PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: AI advertising

  • Ben Thompson on the Future of AI Ads, The SaaS Reset, and The TSMC Bottleneck

    Ben Thompson, the author of Stratechery and widely considered the internet’s premier tech analyst, recently joined John Collison for a wide-ranging discussion on the Stripe YouTube channel. The conversation serves as a masterclass on the mechanics of the internet economy, covering everything from why Taiwan is the “most convenient place to live” to the existential threat facing seat-based SaaS pricing.

    Thompson, known for his Aggregation Theory, offers a contrarian defense of advertising, a grim prediction for chip supply in 2029, and a nuanced take on why independent media bundles (like Substack) rarely work for the top tier.

    TL;DW (Too Long; Didn’t Watch)

    The Core Thesis: The tech industry is undergoing a structural reset. Public markets are right to devalue SaaS companies that rely on seat-based pricing in an AI world. Meanwhile, the “AI Revolution” is heading toward a hardware cliff: TSMC is too risk-averse to build enough capacity for 2029, meaning Hyperscalers (Amazon, Google, Microsoft) must effectively subsidize Intel or Samsung to create economic insurance. Finally, the best business model for AI isn’t subscriptions or search ads—it’s Meta-style “discovery” advertising that anticipates user needs before they ask.

    Key Takeaways

    • Ads are a Public Good: Thompson argues that advertising is the only mechanism that allows the world’s poorest users to access the same elite tools (Search, Social, AI) as the world’s richest.
    • Intent vs. Discovery: Putting banner ads in an AI chat (Intent) is a terrible user experience. Using AI to build a profile and show you things you didn’t know you wanted (Discovery/Meta style) is the holy grail.
    • The SaaS “Correction”: The market isn’t canceling software; it’s canceling the “infinite headcount growth” assumption. AI reduces the need for junior seats, crushing the traditional per-seat pricing model.
    • The TSMC Risk: TSMC operates on a depreciation-heavy model and will not overbuild capacity without guarantees. This creates a looming shortage. Hyperscalers must fund a competitor (Intel/Samsung) not for geopolitics, but for capacity assurance.
    • The Media Pond Theory: The internet allows for millions of niche “ponds.” You don’t want to be a small fish in the ocean; you want to be the biggest fish in your own pond.
    • Stripe Feedback: In a candid moment, Thompson critiques Stripe’s ACH implementation, noting that if a team add-on fails, the entire plan gets canceled—a specific pain point for B2B users.

    Detailed Summary

    1. The Geography of Convenience: Why Taiwan Wins

    The conversation begins with Thompson’s adopted home, Taiwan. He describes it as the “most convenient place to live” on Earth, largely due to mixed-use urban planning where residential towers sit atop commercial first floors. Unlike Japan, where navigation can be difficult for non-speakers, or San Francisco, where the restaurant economy is struggling, Taiwan represents the pinnacle of the “Uber Eats” economy.

    Thompson notes that while the buildings may look dilapidated on the outside (a known aesthetic quirk of Taipei), the interiors are palatial. He argues that Taiwan is arguably the greatest food delivery market in history, though this efficiency has a downside: many physical restaurants are converting into “ghost kitchens,” reducing the vibrancy of street life.

    2. Aggregation Theory and the AI Ad Model

    The most controversial part of Thompson’s analysis is his defense of advertising. While Silicon Valley engineers often view ads as a tax on the user experience, Thompson views them as the engine of consumer surplus. He distinguishes between two very different types of advertising for the AI era:

    • The “Search” Model (Google/Amazon): This captures intent. You search for a winter jacket; you get an ad for a winter jacket. Thompson argues this is bad for AI Chatbots because it feels like a conflict of interest. If you ask ChatGPT for an answer, and it serves you a sponsored link, you trust the answer less.
    • The “Discovery” Model (Meta/Instagram): This creates demand. The algorithm knows you so well that it shows you a winter jacket in October before you realize you need one.

    The Opportunity: Thompson suggests that Google’s best play is not to put ads inside Gemini, but to use Gemini usage data to build a deeper profile of the user, which they can then monetize across YouTube and the open web. The “perfect” AI ad doesn’t look like an ad; it looks like a helpful suggestion based on deep, anticipatory profiling.

