Nathan Lambert hosts this weekly exploration of AI's maturation, examining how the industry is transitioning from startup-phase experimentation to critical infrastructure status. The episode covers the growing demand crunch for AI tokens, shifting business models, and major developments in enterprise AI adoption.

Key developments include GitHub's move to usage-based pricing, record-breaking cloud growth numbers, Anthropic's massive funding round, and government intervention in AI model deployments. The discussion also covers product evolution in AI harnesses and tools like Codex and Cursor, plus the curious case of OpenAI's goblin problem in model training.

The Token Demand Crunch Reshapes AI Economics

GPU rental prices jumped 40% over six months, driven by what Oguzerkin calls "real token demand" rather than speculative bubble activity, with top AI labs generating nearly $60 billion aggregate annual revenue.

Dylan Patel from Semi-Analysis noted that "even the tier two or tier three labs are going to be sold out of tokens," making model ranking debates irrelevant in a supply-constrained market.

Andy Jassy described Tranium demand: "We have such demand right now for Tranium from various companies who will consume as much as we make," highlighting the physical constraints on token production.

OpenAI CFO Sarah Fryer characterized the situation as "a vertical wall of demand, with compute being the bottleneck," forcing the end of flat-price subscription models.

Business Model Evolution from Subsidies to Usage

GitHub shifted Copilot to usage-based billing after Chief Product Officer Mario Rodriguez explained: "A quick chat question and a multi-hour autonomous coding session can cost the user the same amount... the current premium request model is no longer sustainable."

Satya Nadella announced that "any per-user business of ours, whether it's productivity or coding or security, will become a per-user and usage business," signaling Microsoft's broader pricing strategy shift.

Apple's Mac Mini shortage exemplifies hardware constraints, with Tim Cook discussing supply issues on earnings calls as demand outstrips production capacity.

Companies must develop sophisticated token allocation strategies, using premium models only when necessary and cheaper alternatives for routine tasks to maintain cost discipline.

Cloud Giants Post Record AI-Driven Growth

AWS grew 28% year-over-year, its best performance since climbing out of 2021 trough, while Microsoft Azure surged 40% year-over-year.

Google Cloud "absolutely spanked analyst estimates" with 63% year-over-year growth, resulting in Google's second-biggest one-day market cap jump in history.

Google Cloud's backlog growth appears "exponential," with analyst Joseph Carlson noting "this is so crazy it literally looks fake."

Google's cost-to-quality ratio advantage positions them well as companies seek cheaper model alternatives, with Signal noting "we use Gemini heavily because the cost to quality ratio has been absurd for a lot of tasks."

Anthropic's Massive Funding and Market Dynamics

Anthropic began talks to raise over $900 billion, potentially exceeding OpenAI's $825 billion valuation, with investors given just 48 hours to submit allocation requests for a $50 billion round.

Secondary market trading shows Anthropic stock already trading higher than OpenAI, with some shares reportedly reaching trillion-dollar valuations.

Investment logic centers on belief that "there's about a half dozen companies that are writing the story of the future, and there's basically no way that they're not going to be more valuable in the future than they are today."

Microsoft and OpenAI updated their partnership, with Microsoft securing free model access for five more years while OpenAI gains freedom to sell through AWS and Google Cloud.

Government Intervention in AI Model Deployment

White House initially planned to unwind Anthropic's supply chain risk designation and deploy Mythos models to government agencies, including executive order discussions for safe deployment.

Administration officials ultimately opposed Mythos expansion due to national security concerns and fears that Anthropic lacks compute capacity to serve many entities without hampering government access.

Dean Ball characterized this as "the very first case that we know of of the U.S. government restricting rollout of a new AI model based on policy considerations," calling it "an informal, highly improvised licensing regime."

Ball concluded that "the training wheels have come off on AI policy. The trial runs are over," marking a shift toward formal AI governance.

AI Harness Evolution and Product Maturation

Codex received a major upgrade with role-based UI asking users to select from finance, product, marketing, operations, sales, data science, design, student, or other work types for personalized task suggestions.

Codex chose a unified interface approach versus Anthropic's split between Claude Code and Claude Cowork, betting that "knowledge workers will strive to be more technical to unlock their newly discovered wizard powers."

Cursor gained momentum with Lenny Rachitsky noting it's "more fun to work within Cursor than the native Codex or Claude Code apps... and obviously easier to play with new competing models as they come out."

The shift represents movement from "hobbyist PC era to call it the Apple II Plus era of personal computers," with better integrated harnesses replacing hand-built solutions.

The Curious Case of OpenAI's Goblin Problem

GPT models developed an obsession with mentioning "goblins, gremlins, raccoons, trolls, ogres, pigeons" in responses, leading to explicit prompt instructions: "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless absolutely relevant."

OpenAI traced the issue to "nerdy personality" training that encouraged "cute references to various creatures," which then contaminated other personality types through reinforcement learning spillover.

The goblin problem illustrates how "when models are built on top of other models rather than starting from scratch, weird from reinforcement learning in one can have multiplying effects in others."

This discovery prompted OpenAI's research team to develop new tools for auditing model behavior and fixing problems beyond obvious alignment and safety issues.