OpenAI optimizes ChatGPT almost entirely for long-term retention, not revenue or raw user count.
Turley says that if forced to allocate 100 points across every product metric, he would put nearly all of them on long-term retention, because it is the real signal that ChatGPT is solving people's problems durably. Revenue, he argues, follows from that rather than being a target in itself, and he points to OpenAI's decision to unlock GPT-4 from behind a paywall as a case where optimizing for access and customer value ended up being revenue-positive anyway.
retention-over-growth
ChatGPT's smiling retention curve comes from search and personalization, not one single feature.
Turley attributes the shift from a "worky" product with weekend and summer usage dips to a mobile-first, always-relevant product mainly to two investments: search (which supplies daily-value use cases) and personalization (which makes the product get to know the user over time). He frames retention gains broadly as many small systematic improvements rather than any single breakthrough.
retention-over-growth
Distribution alone did not determine the consumer AI winner, contradicting Turley's own prior expectations.
Turley says he expected AI to follow the historical winner-take-all pattern of search (Google), mobile (Apple), and social (Meta), where the incumbent with the most distribution wins. Instead OpenAI reached roughly 900 million weekly active users despite Google and Meta's much larger existing user bases, showing that distribution was necessary but not sufficient.
ai-distribution-and-competition
OpenAI attributes its historical growth to roughly equal thirds: friction removal, core product investment, and raw model improvement.
Turley breaks down growth drivers into three roughly equal buckets: classic friction removal (e.g., removing the mandatory login wall before using ChatGPT), core product investments built jointly by research and product teams (search, personalization, writing UI blocks), and step-change model upgrades (GPT-3.5 to GPT-4, GPT-4 coming out from behind the paywall, and continuous point releases like 5.3/5.4).
ai-distribution-and-competition
ChatGPT's next stage of growth depends on becoming more proactive and action-taking, not just conversational.
Turley argues most people do not naturally know how to delegate problems to an AI system, so the product needs affordances that make its value obvious rather than requiring users to discover prompts on their own. He describes two converging shifts: chat should be able to take actions (not just answer questions) and chat should proactively surface value without being prompted, and says combining the two is what makes a product feel like a genuine "super assistant."
agentic-and-proactive-ai
General-purpose agentic AI is bottlenecked by trust, which requires early real-world attempts even if they mostly fail, and domain-specific agents like Codex show the pattern already working.
Turley says OpenAI's earlier ChatGPT agent launched slightly too early; the model wasn't good enough to build user trust, so people didn't try using it for real problems. He contrasts this with ChatGPT's original text launch and with Codex, where engineers who no longer open an IDE prove domain-specific coding agents have already crossed an escape-velocity threshold, and expects the same testable, RL-friendly pattern to extend to other domains before general-purpose agents mature.
agentic-and-proactive-ai
OpenAI expects subscription pricing to evolve because flat-rate access breaks down as power users consume disproportionate compute.
Turley compares the current situation to an unlimited electricity plan: it stops making economic sense once usage variance across users gets large enough. He frames OpenAI's original move to subscriptions as an accident of scaling GPT-4 rather than a deliberate monetization strategy, and says the SKUs will keep changing as compute-intensive breakthroughs like test-time compute increase the gap between casual and power users.
ai-business-model-evolution
OpenAI is piloting ads specifically to extend access to users who cannot or will not pay via subscription, not primarily for revenue.
Turley says subscriptions fail in many markets where people lack or don't use credit cards, so ads are framed internally as an access mechanism to bring intelligence to more of the world, with heavy internal work on principles (answer independence, privacy) done before the pilot launched. He notes the most common support inquiry about ads has been advertisers asking how to buy them, not users asking how to disable them.
ai-business-model-evolution
GPU capacity, not revenue-per-GPU math, is OpenAI's real allocation constraint, and it is getting worse, not better, as usage matures.
Turley says OpenAI prioritizes serving existing users reliably first, then allocates new capacity partly on a naive revenue basis and partly on high-conviction bets on entirely new capabilities (deep research had no provable demand before launch). He notes GPU demand is rising even as unit prices fall, because the value delivered per user keeps increasing, particularly in enterprise token consumption, and unlike hiring humans, GPU supply is a hard zero-sum resource with no simple scaling lever.
ai-business-model-evolution
The most valuable differentiator for OpenAI going forward is the team's ability to keep synthesizing usable products out of raw model capability, since any given feature can be copied.
Turley argues that competitors can replicate individual ChatGPT features, so durable advantage comes from the cross-functional team (research, engineering, design) repeatedly finding the intersection of what's useful and what's newly possible, faster than competitors can copy. He cites OpenAI's internal "Code Red" focus period, triggered by competitive pressure including Google's strong model and public switching by figures like Marc Benioff, as the mechanism the company uses to realign the whole org around fundamentals like reliability and latency.
ai-distribution-and-competition
Curiosity, not any specific technical skill, is the durable human advantage in an AI-saturated world.
Turley's advice to students is that if a machine can answer any question, the differentiator becomes asking good questions, which only comes from genuinely pursuing what excites you rather than optimizing for a career path. He connects this to his own path into OpenAI, which started from being "nerd sniped" out of curiosity rather than a deliberate plan.
human-skills-in-the-ai-era
Clear writing remains a valuable skill even as AI gets better at writing, because articulating precise intent to a model is itself a writing and thinking skill.
