Aryan's Brief

SemiAnalysis

GPT-5 Set the Stage for Ad Monetization and the SuperApp

Aug 13, 2025 · 30 min · analyzed May 27, 2026
#business

Source →

1) Core thesis

GPT-5’s “router” — the mechanism that dynamically allocates queries between cheap mini-models and expensive reasoning models — is not primarily a performance feature. It is the infrastructure for monetizing ChatGPT’s 700M+ free users through agentic purchasing with transaction take-rates, potentially bypassing Google’s search ad model entirely.

2) Claim and Evidence

  • Claim: The router is an economic sorting mechanism that can differentiate low-value queries from high-commercial-intent queries, enabling variable compute spend based on expected conversion value.

  • Evidence: “A trivial information query like ‘Why is the sky blue?’” can be routed to GPT-5 mini with near-zero cost. A “highly commercial query like ‘What is the best DUI lawyer near me’” — worth thousands in conversion value — could justify “$50 dollars of compute” with agentic outreach to lawyers [~line 75]. The router “continuously learns on preference rates” and “just takes a single additional attribute to begin the path to monetization: the commercial value of the query.”

  • Strength: moderate — the technical architecture of the router is real and described in OpenAI’s release notes, but the $50 compute figure and the lawyer scenario are speculative, not observed behavior.

  • Claim: OpenAI’s hiring of Fidji Simo (ex-Facebook monetization, ex-Instacart CEO) as head of Applications, combined with Sam Altman’s tone shift on ads, signals an imminent monetization push.

  • Evidence: Simo “might be one of the most qualified individuals alive to turn high-intent internet properties into ad products.” Altman went from “I hate ads… ads plus AI are uniquely unsettling” to “I am not totally against it… maybe if you click on something… we’ll get a bit of transaction revenue.” Instacart launched agent-checkout features while Simo was there, and she joined OpenAI shortly after [~line 87].

  • Strength: strong — the personnel move and Altman’s documented statements are verifiable facts. The inference about intent is reasonable but not proven.

  • Claim: Agentic purchasing via take-rates is a fundamentally different model from search ads because AI has “scaling marginal costs” — more compute produces better answers, unlike search’s fixed-supply response.

  • Evidence: “For the first time, the more you spend the better your result is because of CoT reasoning tokens… There is a somewhat direct relationship between more money, more compute, and a better answer.” Search shows the same page ranking regardless of query difficulty. With routing, ChatGPT can dynamically allocate compute. OpenAI is already partnering with Stripe, Visa, PayPal, Shopify, Instacart, Booking.com [~line 107].

  • Strength: moderate — the scaling-marginal-costs distinction from search is real and insightful. But the partnership list includes standard API integrations, not necessarily agentic-commerce deals.

  • Claim: AI labs are actively building agentic purchasing capability, paying startups “hundreds of thousands of dollars to spin up replicas of popular sites like DoorDash and Amazon to RL agents on successfully completing end-to-end transactions.”

  • Evidence: Coupled with the Instacart agent-checkout feature, OpenAI-Shopify checkout integration already in development. “It is not a question of if, but when this capability happens” [~line 89].

  • Strength: moderate — the Instacart and Shopify integrations are concrete. The RL-training-on-site-replicas claim is second-hand and unverifiable.

3) Mechanisms

The causal chain: free users → unified router → query intent classification → variable compute allocation → agentic action (purchasing, booking, scheduling) → transaction take-rate revenue. The key insight is that this model aligns incentives: the user gets a better answer for high-stakes queries, the merchant gets a high-conversion lead, and OpenAI captures a fee. Unlike display ads, there’s no intrusive ad placement — the model’s output is the product. The second-order mechanism: this collapses the traditional purchase funnel (search → research → compare → purchase) into a single AI-mediated action, cutting out Google’s ad auction entirely. Google’s moat was being the starting point for commercial queries. If ChatGPT becomes that starting point, Google’s search ad revenue is structurally threatened even if query volumes don’t decline.

4) Concrete actions

  • If building products that depend on search-driven customer acquisition: prepare for a world where purchase-intent queries shift from Google to AI assistants. The affiliate/take-rate model means you may pay OpenAI instead of Google for the same customer.
  • If investing in ad-tech or digital advertising: the distinction between “search ads” and “agentic commerce” is not semantic — the ad auction model (cost-per-click) doesn’t translate to agentic purchasing (cost-per-conversion). Companies positioned for conversion-based pricing benefit.
  • For product teams: OpenAI’s partnership list (Stripe, Shopify, Instacart, Booking.com) is the early-adopter map. If you’re not on it, your competitor likely is. Agentic checkout integration is a 2025-2026 priority, not a 2027 roadmap item.
  • If evaluating ChatGPT’s competitive position: the free user monetization path makes the “Pro/Plus subscribers vs. free users” framing obsolete. The real revenue opportunity is the 700M unmonetized users, not the single-digit millions of paying subscribers.

5) Delta vs prior episodes

This article predates other SemiAnalysis digests in the archive but overlaps with themes in the Amazon/Trainium piece (Sep 2025): both examine how AI labs’ business models drive infrastructure decisions. The Trainium article focuses on how Anthropic’s training spend flows to cloud providers; this article focuses on how OpenAI’s consumer monetization shapes product architecture. Both share the thesis that AI’s business model evolution determines hardware/infrastructure winners. What’s notable: this article’s prediction that “ChatGPT will become an agent to help users make purchasing decisions” has aged well — agentic commerce demos from OpenAI, Google, and Anthropic in late 2025/early 2026 have validated the direction.

6) Red flags

  • The entire article is an investment thesis: “OpenAI is firmly knocking on the door of technology giants Google and Meta.” The author doesn’t disclose whether SemiAnalysis or its subscribers hold positions in any of these companies, though this is standard for the publication’s model.
  • The “700M+ free users” figure is presented as fact — OpenAI hasn’t published official MAU figures consistently. The comparison to WhatsApp, Wikipedia, and Instagram assumes ChatGPT’s engagement depth matches these platforms, which it almost certainly doesn’t.
  • The article dismisses display ads (“Perplexity has tried this, and it doesn’t seem to be going that well”) but provides no data on Perplexity’s ad performance. This is stated as fact without evidence.
  • “AI labs are paying startups hundreds of thousands of dollars to spin up replicas of popular sites” — this is a significant claim with no named sources, companies, or dollar figures beyond the vague range. If true, it’s a major story. If false, it’s fabrication.
  • The agentic purchasing end state assumes seamless integration across payment systems, merchant APIs, and user authentication that does not exist today. The article acknowledges this (“we are very far from that future”) but then writes as if it’s imminent.
  • The “only one company is growing users at a meaningful rate. It’s OpenAI” claim in the final paragraph is unsupported. Meta, TikTok, and YouTube all report user growth.

7) Open questions

  • If agentic purchasing with take-rates becomes the model, what’s the take-rate? 3% like credit cards? 15% like app stores? The economics of the entire model depend on this number, and the article doesn’t estimate it.
  • What happens to the router’s incentives if it’s optimizing for both user satisfaction AND commercial conversion? The alignment problem isn’t just technical — it’s a business-model conflict.
  • How does Google respond? Google has its own LLMs (Gemini), its own commerce integrations (Google Shopping, Pay), and its own massive user base. The article frames this as OpenAI vs. Google but doesn’t model Google’s countermove.
  • The article focuses entirely on OpenAI. What about Anthropic, Meta, and xAI — all of whom have their own consumer AI products and monetization challenges?