Divya van Mahajan

Who Thinks There's an AI Bubble?

ai-daily-brief-podcast

Who Thinks There’s an AI Bubble?

Overview

This episode of the AI Daily Brief (dated August 12, 2025) covers two major topic areas: a headline segment on NVIDIA’s unprecedented revenue-sharing deal with the U.S. Government for Chinese chip sales, and a main segment examining the changing conventional wisdom around AI — specifically whether AI markets constitute a bubble. The host (unnamed in the transcript) argues that lazy historical analogies to past tech bubbles do not adequately account for the genuine, demonstrated demand driving AI adoption. The episode also touches on GitHub’s restructuring, Anthropic’s revenue composition, and the emergence of AI-focused hedge funds.

Source video: URL not provided.

Prerequisites

  • Basic familiarity with AI terminology (LLMs, inference, tokens, ARR)
  • General understanding of U.S.-China technology export controls and their history
  • Awareness of major AI companies: OpenAI, Anthropic, NVIDIA, AMD, Google, Microsoft/GitHub
  • Familiarity with venture capital and hedge fund mechanics
  • Some context on prior tech boom-bust cycles (dot-com, Web 2.0, crypto) is helpful for evaluating the bubble argument

Main Points

1. NVIDIA and AMD’s Revenue-Sharing Deal for China Access

  • NVIDIA and AMD have reportedly agreed to give the U.S. Government 15% of revenues from Chinese chip sales in exchange for export licenses, particularly for NVIDIA’s H20 chip.
  • The deal originated from a negotiation between Jensen Huang and President Trump; Trump initially requested 20%, Huang countered with 15%, and AMD was brought in under the same terms.
  • Export control experts describe the arrangement as unprecedented — no U.S. company has previously paid a share of revenues to obtain export licenses.
  • Critics across the political spectrum object on different grounds:
    • National security hawks argue it amounts to paying a bribe to enable a security risk.
    • Business commentators call it “corporatism” or “mafia-style sovereign risk.”
    • Some Wall Street analysts, however, view it positively as converting binary ban-risk into a predictable, modelable cost.
  • China has responded by officially discouraging its firms from using H20 chips, particularly for government and national security applications.
  • Back-of-envelope math: NVIDIA could sell ~$23 billion in chips to China annually, making the U.S. Government’s 15% stake worth ~$3.5 billion per year.

2. NVIDIA’s Work Beyond Chips: Cosmos World Models

  • NVIDIA unveiled Cosmos Reason, a 7-billion-parameter reasoning vision-language model, expanding their Cosmos family of world models.
  • The model is designed for embodied AI use cases: data curation, robot planning, and video analytics.
  • The host notes that embodied AI and robotics receive less media attention than LLMs but represent a comparably large transformative opportunity.

3. GitHub Restructuring and Microsoft’s AI Strategy

  • GitHub CEO Thomas Domke is stepping down after nearly four years; he announced plans to become a founder again.
  • His role will not be replaced; GitHub will be folded into Microsoft’s core AI engineering team, led by former Meta executive Jay Parikh (separate from Mustafa Suleiman’s consumer AI team).
  • The move signals Microsoft’s recognition that GitHub Copilot is a core competitive asset in enterprise AI, not merely a developer relations tool.
  • GitHub Copilot launched in late 2021, over a year before ChatGPT; Microsoft claims 20 million all-time users, though active user figures are undisclosed.

4. Anthropic’s Revenue and Coding Dominance

  • Anthropic reached $5 billion ARR, driven heavily by Claude 4 and its adoption in AI coding.
    • $3.1 billion comes from API revenue — slightly exceeding OpenAI’s API revenue at the time.
    • Half of API revenue ($1.4–1.5 billion) comes from just two customers: Cursor and GitHub Copilot.
    • Claude Code alone is generating ~$400 million ARR, reportedly doubling in a matter of weeks.
  • For comparison, OpenAI doubled its total ARR from $6B to $12B in roughly six months; Anthropic grew 5x from $1B to $5B in seven months.
  • Anthropic’s growth is assessed as highly dependent on coding dominance: almost every major coding assistant defaults to Claude 4 Sonnet.
  • GPT-5’s competitive pricing (approximately 10x lower than anticipated) is viewed as a direct competitive move targeting Anthropic’s coding revenue.

5. The Rise of AI-Focused Hedge Funds

  • Leopold Aschenbrenner, a 23-year-old former OpenAI researcher, raised $1.5 billion for a hedge fund called Situational Awareness (named after his 165-page 2024 essay of the same name).
  • The fund was up 47% after fees in the first half of 2025, versus the S&P 500’s 6% and tech hedge fund indexes up ~7%.
  • Strategy: long positions in semiconductor, infrastructure, and power companies benefiting from AI build-out; startup investments (including Anthropic); short bets on industries expected to be disrupted by AI.
  • Investors include Patrick and John Collison (Stripe founders), Daniel Gross, and Nat Friedman.
  • The WSJ draws a comparison to ESG-themed hedge funds, many of which have since closed; the host argues AI is fundamentally different due to genuine, massive demand rather than policy-driven or artificial demand.