    3. The “End” of SaaS and Seat-Based Pricing

    Is SaaS canceled? Thompson argues that the public markets are correctly identifying a structural weakness in the SaaS business model: Headcount correlation.

    For the last decade, SaaS valuations were driven by the assumption that companies would grow indefinitely, hiring more people and buying more “seats.” AI disrupts this.

    “If an agent can do the work, you don’t need the seat. And if you don’t need the seat, the revenue contraction for companies like Salesforce or Box could be significant.”

    The “Systems of Record” (databases, HR/Workday) are safe because they are hard to rip out. But “Systems of Engagement” that charge per user are facing a deflationary crisis. Thompson posits that the future is likely usage-based or outcome-based pricing, not seat-based.

    4. The TSMC Bottleneck (The “Break”)

    Perhaps the most critical macroeconomic insight of the interview is what Thompson calls the “TSMC Break.”

    Logic chip manufacturing (unlike memory chips) is not a commodity market; it’s a monopoly run by TSMC. Because building a fab costs billions in upfront capital depreciation, TSMC is financially conservative. They will not build a factory unless the capacity is pre-sold or guaranteed. They refuse to hold the bag on risk.

    The Prediction: Thompson forecasts a massive chip shortage around 2029. The current AI boom demands exponential compute, but TSMC is only increasing CapEx incrementally.

    The Solution: The Hyperscalers (Microsoft, Amazon, Google) are currently giving all their money to TSMC, effectively funding a monopoly that is bottlenecking them. Thompson argues they must aggressively subsidize Intel or Samsung to build viable alternative fabs. This isn’t about “patriotism” or “China invading Taiwan”—it is about economic survival. They need to pay for capacity insurance now to avoid a revenue ceiling later.

    5. Media Bundles and the “Pond” Theory

    Thompson reflects on the success of Stratechery, which was the pioneer of the paid newsletter model. He utilizes the “Pond” analogy:

    “You don’t want to be in the ocean with Bill Simmons. You want to dig your own pond and be the biggest fish in it.”

    He discusses why “bundling” writers (like a Substack Bundle) is theoretically optimal but practically impossible.

    The Bundle Paradox: Bundles work best when there are few suppliers (e.g., Spotify negotiating with 4 music labels). But in the newsletter economy, the “Whales” (top writers) make more money going independent than they would in a bundle. Therefore, a bundle only attracts “Minnows” (writers with no audience), making the bundle unattractive to consumers.

    Rapid Fire Thoughts & “Hot Takes”

    • Apple Vision Pro: A failure of imagination. Thompson critiques Apple for using 2D television production techniques (camera cuts) in a 3D immersive environment. “Just let me sit courtside.”
    • iPhone Air: Thompson claims the new slim form factor is the “greatest smartphone ever made” because it disappears into the pocket, marking a return to utility over spec-bloat.
    • Tik Tok: The issue was never user data (which is boring vector numbers); the issue was always algorithm control. The US failed to secure control of the algorithm in the divestiture talks, which Thompson views as a disaster.
    • Crypto: He remains a “crypto defender” because, in an age of infinite AI-generated content, cryptographic proof of authenticity and digital scarcity becomes more valuable, not less.
    • Work/Life Balance: Thompson attributes his success to doubling down on strengths (writing/analysis) and aggressively outsourcing weaknesses (he has an assistant manage his “Getting Things Done” file because he is incapable of doing it himself).

    Thoughts and Analysis

    This interview highlights why Ben Thompson remains the “analyst’s analyst.” While the broader market is obsessed with the capabilities of AI models (can it write code? can it make art?), Thompson is focused entirely on the value chain.

    His insight on the Ad-Funded AI future is particularly sticky. We are currently in a “skeuomorphic” phase of AI, trying to shoehorn chatbots into search engine business models. Thompson’s vision—that AI will eventually know you well enough to skip the search bar entirely and simply fulfill desires—is both utopian and dystopian. It suggests that the privacy wars of the 2010s were just the warm-up act for the AI profiling of the 2030s.