Turley argues that although AI will match or exceed human writing quality, expressing exactly what you want from a machine still forces clarity of thought, so professions built on precise writing and thinking stay well-positioned. He adds there will be a permanent need for high-quality, trusted, authoritative source content even as AI tools get better at helping people discover it.
human-skills-in-the-ai-era
Books referenced
Running Down a Dream - Bill Gurley - Referenced when Turley named curiosity as the most important permanent skill in the AI era
Media referenced
NotebookLM - other - Praised by both host and guest as an innovative, differentiated product that transforms information into new mediums for learning
Codex - other - Cited as OpenAI's proof that domain-specific coding agents have already reached escape velocity with real users
Pulse - other - OpenAI's early proactive product, discussed as a first step toward a chat interface that prompts the user instead of the reverse
Deep Research - other - Cited as OpenAI's first agentic consumer product and an example of shipping to discover unproven demand
OpenClaw - other - A project by Peter (recently hired at OpenAI) that Turley says clarified an embodied, stateful, multi-surface AI assistant vision for the whole industry
Companies
OpenAI - Turley's employer; maker of ChatGPT, discussed throughout as the subject company
Instacart - Turley's employer immediately before OpenAI, delivering groceries before he moved to "delivering AGI"
Dropbox - Where Turley worked with Joanne, the OpenAI contact who got him off the DALL-E 2 waitlist and into an interview
Google - Cited as the historical winner-take-all precedent in search (90%+ market share) that Turley expected, wrongly, would repeat in AI via Gemini distribution
Apple - Cited alongside Google/Meta as a distribution incumbent, and named as one of OpenAI's two major recent partnerships (with Gemini bundling)
Meta - Cited as the social-market winner-take-all precedent that did not repeat for OpenAI in consumer AI despite Meta's distribution advantage
Uber - Used as the analogy for how ChatGPT's power-user pricing problem resembles the 2015-era Uber/Lyft subsidized-growth dynamic
Lyft - Same Uber-era pricing analogy for heavy users consuming outsized value on flat subscriptions
Reliance - Named as one of OpenAI's two big recent distribution partnerships, reaching a large Indian user base
Techniques and frameworks
Code Red - OpenAI's internal cross-team focus mechanism, invoked end of last year to prioritize reliability, latency, and personalization ahead of new capabilities; declared over with the 5.3/5.4 launches
Smile curve retention - The shape of ChatGPT's user retention/reactivation curves, cited as evidence the product keeps pulling lapsed users back rather than losing them permanently
Progressive disclosure (macOS model) - Turley's stated design inspiration for ChatGPT: simple and magical for novices, but with all the knobs exposed for power users, same as macOS
Summary
Apoorv Agrawal interviews Nick Turley, OpenAI's Head of ChatGPT, about how a product that started as a one-month demo scaled to roughly 900 million weekly active users and where it goes next. Turley opens by rejecting the idea of a single north-star metric: he says he would put nearly all his points on long-term retention if forced to choose, because durable return usage is the real proof a product is solving people's problems, and revenue and growth tend to follow from that rather than being chased directly. He credits ChatGPT's "smiling" retention curve to search and personalization investments that turned a workday-only tool into a mobile-first, always-relevant one, alongside a long tail of unglamorous friction removal (like eliminating the login wall) and step-change model releases, which he breaks down as roughly a third each of the historical growth.
A large portion of the conversation is about why ChatGPT beat the winner-take-all distribution logic Turley himself expected to apply, given Google's and Meta's much larger existing user bases, and what comes next for the "next billion" users. His answer centers on making the product more proactive and capable of taking real actions rather than just answering questions, because most people do not intuitively know how to delegate problems to an AI. He frames the earlier ChatGPT agent as launched slightly too early to earn user trust, contrasts that with Codex's escape velocity in coding, and argues the same domain-specific agentic pattern will extend to other RL-friendly domains before general-purpose agents mature. Proactive products like Pulse are described as an early step toward a model that prompts the user instead of the reverse, ultimately combining action-taking and proactivity into what he calls a genuine "super assistant."
On the business model, Turley says subscription pricing is inherently temporary: as power users consume increasingly large amounts of compute (an "unlimited electricity plan" that doesn't scale), pricing structures will keep evolving, and he frames OpenAI's original move to subscriptions as an accidental byproduct of GPT-4 capacity constraints rather than deliberate strategy. Ad pilots are framed explicitly as an access mechanism for users without credit cards or subscription habits, not primarily as a revenue play, with heavy internal work on principles like answer independence before launch. He is candid that GPU capacity, not clean revenue-per-GPU math, is the company's hardest constraint, and that it is intensifying rather than easing as enterprise token consumption grows, because unlike human hiring, GPU supply is genuinely zero-sum.
The conversation closes with a rapid-fire segment covering Turley's investment interest in AI-enabled professional services companies, praise for NotebookLM as a genuinely differentiated product, and career advice for students: curiosity is the durable "permanent skill" because a machine that can answer any question makes the ability to ask good ones the scarce resource, and clear writing remains valuable because it forces precise thinking even as AI writing quality improves. He also recounts several personal "AGI moments," from GPT-4 unexpectedly writing poetry and working code, to watching a reasoning model swear and self-correct mid-demo in a moment of emergent behavior that "completely blew my mind."
Notable Quotes
"We've got about 10% of the world coming to us now. 90% left to go, right? There's so much more opportunity to reach more people." - Nick Turley
"It's less about chat and more about natural language to me... chat is a great way of expressing your intent. It's a good way of communicating with the machine, but it's not a great output." - Nick Turley
"If the machine can answer all your questions, you better have good questions." - Nick Turley
"GPUs are zero sum, and if you don't have more GPUs you really have to figure out how do you make very very hard trades, and I hate making hard trades for users." - Nick Turley
"I don't think delegation is a natural skill for most... the product it's like a raw appliance." - Nick Turley