Your signal. Your price.
William Cohan explains that regulatory limits under the Dodd-Frank Act forced traditional banks to move long-term loans off their balance sheets. This policy vacuum allowed private credit providers like Apollo, Blackstone, and KKR to dominate the corporate lending space.
Treasury Secretary Scott Bessent argues that the global economy must grow its way out of its massive debt burden by easing banking regulations. John Arnold notes that post-2008 Dodd-Frank capital requirements stifled private credit creation, forced community bank consolidation, and fueled the rapid rise of the shadow banking sector.
In 2021, financial agencies quietly removed the debt-to-income threshold from Dodd-Frank, fueling a wave of high-leverage lending right as the housing cycle neared its peak.
Frank argues that well-defined tasks should run on cheap, specialized models. Conversely, ambiguous tasks require frontier models to prevent cheaper models from wasting tokens on routing errors.
Frank notes that while OpenAI and Anthropic dominate user engagement, the market still lacks a strong, independent enterprise agent orchestration application.
Kalshi claims it operates under Dodd-Frank legislation as a prediction market, not gambling, arguing for federal preemption over state regulation, but Bloomberg data shows bettors lost $294 million on sports-like "parlay bets" since the year's start.
Frank reports Moonshot's Kimi K3 model, while performing well on benchmarks between Opus/Fable and GPT 5.6, fits the pattern of open-source models lagging the latest frontier. The model is the largest open-source release at 2.8 trillion parameters.
Frank considers NVIDIA investments in Chinese AI firms dubious, referencing Meta's failed Manus acquisition, yet expects Jensen Huang to lobby intensely for NVIDIA chip access given global demand.
The AI frontier's definition has evolved from the smartest model to the smartest model with the lowest inference cost. Frank argues open-source models, despite cheap training, may not compete on intelligence per unit cost.
Frank points to unknown factors for Chinese model companies, including R&D borrowing and government subsidies. DeepSeek's V4 model, priced up to 10x lower on its platform than on Microsoft Azure, suggests significant structural margin differences.
Frank's analysis of OpenRouter data reveals open-source models consume 75% of token volume but only 20% of dollar spend, while closed-source models account for 80% of spending. This dynamic still shows increasing spend on open models year-over-year.
High demand for models like Kimi K3 emphasizes the intensifying need for energy infrastructure to support AI's growing compute requirements. Frank notes cloud companies report demand outpacing supply, driving rapid data center construction and capital expenditure increases.
Frank argues that while demonstrating ROI for AI spending is crucial, overall spending will grow, with optimization further incentivizing investment by improving returns.
Frank uses an analogy, inspired by Brett, to explain that benchmarks do not tell the whole story; users will pay for higher intelligence for ambiguous tasks despite the existence of "good enough" models.
Frank distinguishes that well-defined tasks suit the cheapest adequate model, but ambiguous tasks require smarter models to fill gaps efficiently, potentially becoming more cost-effective than cheaper alternatives.
Frank suggests that the broader scope of poorly defined tasks, compared to well-defined ones, drives the demand for smarter models, as illustrated by UiPath's business scale.
Frank stresses the critical need for intuitive, rapidly improving, and integrated agent products, noting a current void for enterprise-ready, independent applications beyond ChatGPT or Claude.
Frank notes that Kalshi has launched prediction markets for compute, including H100 GPU prices. He predicts H100 prices will decrease next month due to the ramping capacity of Blackwell and Vera Rubin architectures.
Frank notes Gemini 3.5 Pro's delayed release as Google likely responds to competitive launches from OpenAI, Anthropic, Meta, and XAI.
Frank's benchmark chart shows OpenAI and Anthropic lead frontier intelligence, with XAI and Meta delivering near-frontier performance at lower cost.
Frank estimates Anthropic's run rate revenue is $60-70 billion, heading toward $100 billion by year-end.
Frank contrasts that with OpenAI revenue in the mid-tens of billions, while open-source AI companies operate in the hundreds of millions to low billions.
Frank describes enterprise pricing tiers: all-you-can-eat pro plans for small firms create lock-in, while per-token enterprise plans drive cost-conscious routing.
Frank does not believe Grok will catch OpenAI or Anthropic to become the number-one model in the next year and a half.
Frank predicts Grok will gain users and see token growth exceeding user growth due to its competitive performance-per-cost.