Divya van Mahajan

OpenAI's New Deal

ai-daily-brief-podcast

Study Document: OpenAI’s Industrial Policy Proposal & AI Industry News

Source: AI Daily Brief – Episode: “2026-04-08-openais-new-deal” Host: Nathaniel Whittemore (AI Daily Brief) URL: Not available

Overview

This episode of the AI Daily Brief covers two major threads. First, a set of headline news items about the competitive and financial landscape of the leading AI labs—particularly Anthropic’s explosive revenue growth and a new compute partnership with Google and Broadcom. Second, and more extensively, a critical analysis of OpenAI’s newly released policy document, Industrial Policy for the Intelligence Age, which proposes a framework for managing the societal transition to AI. The host argues that while the document raises important policy discussions, it fails both as a public relations instrument and as a substantive policy commitment, arriving at a moment of worsening public sentiment toward AI.

Prerequisites

  • Basic familiarity with the leading AI labs: OpenAI, Anthropic, Google DeepMind, Meta AI
  • Understanding of startup financial metrics: ARR (Annual Recurring Rate), cash flow, burn rate, IPO processes
  • Awareness of the general public discourse around AI and labor displacement
  • Familiarity with U.S. policy concepts: wealth funds, tax policy, social safety nets, labor rights
  • General knowledge of AI hardware infrastructure: TPUs, data centers, GPU/chip supply chains
  • Understanding of Goodhart’s Law (metrics becoming targets cease to be good measures)

Main Points

1. Anthropic’s Revenue Surge to $30 Billion ARR

  • Anthropic announced a $30 billion annualized run rate (ARR), a 3x increase since end of 2024 and a 58% increase since February 2026.
  • This growth rate—approximately 9,700% annualized—is described as the fastest revenue growth at this scale in history; NVIDIA’s best comparable quarter was 1,240% annualized growth.
  • Anthropic’s revenue is almost entirely from enterprise customers; 1,000 enterprise customers now have annual spend above $1 million, doubling from 500 in under two months.
  • By contrast, OpenAI’s revenue skews more toward consumers, meaning it carries inference costs for a large free user base that Anthropic does not.

2. Financial Scrutiny Ahead of IPOs

  • The Wall Street Journal published a deep-dive into both companies’ financials sourced from fundraising disclosures.
  • OpenAI expects to spend ~$30 billion on model training in 2026 (triple the prior year); Anthropic projects training costs to reach ~$28 billion by 2028.
  • Both companies present an alternate profitability metric that excludes training costs; critics compare this to an airline claiming profitability by excluding the cost of jets.
  • OpenAI projects cash flow positivity by 2030; Anthropic forecasts GAAP-style profit by 2028.
  • The default IPO narrative both companies must fight: they will burn massive cash and rely on IPO investors to sustain the business.

3. Anthropic’s New Compute Partnership with Google and Broadcom

  • Anthropic signed an expanded partnership adding 3.5 gigawatts of compute capacity, coming online from 2027, with the majority of data centers in the U.S.
  • The deal deepens Anthropic’s commitment to Google’s TPUs (manufactured by Broadcom), used exclusively for inference; AWS remains the exclusive training cluster partner.
  • For Google, the deal effectively creates a multi-billion dollar external TPU business built around a single customer, addressing prior concerns about Google’s ability to compete with NVIDIA in chip sales.
  • Broadcom gains guaranteed demand tied to Anthropic’s growth trajectory.

4. Google’s Gemma 4 and On-Device AI

  • Google released an AI dictation app, Google AI Edge Eloquent, built on the Gemma 4 small model, running entirely locally on-device with no internet connection required.
  • Gemma 4 was downloaded 2 million times in its first week (vs. 6.7 million for Gemma 3 over a full year; Alibaba’s Qwen 3.5 achieved 27 million since mid-February).
  • The entire Gemma 4 family—down to the 2B parameter model—demonstrates strong agentic performance on mobile; a developer demonstrated Wikipedia querying via agent skills on an iPhone.
  • The host notes this could position Gemma 4 as a candidate for powering Apple’s relaunched Siri (expected summer 2026), though it may not yet be sufficient for a full offline Siri experience.

5. Meta’s Upcoming Model (“Avocado”) and Open-Source Commitment

  • Axios reported that Meta (under AI CEO Alexander Wang) plans to release a proprietary version first for safety reasons, followed by an open-source version.
  • Wang reportedly views Meta as a democratizing force, positioning it against OpenAI and Anthropic, which he sees as increasingly focused on government and enterprise.
  • The model (codenamed Avocado) was previously reported delayed in March; talk of safety concerns may indicate improved performance after additional post-training.
  • Meta acknowledges its model won’t be competitive across all benchmarks but expects specific consumer-appeal strengths.

6. Meta’s Internal “Token Maxing” Culture

  • Meta employees have created an internal leaderboard called CLAWDonomous (using Anthropic’s Claude), tracking top 250 token consumers among 85,000 employees, with ranks like “Session Immortal” and “Token Legend.”
  • The concept of token maxing—measuring AI productivity by token consumption—is being driven from the top; NVIDIA CEO Jensen Huang suggested engineers earning $500K should consume $250K worth of tokens annually.
  • Critics compare this to Mao-era steel-smelting quotas (Goodhart’s Law problem: metric becomes the target, not the outcome).
  • A counterargument: the engineering feat required to actually consume that many tokens is so difficult that incentivizing the attempt is itself valuable; most corporations cannot achieve it regardless.

