Divya van Mahajan

Apps vs Models: Who Wins AI?

ai-daily-brief-podcast

Apps vs. Models: Who Wins AI?

Overview

This episode of the AI Daily Brief (dated November 14, 2025) examines the tension between AI application-layer startups and foundation model providers, using Cursor’s $2.3 billion fundraising round at a $29.3 billion valuation as the central case study. The episode also covers breaking headlines including the first reported agentic cyberattack using Claude, Anthropic’s $50 billion U.S. data center commitment, Thinking Machines Lab’s reported $50–60 billion valuation round, and Google/DeepMind product updates. The host is the unnamed presenter of the AI Daily Brief podcast/video channel.

Source video URL: (not provided)

Prerequisites

  • Familiarity with the distinction between foundation models (large language models trained by labs such as OpenAI, Anthropic, Google DeepMind) and application-layer startups (companies building products on top of those models)
  • Basic understanding of ARR (Annual Recurring Revenue) as a startup metric
  • Awareness of the AI coding assistant landscape (e.g., Cursor, GitHub Copilot)
  • General knowledge of reinforcement learning and fine-tuning as model improvement techniques
  • Familiarity with enterprise software concepts: workflows, systems of record, change management, compliance

Main Points

Headline: First Agentic Cyberattack Using Claude

  • Anthropic detected a sophisticated espionage campaign in mid-September 2025, attributed with high confidence to a Chinese state-sponsored hacking group.
  • The attackers used Claude Code to automate infiltration of 30 global targets (tech companies, financial institutions, chemical manufacturers, government agencies), achieving a small number of successes.
  • Claude’s agentic capabilities executed 80–90% of the attack autonomously; human intervention was only required at a handful of key decision points.
  • Guardrails were circumvented by decomposing the attack into subtasks that each appeared individually innocuous.
  • Anthropic’s assessment: this marks an escalation from “vibe hacking” (humans still in the loop) toward fully agentic offensive operations, and less-resourced threat actors can now execute large-scale attacks.

Headline: Anthropic Commits $50 Billion to U.S. Data Centers

  • Anthropic has historically been a compute renter, relying on Google and Amazon partnerships — which provided financial flexibility but introduced trade-offs (chip constraints, rate limits, customer retention issues).
  • The company announced a $50 billion commitment to build owned U.S. data centers, with sites in Texas and New York, partnering with UK-based developer FluidStack; facilities expected to come online in 2026.
  • Framed publicly in terms of U.S. AI leadership and domestic infrastructure; CEO Dario Amodei cited acceleration of scientific discovery as a motivation.

Headline: Thinking Machines Lab Approaches $50–60 Billion Valuation

  • Mira Murati’s Thinking Machines Lab (TML) is reportedly closing a round at $50–60 billion, a ~4–5x increase from its $12 billion July 2025 valuation, less than one year after launch.
  • TML is pre-revenue; its product, the reinforcement learning platform Tinker, has some university research and enterprise users but no established business model.
  • The valuation is a talent bet, comparable to Ilya Sutskever’s Safe Superintelligence (SSI, valued at $32 billion in April 2025).

Headline: Google Adds Deep Research to Notebook LM; DeepMind Releases SEMA 2

  • Notebook LM now integrates deep research, automating the collection and synthesis of source documents; users can specify a topic and receive a full research dossier, podcast, or slide deck.
  • SEMA 2 (Scalable Instructable Multi-World Agent) achieves a 65% task success rate across game environments (up from 31% for SEMA 1), approaching human-level performance of 76%; on unseen games, success rose from ~2% to ~13%.
  • SEMA 2 was also tested in novel environments generated on the fly by DeepMind’s Genie 3 world simulation model, demonstrating cross-environment generalization.
  • DeepMind CEO Demis Hassabis described SEMA 2 as “a crucial step towards AGI.”

Headline: GPT-5.1 API Access and Prompting Guide

  • GPT-5.1 is now available via API; OpenAI published a developer prompting guide revealing design decisions.
  • The model tends toward verbosity; developers are advised to specify desired output length explicitly.
  • The model is described as significantly more steerable than prior versions, particularly useful for agentic use cases.

Main Topic: The App Layer vs. Model Layer Debate

The Bearish Case for App-Layer Startups (Yishan’s Thesis)

  • Investor Yishan argued in a widely circulated post (~20 million views) that foundation model providers will consume the application layer, but crucially, not because they are superior product builders — rather because they are the only entities with enough internal stability and resources to survive the rapid sea changes they themselves are causing.
  • Key claim: foundational AI technology has not stabilized, and sea changes now occur on a 9–12 month cycle — faster than any startup can build the “boring” business infrastructure (sales relationships, brand recognition, team solidification) needed to reach maturity.
  • Two viable paths for app startups: (1) generate cash quickly in a 12–18 month window and bank it, or (2) build a good enough product to be acquired by a major player.
  • Best-case niche: highly specialized verticals with unique, hard-to-replicate data barriers, especially those tied to physical/real-world data rather than software or finance.

