The Next Distribution Platform Is Coming, and ChatGPT Is the Likeliest Host

Balfour opens with a point he says he has spent his career repeating: building a great product is necessary but not sufficient, and what separates winners is distribution. He cites an Andreessen Horowitz partner, Alex Rampell, who argued roughly a decade ago that startups are a game of getting distribution before the incumbent can copy you. In Balfour's telling, that race has become much harder.

He names three reasons. Incumbents copy faster, so the window to reach escape velocity has shrunk. Organic distribution has dwindled, with SEO and clicks declining and social platforms such as LinkedIn and X sending less traffic to outside sites. And AI makes code generation so easy that competition among startups feels infinite, with YC producing "six of the same thing" every cohort.

There are exceptions, and Balfour points to Cursor, which he says overtook GitHub Copilot's market share in nine months or less. AI acted as a spark that excited early adopters. But the most reliable way to beat incumbents, he says, is when a new distribution platform emerges, because startups move faster than incumbents to exploit it.

Balfour draws on Casey Winters' observation that the AI shift is a technology shift without a distribution shift yet. Historically, big technology shifts such as the internet, cloud, mobile and social came first, with new distribution platforms arriving a little later. Two years on, he says, the ingredients for a new platform are all in place, and Lenny Rachitsky adds that the usual channels like SEO, paid and word of mouth are saturated, so an unsaturated one is a rare opportunity.

The cycle starts at Step Zero, where there is consensus a huge new category is coming but no clear winner, and five to seven players fight fiercely because these markets end in monopolies or duopolies. In Step One, someone identifies a moat and needs an ecosystem of third-party developers and content creators to build it faster. They get distribution in exchange. Step Two is the platform opening up, and Step Three is the closing.

In the closing phase, platforms lock down for monetization, growth, or to avoid being disrupted by their own ecosystem. They either shut access, build first-party apps to absorb the best use cases, or artificially depress the organic reach they once granted so that developers move to paid. Balfour is also confident enough to separate his two predictions: he may be wrong that ChatGPT is the platform, but he is very confident the four-step cycle will repeat.

Balfour's first example is Facebook, where one of his early companies rode the social gaming boom and lived the whole cycle in a short period. In 2007 Facebook was smaller than MySpace and Friendster but saw the direct network effects early. It offered developers a canvas, freedom to monetize and access to notifications and feed, taking only the sidebar ad space. Later it took a cut, suppressed organic channels and finally absorbed top use cases such as photos and events into first-party features.

He stresses that none of this is evil. The cycle comes from competitive and capitalistic pressure: platforms must keep growing, must avoid enabling their own disruption, and sometimes closing really is better for users when feeds get spammy. His advice is simply "you just need to know how to play the game."

Google ran the same cycle over a much longer period. It identified data moats, got web developers to optimize for its search, then turned more real estate into ads and absorbed high-value verticals like travel. Mobile did it too, with the App Store and growing restrictions, and LinkedIn did it twice, first with company pages and then with personal profiles and thought-leader ads. Balfour notes that the cycles seem to be getting shorter, which leaves less time to play.

Balfour says the natural reaction on first hearing this is "screw them, I'm not playing that game," and he understands why. Many companies were killed when Facebook's platform collapsed, Apple's 30% cut wrecked some business models, and plenty of companies built on SEO loops are in serious trouble now. Rachitsky adds that he thinks of Zynga, which grew on Facebook and became a massive company.

His counterargument uses ChatGPT's new deep research connectors, one of which is for his former employer HubSpot. In isolation, making your data accessible inside ChatGPT looks irrational. But competitors will join, customer expectations will change, and so it becomes a prisoner's dilemma with no way to opt out. Being early is better than being late, and the hard part is anticipating the final stage and sequencing away before it arrives.

Balfour argues that we are in Step Zero now, with ChatGPT, Claude and Gemini in a fierce battle and no clear winner. He thinks the moat is context and memory: side by side, models produce similar results, so the difference is which one holds more of your context, and memory creates a flywheel as usage feeds personalization. ChatGPT, he says, was first to memory and has invested heavily in context connectors.