6. Wall Street Shorting AI-Disrupted Companies

  • Alongside long AI positions, investors are actively shorting companies perceived as AI disruption targets.
  • Bank of America published a list of 26 companies most at risk — notably skewed toward recent-decade tech successes rather than legacy industries:
    • Examples: Wix (web development), Shutterstock (stock imagery), Adobe (creative software), Gartner (research and advisory).
  • Gartner is highlighted as a case study: after cutting revenue forecasts, the stock fell 30% in five days (its worst weekly drop on record) and has been cut in half year-to-date. Morgan Stanley cited the results as “adding fuel to the AI disruption case.”
  • The pace of potential disruption is noted as historically unusual; prior waves (PC, internet) moved more slowly.

7. The AI Bubble Debate

  • The episode’s central argument addresses whether AI is a bubble, specifically responding to a post by Antonio Garcia Martinez drawing structural analogies between current AI and previous tech bubbles.
  • Martinez’s argument (as summarized by the host):
    • Every tech bubble is inflated by external liquidity subsidizing unsustainable growth.
    • AI is using VC and inflated equity to subsidize compute costs, artificially inflating consumer usage.
  • The host’s rebuttal has two components:
    1. Conceded point: AI may be currently underpriced relative to its eventual equilibrium price due to capital subsidization — analogous to how VC subsidized cheap Uber rides in the 2010s.
    2. Disputed point: The subsidy argument does not explain away genuine demand. ChatGPT growing from 0 to 700 million users in ~2.5 years cannot be attributed solely to low pricing. Coding tool customers are consuming as much of the product as they can get — indicative of real value, not subsidy-driven behavior.
  • Additional data points against bubble framing:
    • Google’s token processing jumped from 480 trillion in May to 980 trillion in July 2025 (104% growth in two months).
    • Inference costs have dropped dramatically, but demand is growing faster than costs fall.
    • Platforms like Replit adjusted pricing models (shifting from per-request to compute-time billing) and margins recovered — indicating a business model adaptation issue, not a fundamental demand problem.
  • The host concludes: markets may be overpricing certain stocks, and inflated expectations will at times go unmet, but broad “AI bubble” analogies borrow too lazily from historical precedents that don’t map cleanly onto the current situation.

Key Concepts

  • H20 chip: NVIDIA’s less-advanced GPU variant, previously subject to U.S. export controls on sales to China; now potentially licensable under the new revenue-sharing arrangement.
  • Export license revenue-sharing: The unprecedented arrangement whereby NVIDIA and AMD pay the U.S. Government 15% of Chinese chip sales revenue in exchange for permission to sell.
  • Cosmos Reason: NVIDIA’s 7B-parameter reasoning vision-language model designed for embodied AI tasks such as robot planning and video analytics.
  • Embodied AI: AI systems integrated into physical devices or robots that interact with the real world.
  • Situational Awareness (essay and fund): Leopold Aschenbrenner’s 2024 165-page document arguing AGI is imminent and foreseeable; also the name of his AI-focused hedge fund.
  • ARR (Annual Recurring Revenue): A metric used to measure subscription/recurring revenue on an annualized basis, used here to compare OpenAI and Anthropic’s growth trajectories.
  • AI disruption shorts: Investment strategy of selling short the stocks of companies whose business models are seen as threatened by AI adoption.
  • Gross margin / negative gross margin: Revenue minus the direct cost of goods sold; some AI coding platforms are reportedly spending more to serve users than they earn, resulting in negative gross margins.
  • Inference cost: The computational cost of running a trained AI model to generate outputs; has fallen dramatically but is outpaced by rising demand.
  • Vibe coding: Colloquial term for AI-assisted coding, particularly where non-developers use AI tools to write code.
  • ESG funds: Environmental, Social, and Governance investment funds; cited as a cautionary comparison case where thematic demand outpaced genuine investment opportunity.
  • World models: AI systems that build internal representations of how physical or simulated environments work, used in robotics and embodied AI.

Summary

The episode argues that the conventional wisdom around AI — both its near-term transformative pace and whether it constitutes a financial bubble — is shifting but that much of the skepticism is poorly grounded. On the market structure side, the host uses the emergence of dedicated AI hedge funds (particularly Leopold Aschenbrenner’s Situational Awareness fund, up 47% in H1 2025) and the active shorting of AI-disrupted companies as evidence that Wall Street is beginning to take the AI transition seriously on multiple fronts. Against the backdrop of Anthropic’s heavy coding revenue concentration and GitHub’s absorption into Microsoft’s core AI team, the competitive dynamics in enterprise AI coding are intensifying, with GPT-5’s aggressive pricing posing a direct challenge to Claude’s dominance. On the bubble question specifically, the host concedes that capital subsidization may be temporarily suppressing AI pricing below its equilibrium level, but firmly rejects the broader claim that AI usage is artificially inflated: the scale and growth rate of genuine user demand — 700 million ChatGPT users, 104% token growth at Google in two months, and developer workflows that actively maximize consumption — distinguish AI from prior bubble episodes like Web 2.0 or crypto, where monetizable demand was either thin or fabricated. The episode cautions against applying historical analogies too readily to a technological transition that is, by most measurable indicators, unlike anything markets have previously encountered.