Episode Summary
Executive Summary: Cal Newport and Ed Zitron review AI in 2025 month by month and conclude it was a terrible year for the industry: hype shifted from agents and superintelligence to inference-heavy cost blowups, underwhelming model gains, and a growing bubble narrative. They argue the strongest evidence is financial and operational, not existential: AI products remain expensive, lossy, and difficult to monetize despite massive investment.
Main Topics: DeepSeek and the January AI reset (Priority: 5/5): DeepSeek shocked markets by showing a cheaper path to model training and exposing how much the AI boom depended on Nvidia and the assumption that only U.S. firms could build frontier models. The hosts argue the story was later memory-holed because it threatened the dominant investment narrative. The rise and collapse of the agents narrative (Priority: 5/5): In early 2025, OpenAI and others promoted AI agents as digital labor and the next major use case. Zittron argues this was mainly marketing: coding assistants and vibe coding had narrow utility, while autonomous agents failed to deliver reliable economic value. Scaling limits, GPT-4.5, and the shift to reasoning/test-time compute (Priority: 5/5): February and March showed that brute-force scaling was stalling. GPT-4.5 underwhelmed relative to its hype, and Nvidia/Jensen Huang reframed the era around post-training and inference, which supports Nvidia sales but increases costs for AI providers. AI doom returns via AI 2027 and safety rhetoric (Priority: 4/5): The AI 2027 scenario and similar safety narratives revived fears of superintelligence and extinction, but the hosts dismiss them as speculative, vague, and often self-serving. They argue such claims avoid concrete present-day harms like labor exploitation, pollution, and energy use. GPT-5 disappointment and the bubble becomes visible (Priority: 5/5): GPT-5’s August launch became a major turning point: it was overhyped, then quickly re-framed as less than revolutionary. That underperformance helped trigger more skeptical reporting on the AI bubble, model economics, and unsustainable spending. Financial reality: inference costs, capex, and circular deals (Priority: 5/5): Late 2025 reporting showed enormous inference bills, weak revenue, and circular mega-deals among OpenAI, Nvidia, Oracle, AMD, and Broadcom. The hosts emphasize that reported growth often depended on accounting optics and future promises rather than actual business durability. Sora, Disney, and the desperation for consumer products (Priority: 4/5): OpenAI’s Sora app and Disney’s related deal are framed as attempts to create new revenue and prestige by attaching AI to entertainment and social media. The hosts see this as evidence of desperation and a search for a viable consumer wedge, not proof of broad adoption.
Key Arguments: DeepSeek mattered because it showed frontier-ish models could be trained far more cheaply, undermining the claim that only enormous capital-heavy labs could compete. Agents were heavily marketed as digital labor, but in practice coding agents mostly helped with limited autocomplete and prototype generation, not robust autonomous work. GPT-4.5 and GPT-5 revealed diminishing returns from scaling; any gains increasingly required expensive reasoning and test-time compute rather than straightforward model size increases. The AI industry’s real economics are worse than its public narrative: costs rise with usage, inference is expensive, and many companies appear to lose money on core products. Nvidia benefits from a shift toward inference-heavy workloads because it increases demand for GPUs beyond initial training. AI doom discourse is often vague, speculative, and convenient for status-seeking or fundraising; it does not address current harms like labor exploitation, energy consumption, and data-center pollution. OpenAI, Anthropic, Oracle, and others have used big partnerships and future revenue promises to create the appearance of momentum, even when the underlying math is dubious. The hosts think the most important AI story of 2025 is not intelligence breakthroughs but the exposure of the industry as a capital-intensive hype machine.
Data Points: DeepSeek training cost: $5.3 million - Zittron says DeepSeek R1 was trained for far less than American frontier models, which often cost $50M-$100M+. AI model training cost comparison: $50 million to $100 million+ - Referenced as typical American frontier-model training cost before DeepSeek. GPT-4.5 release date: February 27, 2025 - Altman posted that GPT-4.5 was ready but expensive and GPU-constrained. GPT-4.5 GPU shortage: Tens of thousands of GPUs next week; hundreds of thousands coming soon - Altman said OpenAI was out of GPUs and needed massive additions to roll out GPT-4.5. OpenAI revenue by end of September: $4.3 billion - Used in comparison against much larger inference costs. OpenAI inference cost through end of September: $8.67 billion - Presented as the core financial problem: costs exceeded revenue. Anthropic AWS spend: $2.66 billion - Reported for three quarters; Zittron argued total cloud spend may be around $5B when Google Cloud is included. OpenAI annual revenue expectation: $13 billion - Referenced as projected annual revenue, contrasted with major data-center obligations. OpenAI-Nvidia deal size: $100 billion - Discussed as a headline investment announcement that lacked clear signed details and future build-out timing. OpenAI-Oracle deal size: $300 billion - A future data-center commitment that fueled stock reactions despite OpenAI lacking the cash. OpenAI-AMD deal size: 6 gigawatts - Described as a future buildout in exchange for AMD stock. OpenAI-Broadcom deal size: 10 gigawatts - Another large future infrastructure commitment that expanded the appearance of demand. OpenAI core spending target: Over $150 billion - Referenced as an announced or leaked spending expectation that did not appear to match revenue. GPT-5 launch timing: August 2025 - The launch became a turning point because it failed to meet the transformational expectations Sam Altman had raised. AI data-center capex share of GDP growth: More than consumer spending combined - Mentioned as a striking macroeconomic sign of how much investment the AI buildout was driving. Meta annual revenue: Over $200 billion - Used to show Meta’s scale despite questionable product strategy. TikTok annual revenue: About $30 billion - Used for comparison with Meta’s much larger revenue base. AWS build cost over nine years: About $70 billion - Used to contrast cloud infrastructure economics with AI infrastructure economics. OpenAI training/infrastructure context: 26 gigawatts - Mentioned as the scale of OpenAI’s planned data-center buildout by late 2025.
Pivotal Quotes: "2025 is the year of AI agents" — OpenAI CPO / Axios framing: Used to illustrate how media paraphrasing and startup messaging turned a vague product narrative into a headline-friendly hype cycle. "We really wanted to launch it to plus and pro at the same time, but we've been growing a lot and are out of GPUs." — Sam Altman: Altman’s February 2025 post about GPT-4.5 underscored the GPU shortage and the costliness of scaling. "The whole thing hinges on this idea that they invented an AI that could research how to build an AI they wanted." — Ed Zitron: His critique of AI 2027, arguing the scenario assumes an unproven leap that the piece never explains.
Implications: The episode argues 2025 exposed AI as economically fragile: huge capex, weak unit economics, and overblown claims. Listeners should expect more consolidation, more hype recycling, and more scrutiny of whether AI products can ever become sustainably profitable.