7. OpenAI’s Policy Document: Industrial Policy for the Intelligence Age

  • Released against a backdrop of worsening U.S. public sentiment: 55% of Americans believe AI will do more harm than good (up 11 points YoY); 70% believe AI will reduce job opportunities (up 14 points); AI now has worse public perception than ICE.
  • The document is divided into two sections: Building an Open Economy and Building a Resilient Society.
  • The host’s central critique: the document sits in an “uncanny valley”—too technocratic to function as effective public communication, yet not substantive enough to advance real policy.

Policy Proposals and Host Commentary

ProposalHost’s Assessment
Worker voice in AI transitionImportant concept, but ignores political history; avoids the word “union”; real change requires a labor movement, not a policy memo
AI-first entrepreneurship supportValid as one tool in a broader toolkit; not a solution for displaced workers individually, but increasing entrepreneurship rates is worthwhile
Right to AI (access)Access alone is insufficient without investment in capability; companies spend 12x more on AI infrastructure than on training people to use it
Modernize the tax baseInevitable if wealth shifts from labor to capital; may break traditional partisan lines; could include capital gains increases or automation taxes
Public wealth fundNot inherently bad, but skepticism that small distributed shares move the needle on core sentiment; “window dressing” risk
Accelerate grid expansionHost is enthusiastic, but argues benefits should actively reduce people’s energy costs, not merely avoid increasing them
Efficiency dividends (4-day workweek, etc.)Skeptical of the 4-day workweek as a panacea; more supportive of portable benefits (healthcare, retirement, training) funded by AI-generated wealth
Adaptive safety netsMost technically interesting proposal: use real-time data on wages, job quality, and AI impact to trigger dynamic, targeted social safety net interventions

8. The Core Failure: No Commitments from OpenAI Itself

  • Critics note the document proposes actions for policymakers and others but contains no commitments that would cost OpenAI anything.
  • Specific gaps identified: OpenAI could seed the public wealth fund; accept voluntary energy rate separation; reinstate profit caps it dismantled; pledge equity or funding toward these initiatives.
  • The only concrete OpenAI commitments mentioned: a workshop, fellowships paid in OpenAI product credits, and an email address.
  • The host’s conclusion: the document is counterproductive in the current environment, where 13-page PDFs with no binding commitments do not address the gravity of what the industry itself claims is coming.

9. The Broader Communication Failure of the AI Industry

  • The host argues the AI industry systematically inverts the communication ratio: spending most of its messaging validating risks and concerns, with only brief acknowledgment of benefits.
  • In the absence of a compelling answer to “why are we doing this?”, the public defaults to: “because it will make some people rich.”
  • The “China will do it if we don’t” argument remains too abstract for most audiences.
  • The only viable answer is demonstrating that benefits exceed costs—and simply asserting this “in a hand-wavy way” is no longer sufficient.

Key Concepts

  • ARR (Annualized Run Rate): A projection of annual revenue based on the most recent revenue period, extrapolated forward; both companies note it is calculated differently, complicating direct comparisons.
  • Training costs vs. inference costs: Training costs are one-time (per model generation) capital expenditures to build a model; inference costs are ongoing operational costs each time the model responds to a user.
  • TPUs (Tensor Processing Units): Google’s custom AI accelerator chips, manufactured by Broadcom, used by Anthropic for inference workloads.
  • Token maxing: The practice of maximizing the number of AI tokens consumed as a proxy metric for AI-enhanced productivity.
  • Goodhart’s Law: When a measure becomes a target, it ceases to be a good measure; applied here to token consumption as a productivity metric.
  • Industrial Policy for the Intelligence Age: OpenAI’s policy document proposing frameworks for managing the AI economic transition, covering worker rights, entrepreneurship, taxation, public wealth funds, grid expansion, and social safety nets.
  • Efficiency dividends: OpenAI’s term for redistributing economic gains from AI productivity back to workers and citizens (e.g., reduced work hours, portable benefits).
  • Adaptive safety nets: Dynamically adjusted social support programs informed by real-time data on AI’s economic impact, rather than static program structures.
  • Right to AI: OpenAI’s proposed principle that access to AI tools should be treated as foundational to economic participation, analogous to electricity or internet access.
  • Avocado: Meta’s codename for its next major AI model, expected to release in a proprietary version first followed by open-source.
  • CLAWDonomous: Meta’s internal employee leaderboard tracking token consumption using Anthropic’s Claude.
  • Gemma 4: Google’s latest small open-source model family, notable for on-device deployment capability and strong agentic performance at small parameter sizes.

Summary

The episode covers a pivotal moment in the AI industry characterized by extraordinary growth—Anthropic’s 9,700% annualized revenue growth and a massive new compute infrastructure deal with Google and Broadcom—alongside deepening financial scrutiny of both Anthropic and OpenAI as they approach IPOs. Against this backdrop, the host delivers an extended and sharply critical analysis of OpenAI’s policy document, Industrial Policy for the Intelligence Age, arguing that it fails on two simultaneous fronts: it is too dry and technocratic to shift public opinion at a moment when 55% of Americans believe AI will harm their lives, yet too vague and uncommitted to function as serious policy advocacy. The host engages substantively with each policy proposal—worker voice, entrepreneurship support, the right to AI, tax modernization, a public wealth fund, grid expansion, efficiency dividends, and adaptive safety nets—finding genuine value in some ideas while criticizing the document’s central flaw: OpenAI proposes costs and commitments for governments and others while volunteering nothing concrete itself. The host’s broader argument is that the entire AI industry has a fundamental communication failure: it spends disproportionate time validating risks and concerns while failing to articulate a compelling, specific case for why AI should exist and who will concretely benefit—leaving the public to conclude the only real beneficiaries are the industry’s shareholders.