The AI Chart Everyone Is Getting Wrong
The AI Daily Brief debunks viral misinterpretations of Citadel's 'Token Expenditure Index,' arguing it measures average price paid per million tokens—not total demand or volume—and is drawn exclusively from third-party token routers. The host also covers the SpaceX IPO, Jeff Bezos's Prometheus AI startup, Meta's forced separation from Manus, and Goldman Sachs's bullish $1.1 trillion AI capex forecast for 2027.
Summary
The episode opens with headlines covering several major AI and tech stories before diving into the host's main argument: that a viral chart from Citadel Securities is being widely misread on social media and Wall Street.
**SpaceX IPO:** SpaceX conducted what was described as the largest IPO in history, priced at $135 per share implying a ~$1.8 trillion valuation, making it the seventh largest company in the world. Retail investors submitted over $100 billion in orders, making the retail allocation nearly 7x oversubscribed. The host notes criticism around Goldman Sachs simultaneously conducting the IPO and publishing bullish research, and argues that SpaceX should not be read as a referendum on AI model companies—it's more of an infrastructure/neocloud play with a heavy 'Elon halo' effect.
**Prometheus (Bezos AI Startup):** Jeff Bezos's AI startup Prometheus closed a round valuing it at $41 billion, raising $12 billion from JP Morgan, Goldman Sachs, BlackRock, and Bezos himself. The company aims to build an 'artificial general engineer' that can design and manufacture anything. Bezos dismissed AI jobs apocalypse fears, arguing AI will create a labor shortage by generating 10x more opportunities. The company is also reportedly exploring a $100 billion fund to acquire legacy industrial companies—partly to obtain proprietary manufacturing data that cannot be scraped from the internet.
**Meta/Manus Split:** Meta completed an operational split with Manus after Chinese officials ordered the $2 billion acquisition unwound. Manus staff lost access to Meta systems and vice versa. Manus is now attempting to raise $1 billion for a buyback with uncertain prospects. The episode notes this has had a chilling effect on Chinese startups using 'red-chip' foreign corporate structures, with Beijing reportedly seizing passports of key researchers and executives.
**TSMC Backlogs / Google:** Due to TSMC's years-long waitlist, Google is evaluating Samsung's 2nm process for components of its next-gen TPUs and has placed orders with Intel for advanced packaging. This signals the emergence of a complex supply chain where TSMC handles the most advanced processors but other chipmakers handle less sensitive components.
**KKR/NVIDIA Helix:** KKR and NVIDIA announced a $10 billion data center construction venture called Helix Digital Infrastructure, with Kuwait Sovereign Wealth as a capital partner and Vistra providing energy. Former AWS CEO Adam Salipsky will lead the venture.
**Goldman Sachs Capex Forecast:** Goldman Sachs argues consensus AI capex forecasts are too conservative. While the median Wall Street analyst expects $920 billion in AI data center spending in 2027, Goldman's baseline is $1.1 trillion, with a bullish scenario of $1.4 trillion. Goldman expects token consumption to grow 24x through 2030 driven by agent deployment.
**Main Segment – The Token Expenditure Chart:** The host takes on the viral Citadel Securities 'tokenomics' note and Silicon Data's LLM Token Expenditure Index. He argues that social media commentators and even some Wall Street analysts are fundamentally misreading the chart. The index does not measure total token demand, total token volume, or total token expenditure—it measures the usage-weighted average price paid per million tokens. The downward trend in the chart simply means the average price buyers were paying per million tokens in mid-June had declined from a peak in early June, back to approximately early May levels.
Further, the host points out that Silicon Data's index draws exclusively from third-party token routers—platforms whose entire purpose is to route usage toward cheaper models—meaning the dataset is structurally biased toward showing a shift away from expensive tokens. It captures none of the direct customer-to-OpenAI or customer-to-Anthropic relationships that represent the vast majority of token expenditure.
The host also cites RAMP financial data showing that the median company spends only $11.38 per employee per month on AI, while even the top 10% spend only $610/month. This context makes it nearly impossible, in his view, for any efficiency-driven shift toward cheaper tokens to outweigh the sheer growth in total AI consumption as the rest of the market catches up. He concludes that what the market is experiencing is rational resource allocation, not a bubble bursting—and that Citadel's own note does not actually make the catastrophic claims being attributed to it.
Key Insights
- The host argues that the Silicon Data LLM Token Expenditure Index measures only the usage-weighted average price paid per million tokens—not total demand, volume, or expenditure—making the viral 'token panic' narrative a fundamental misreading of the chart.
- The host points out that Silicon Data's index is drawn exclusively from third-party token routers, platforms specifically designed to route usage to cheaper models, meaning the data structurally overrepresents shifts away from expensive frontier tokens and captures none of direct lab-to-customer relationships.
- RAMP data cited by the host shows the median company spends only $11.38 per employee per month on AI, suggesting that total growth in AI consumption will massively outweigh any revenue loss from customers shifting to cheaper token options.
- The host contends that Citadel's own tokenomics note does not argue frontier AI demand will crater—it argues that expensive AI will concentrate among firms with the balance sheets, research depth, and operating domain to use it most effectively, which he characterizes as market rationalization rather than bubble-bursting.
- The host argues that SpaceX's IPO should not be read as a pricing event for AI model companies, because SpaceX's late pivot to being a 'neocloud' infrastructure provider changes the narrative, and because Elon Musk's personal 'market halo' makes the company a poor proxy for the broader AI sector.
- Bezos's Prometheus reportedly plans a $100 billion fund to acquire legacy industrial companies specifically to obtain proprietary manufacturing data that cannot be scraped from the internet—the host frames this as 'how acceleration escapes the screen and enters atoms.'
- The host notes that Beijing's crackdown on Manus and its 'red-chip' corporate structure—including reportedly seizing passports of key researchers—signals that the previously common practice of Chinese startups decamping to Singapore before seeking foreign capital is now effectively over.
- Goldman Sachs analysts argue that consensus 2027 AI capex forecasts of ~$920 billion are too conservative, projecting a baseline of $1.1 trillion and a bullish scenario of $1.4 trillion, driven by an expected 24x increase in token consumption through 2030 from widespread agent deployment.
Topics
Full transcript available for MurmurCast members
Sign Up to Access