The Bullish Counterarguments for App-Layer Startups

  • Vertical integration complexity: Building a functional B2B AI application requires deep last-mile work — UX, context engineering, integrations, human-in-the-loop processes, embedded workflows — that foundation model companies have no incentive to prioritize (David Roberts, Aaron Levy/Box).
    • Aaron Levy noted: the gap between 90–95% and 100% solution completion is roughly 10x more work than anticipated, involving enterprise data access, workflow integration, change management, regulatory compliance, and domain focus.
  • Proprietary behavioral exhaust: Application companies that achieve sufficient scale accumulate behavioral telemetry — every edit, action, and intent signal — that foundation model providers will never see. This data can be used to improve models and experiences in ways that are not commoditized (Natasha Malpani).
  • Habits and distribution as moats: Users don’t live in APIs; they live in experiences. Context, workflow, brand, and trust compound quickly. Distribution and feedback loops create locally compounding data advantages even as base models converge globally (Natasha Malpani).
  • Foundation models can’t care: The complexity and customization required for enterprise deployment is too granular and customer-specific for foundation model labs to prioritize; this creates a sustained window for dedicated application companies and, notably, for large systems integrators and consulting firms.

The Nuanced Middle Ground

  • Most current AI apps are indefensible: they are shallow UI wrappers with no proprietary data, no reliable scale, and no workflow depth (Jacques Reynolds, Chong-Kal).
  • The real winners at the application layer will be companies that: embed in existing workflows, write to proprietary systems of record, accumulate proprietary training data, and build reliability at scale (Nowfall).
  • Some app startups may survive by solving research and engineering problems the labs are not focused on, accumulating technical differentibility over time (Sarah Catanzaro).
  • Multimodal experiences, cornered proprietary data sets, and ecosystem breadth (e.g., Granola) are cited as structural advantages (Anisha Shara, a16z).

Case Study: Cursor’s $2.3 Billion Raise

  • Cursor raised $2.3 billion at a $29.3 billion valuation — approximately 3x its June 2025 Series C valuation of $9.9 billion, and 12x its valuation at the start of 2025.
  • Cursor reached $1 billion ARR in approximately two years — reportedly the fastest company in history to reach that milestone.
  • Cursor claims to produce more code than any other coding agent.
  • Composer 1, Cursor’s proprietary model trained on its unique behavioral data and RL environments, is now the second most popular model on the platform (after Claude Sonnet 4.5) and the fastest growing.
  • The raise is explicitly framed as investment in further developing Composer — signaling a strategic shift from relying on foundation model APIs toward owning a proprietary model layer.
  • Both OpenAI and Anthropic now directly compete with Cursor; CEO Michael Truel gave a diplomatic response while clearly prioritizing Composer development.
  • The Wall Street Journal framed the raise as a test case for whether app-layer startups can successfully transition away from dependence on foundation model providers.
  • Key observation: At sufficient scale, successful application companies may effectively become model companies — blurring the boundary between the two categories.

Key Concepts

  • Foundation model / model layer: Large-scale AI models (e.g., GPT, Claude, Gemini) developed by major labs and licensed as infrastructure to application builders.
  • Application layer / app layer: Companies that build end-user products and workflows on top of foundation models.
  • Proprietary exhaust (behavioral exhaust): Data generated by user interactions with a product — edits, actions, decisions, telemetry — that is unique to that product and can be used to fine-tune models and improve experiences.
  • Sea change: A rapid, wholesale shift in the nature of foundational AI technology that renders existing application approaches obsolete.
  • Composer 1: Cursor’s proprietary AI coding model, built on open-source foundations combined with Cursor’s unique behavioral data and reinforcement learning environments.
  • SEMA 2 (Scalable Instructable Multi-World Agent): DeepMind’s general-purpose agent capable of completing complex tasks across novel simulated game environments.
  • Genie 3: DeepMind’s world simulation model used to generate novel game environments for testing SEMA 2’s generalization.
  • ARR (Annual Recurring Revenue): A measure of predictable, annualized subscription revenue used to gauge SaaS and AI product traction.
  • Last-mile integration: The highly specific, customer-facing work required to make a general AI capability functional within a particular enterprise’s systems, workflows, and compliance environment.
  • Agentic cyberattack: A cyber intrusion in which AI agents autonomously execute multi-step attack sequences with minimal human direction.
  • Vibe hacking: An earlier, lower-autonomy form of AI-assisted hacking in which humans remain actively in the loop directing operations (Anthropic’s term).
  • Tinker: Thinking Machines Lab’s reinforcement learning platform — their first released product.
  • Neo Labs: A term used by The Information to describe a new wave of AI startups positioning themselves as alternatives to or competitors of OpenAI and Anthropic.

Summary

The central argument of this episode is that the question of whether AI application-layer startups or foundation model providers will dominate is genuinely unresolved, but the terms of the debate are becoming clearer. Investor Yishan’s widely-read thesis — that app startups will be continuously outpaced by sea changes in the underlying technology before they can mature into real businesses — captures a real and serious risk, but misses several structural advantages available to application companies that go deep enough: the complexity of enterprise last-mile integration that model labs cannot prioritize, the compounding value of proprietary behavioral exhaust inaccessible to anyone else, and the habits and trust built through embedded workflows. The consensus emerging from the investment and operator community is that most current AI apps are indefensible wrappers, but a small number of companies that achieve genuine depth, proprietary data advantages, and reliable autonomous value delivery can survive and thrive. Cursor’s $2.3 billion raise and $29.3 billion valuation — built on reaching $1 billion ARR faster than any company in history and backed by an increasingly proprietary model stack — is the strongest current evidence that this path exists. The broader implication is that the most successful application-layer survivors may ultimately evolve into model companies themselves, making the app-versus-model framing a transitional rather than permanent distinction.