token expenditure index — amplified by Citadel's note and viral X posts (Zero Hedge's "tokconomics = panic," Real Vision's "the whole setup depends on this") — actually tracks only the average price paid per million tokens across third-party routers, which fell in mid-June back to early-May levels. Whittemore reframes this as exactly the Token Scarcity dynamic the show has been tracking: as use moves from assisted to agentic, demand outpaces a physically constrained token supply, so every AI company is now "in the token efficiency business." He argues total token-volume growth will dwarf any savings from cheaper baskets — citing Ramp data that the median firm spends just $11.38/employee/month on AI vs. $75k for the top 1% — so the buildout thesis holds. The headlines cover the largest IPO in history (SpaceX), Prometheus' $41bn round, a $10bn KKR–Nvidia data-center JV, and Goldman calling consensus AI capex "rookie numbers."

Key claims

  • The chart critique. The Silicon Data "LLM token expenditure index" should have been named the token expenditure price index: it's a usage/expenditure-weighted average price per million tokens, not demand, not volume, not total spend. Silicon Data itself clarified this on X ("marginal willingness to pay"). The mid-June dip just means the average price paid fell from an early-June peak back to early-May levels.
  • The data is biased toward cheapness. The index draws only from third-party token routers — whose whole market purpose is routing use cases to lower-cost models — so it overstates any real shift away from frontier models. Whittemore therefore disagrees even with Silicon Data's "willingness to pay" framing; he calls it a useful leading indicator of where advanced users are heading, not the average buyer's experience.
  • Token maxing → token panic → token scarcity. Follows recent caps news (Walmart capping internal AI use, Uber's ~$1,500/employee/month cap after blowing its budget). The genuine signal is the bifurcation of "frontier vs. everyday AI usage" — Citadel argues the most expensive inference will concentrate among firms with the balance sheets, research depth, and operating domain to exploit it. Whittemore: that's "an AI market rationalizing," not a bubble popping.
  • Volume dwarfs price-mix shifts. Ramp's data: top 1% of firms spend ~$75k/employee/yr on AI; top 10% ~$610/month; the median firm just $11.38/month. If the median rose toward Uber's $1,500 cap, total market growth would dwarf any revenue lost to efficiency. Reported OpenAI price cuts (preempting Anthropic) wouldn't tank the industry — Max Weinbach estimates ~70% margins on inference-intensive API tokens, room to cut ~60% and stay profitable.
  • SpaceX record IPO. Largest IPO in history: $135/share flat pricing, ~$1.8trn valuation (7th-largest company, ahead of Aramco/Tesla/Meta). Retail orders exceeded $100bn against a $75bn raise (retail allocation cut 30%→20%, ~7x oversubscribed). 2025 financials: $5bn loss on $18.7bn revenue. Goldman (also running the IPO) forecast $474bn revenue by 2030. Whittemore's read: priced as a neocloud / infrastructure play plus an "Elon halo," not as an AI-model company — so not a clean referendum on the coming OpenAI/Anthropic IPOs.
  • Bezos-backed Prometheus. $12bn round at a $41bn valuation (JP Morgan, Goldman, BlackRock, Bezos). Building an "artificial general engineer" to design/manufacture anything; eyeing a ~$100bn industrial-buyout fund (PE rollup model on manufacturing) — because the physical economy "can't be scraped," you acquire the factories that generate the data.
  • KKR–Nvidia "Helix" JV. A $10bn data-center construction company (Helix Digital Infrastructure), KKR + Kuwait sovereign wealth as capital, Nvidia supplying chips, Vistra supplying power, ex-AWS CEO Adam Selipsky leading. One of several (cf. Broadcom–Apollo–Blackstone). Counterpoint: JLL reports ~half of US data-center projects are delayed — see Data Center Backlash.
  • Goldman: consensus capex is "rookie numbers." Median Wall Street estimate ~$920bn AI data-center spend next year; Goldman expects $1.1trn (base) / $1.4trn (bull) for 2027, assuming token consumption rises 24x through 2030 as agents deploy, with higher input costs pushing nominal capex even higher. Whittemore is "firmly in the Goldman camp."

Cross-vault relevance

tip This directly extends the user's Token Maxing thesis toward "token scarcity." Whittemore's "every AI company is now in the token efficiency business" and the "token subsidy → token scarcity" framing are the same arc as the user's Managing Enterprise IT Development in the Era of Token Scarcity query page — strongly corroborating it. The enterprise takeaway is concrete: the era of cheap, unmetered tokens is ending; agentic adoption (the Agentic Loop) is what drives demand past constrained supply, forcing IT leaders into mixed-basket model routing, token budgets/caps, and ROI scrutiny — exactly the discipline P&G-scale IT will need. The Ramp data ($11.38 median vs. $75k top 1%) is a useful planning yardstick for where most enterprises actually sit on the adoption curve.

Wikilinks

The AI Daily Brief · Nathaniel Whittemore · Token Maxing · Token Scarcity · Managing Enterprise IT Development in the Era of Token Scarcity · Token Expenditure Price Index · Agentic Loop · AI Capex Supercycle · AI Bubble · SpaceX · OpenAI · Anthropic · Nvidia · Goldman Sachs · KKR–Nvidia Helix Venture · Data Center Backlash · Giga-IPOs and the Stockmarket (Economist)