    Furthermore, the TSMC warning should be a flashing red light for investors. If the physical layer of compute cannot scale to meet the software demand due to corporate risk aversion, the “AI Bubble” might burst not because the tech doesn’t work, but because we physically cannot manufacture the chips to run it at scale.

  • Unlocking the Future of Audio: NVIDIA’s Fugatto Transforms Sound Synthesis and Transformation

    NVIDIA’s Fugatto is a generative AI model for advanced audio synthesis and transformation. Using text and audio as inputs, it creates or modifies music, voices, and sounds with precision. Features include ComposableART for combining attributes like emotion and accent, and temporal interpolation for evolving soundscapes. Trained on 50,000+ hours of curated datasets, Fugatto powers applications in music production, gaming, language learning, and advertising. It supports emergent capabilities, like creating unheard sounds or blending tasks, positioning itself as a versatile tool for audio innovation.


    NVIDIA’s Fugatto, or Foundational Generative Audio Transformer Opus 1, is a groundbreaking generative AI model redefining audio synthesis and transformation. This advanced model seamlessly combines audio and text to create versatile outputs, from unique soundscapes to voice modulation, offering unprecedented flexibility to industries like music, gaming, and education.

    Key Features of Fugatto

    1. Multimodal Capabilities

    • Inputs: Text, audio, or a combination.
    • Outputs: Music snippets, modified voices, or entirely new sounds.

    Fugatto’s versatility allows users to generate diverse outputs. Whether it’s creating the sound of a barking saxophone or fine-tuning a voice’s emotion and accent, the possibilities are vast.

    2. Composable Audio Representation Transformation (ComposableART)

    • Customization: Combines multiple attributes (e.g., emotions, accents) into unique outputs.
    • Temporal Interpolation: Enables dynamic changes over time, such as simulating a storm that transitions to calm.

    This technique gives artists and developers granular control over their audio creations.

    How Fugatto Works

    Data and Training

    Fugatto was trained using NVIDIA’s DGX systems on over 50,000 hours of curated audio datasets. The training leveraged:

    • Free-form Instructions: Generated via large language models (LLMs).
    • Synthetic Captioning: Augmented datasets with AI-generated descriptions for better context and task diversity.

    Advanced Modeling Techniques

    • Optimal Transport Conditional Flow Matching (OT-CFM): Powers Fugatto’s ability to synthesize and transform audio precisely.
    • Adaptive Layer Norm and Specialized Architectures: Enable robust performance across a variety of audio tasks.

    Real-World Applications

    1. Music Production
      • Rapidly prototype music ideas by modifying style, instruments, or vocals.
      • Enhance existing tracks with effects or improved quality.
    2. Gaming
      • Dynamically adapt game soundtracks based on player interactions.
      • Generate unique audio assets on the fly for immersive experiences.
    3. Language Learning
      • Personalize lessons with voices that mimic familiar accents or tones.
      • Create engaging, adaptive audio content for learners.
    4. Advertising and Media
      • Localize campaigns by adjusting accents and emotional tones for regional markets.
      • Create novel sound effects to enhance brand identity.

    Emergent Capabilities: Beyond Conventional Audio Models

    Fugatto excels where traditional models fall short:

    • Emergent Sound Generation: Create sounds beyond the scope of its training data, such as a cello that mimics a human voice.
    • Task Composition: Combine previously unrelated tasks, like speech synthesis paired with environmental soundscapes.

    The Future of Audio AI

    Fugatto represents a leap toward unsupervised multitask learning in audio. As NVIDIA continues to refine this model, potential enhancements include:

    • Improved Dataset Scaling: Incorporating more diverse datasets to unlock new creative potentials.
    • Latent Representations: Supporting stereo and low-frequency audio for richer soundscapes.

    Just Wow

    Fugatto isn’t just a tool—it’s a creative partner for anyone working with sound. From revolutionizing the music industry to enhancing the gaming experience, this model is poised to set new benchmarks in generative AI. Whether you’re a producer, developer, or educator, Fugatto opens doors to unprecedented possibilities in audio creation.

    For more details and sound demos, visit Fugatto’s official website.