Ben Thompson on OpenAI, Commerce, and Ads

There's a certain kind of tech conversation that only happens when two people who actually build things stop performing for an audience and start thinking out loud. This was that. Powered by Guinness.

Thompson and Collison covered life in Taiwan, the TikTok deal, and whether sports teams become more valuable in an AI world (they do). But the substance that matters most for anyone building in or around AI right now lives in three sections: how ads should work in AI products, the four levels of agentic commerce, and what's actually happening to SaaS.

OpenAI Is Doing Ads Wrong — And Thompson Can't Stop Talking About It

Thompson's frustration with OpenAI's ad strategy is visceral. ChatGPT launched banner ads targeted to the context of your conversation. Thompson thinks this is backwards.

The problem: if an ad is clearly connected to your answer, you start questioning the answer. "My T-shirt answers questions that my T-shirt is raising." The better model, Thompson argues, is Meta's approach — build a deep understanding of the user and show them things they didn't know they wanted, completely disconnected from the current query.

His ideal for Google: never put ads in Gemini at all. Use Gemini conversations to sharpen the targeting across YouTube, Search, and every other Google property. The signal from what you discuss with AI is extraordinarily rich — but the ad should manifest elsewhere.

OpenAI's core problem? They only have one surface to monetize: ChatGPT itself. And they're two to three years behind where they need to be.

The Four Levels of Agentic Commerce

Collison laid out a framework for how agentic commerce actually develops, bottom-up rather than top-down. Thompson mostly agreed, then added a crucial layer.

Level 1: Replace form-filling. You find the winter jacket you want, paste the URL into ChatGPT, and say "buy this for me." No clicking around unfamiliar sites. Just execution. Thompson calls this "just in time UI" and is genuinely optimistic about it.

Level 2: Better discovery and search. Keyword search is broken for anything you can't name precisely. Natural language product search — "I need a jacket for 20-degree weather, I like this style, here's my budget" — is already manifesting.

Level 3: Persistent user profiles. Pin things you like as you go. Share a Pinterest board of styles. Let the AI build an understanding of your preferences that compounds over time.

Level 4 (Thompson's addition): Anticipatory commerce. This is where it gets interesting. Imagine it's October 1st and ChatGPT shows you an ad for a winter jacket perfectly suited to you — not because you searched for one, but because the system knows winter is coming and you haven't prepared. This is what Meta already does on Instagram. Thompson's argument is that OpenAI should be building this, and they should have started three years ago.

The uncomfortable implication: Zuck may have been right when he said Facebook advertising is already the world's largest and most successful agent. You tell it what a customer acquisition is worth. It delivers. Everything else is autopilot.

For agents to operate at every level of this stack, they need access to tools, payments, and inference without per-vendor wiring. That's the infrastructure gap ATXP agents are designed to close.

SaaS Isn't Dead, But the Growth Story Might Be

Collison asked the question directly: the public markets say SaaS is canceled. Are they right?

Thompson's answer is more nuanced than either side of the debate. The "going to zero" crowd is wrong — systems of record like Workday and Stripe Billing aren't getting Claude Code'd out of existence anytime soon. But the growth assumption that underpins SaaS valuations and compensation structures is genuinely at risk.

Two pressures converge:

• Seat-based pricing meets headcount compression. If companies are growing slower — or shrinking — the per-seat model mechanically contracts.
• AI enables smaller teams and self-serve. Thompson sees a world of more individual entrepreneurs and small teams who don't fit the Salesforce-driven, seat-based model. They might return to self-serve tools, or just build what they need.

The shift from "growth company" to "stable business" is a valuation haircut even if revenue holds. Combined with compensation structures predicated on indefinite growth, it's a structural problem.

This is exactly why pay-as-you-go matters for agents. No seats, no contracts — just usage-based execution that scales with what the agent actually does.

The TSMC Brake: A 2029 Chip Crunch Nobody's Pricing In

This might be Thompson's strongest conviction of the past two years. TSMC is a monopolist in logic chip fabrication, and they are rationally under-building capacity because the downside of overcapacity is catastrophic for fabs.

After ChatGPT launched, TSMC actually decreased CapEx for two consecutive years. Even their increase to $60 billion this year is a smaller percentage jump than last year.

Thompson's prediction: a massive chip shortage around 2029, especially as agentic AI multiplies compute density far beyond what human-driven interactions require. An agent doing commerce on your behalf generates orders of magnitude more computation per unit of time than you Googling.

His prescription: hyperscalers need to invest in making Intel and Samsung credible competitors. Not as charity — as economics. The foregone revenue from chip shortages in 2029 will dwarf what it costs to diversify the supply chain now. Geopolitical insurance comes free.

The Thread That Connects Everything

There's a pattern running through the entire conversation that's worth naming: the gap between what's rational for individuals and what's optimal for the system.

TSMC rationally under-builds. Newsletter writers rationally refuse bundles. Cable channels rationally go direct-to-consumer. OpenAI rationally ships the easiest ad product. In every case, the individually rational decision creates a worse collective outcome.

Thompson doesn't resolve this tension — he just keeps making it visible. That's the value.

The full interview is on Stratechery and the Cheeky Pint YouTube. It's worth the full 90 minutes.