To the objection that Google has Chrome and enormous distribution, he answers that historically winners were not those with the biggest distribution at a given moment but those with the best retention and engagement. He cites retention data published by Menlo Ventures' Deedy Das showing ChatGPT's curves level off higher than rivals and are shifting up over time, with the rare "smile curve" he associates with winners like Slack.

The third reason is signals of a coming third-party platform. OpenAI has been hiring for what look like agent-platform roles, and Balfour expects a value exchange where developers get context, memory and distribution in return for integrating. It is already forming preferred partnerships with larger companies, which lends the platform credibility. He adds that ChatGPT has roughly 10x the monthly active users of Claude, which makes it the logical place for a developer to spend scarce resources.

Asked for a fallback, Balfour says Apple is best positioned on paper because its devices see everything and it owns the ultimate view of your context, though there are no external signals it is executing on that. Behind Apple he puts Google, which owns email, search, Chrome and Android. But he suspects many Gemini users are accidental fly-by clicks, and says the window for Google is small if ChatGPT keeps pressing its advantage.

Rachitsky relays that Anthropic's Mike Krieger told him ChatGPT had caught lightning in a bottle, so Anthropic is focusing on what Claude is best at: developer tools, coding and backend work. Balfour welcomes this and says smaller platforms exist and will keep emerging, all following the same cycle. His example is Udemy, whose creator revenue share started around 80% and was later cut sharply.

He expects Cursor to build an agent platform for developers, and tools like Notion, Airtable or Monday.com to try the same. Rachitsky adds that niche channels like LinkedIn can still be very valuable for a specific audience. Balfour is clear his prediction has no bearing on which product he likes most, since he uses both Claude and ChatGPT.

Rachitsky offers a tangent from Bret Taylor of Sierra: agents are attractive partly because their impact can be attributed, so outcome-based pricing is possible, such as a dollar per resolved case. Balfour asks whether that lasts, since competition will eventually erode the price unless something else creates durable pricing power. Rachitsky suggests proprietary data could be that moat, and Balfour agrees if it is paired with a second advantage.

On timing, Balfour warns that AI timelines are always shorter than expected, but expects the next major steps within six months. ChatGPT's recent Agent mode is a general-purpose agent that introduces users to agents, but he thinks no horizontal tool can serve all use cases because people struggle with generic tools and specific cases need specific UI and data.

He expects first an announcement of an initial 10 to 20 preferred partners acting as a credibility card, then a broader opening where OpenAI defines the value exchange. A second vector is the replacement for search, with deeper attribution, native shopping and new monetization mechanisms that cover the free tier without relying only on subscriptions. He admits he does not know what acronym, AEO or GEO, the industry will settle on for this.

Balfour says we have just entered the casino and now have to decide where to place chips. History shows what matters: in the iOS and Android era, betting only on Android probably lost, while iOS alone could still win, so iOS had to be part of any winning strategy. A late-stage company can afford to spread bets and then back the winner, though the risk is that incumbents wait too long.

Startups do not have that luxury. With scarce resources and attention, they have to choose one platform and go all in, accepting higher risk and higher reward. Rachitsky summarizes the practical advice as integrating with whatever ChatGPT or others launch, whether login, search or memory connectors. Balfour boils it down to two rules: play the game, and make it a focused bet, since the failures are those who played multiple games at once.

Rachitsky asks whether HubSpot should really give its data to ChatGPT, and Balfour says yes, though he has not spoken to anyone there. Even without a clear exit plan, it is better to be early and figure out the exit along the way than to join late. He also points out that Rachitsky himself hitched to Substack early with a focused bet and benefited disproportionately, even though it felt too late when he started six years ago.

His criteria for choosing a platform come in four parts. First, retention and depth of engagement matter more than vanity user counts. Second, user quality and ability to monetize, as with Android's large device share but much smaller share of dollars versus iOS. Third, the value exchange, meaning what the platform offers in return for building on it. Fourth, pure scale, since a 200x gap probably means you choose the bigger platform.

Once you enter, you should immediately plan your exit. Balfour suggests owning an important part of the user workflow, accumulating specialized data and context that the major platforms lack, or building micro network effects. Rachitsky notes that this is the same logic as defensibility for a so-called GPT wrapper company.

Asked what founders can do right now, Balfour says it may be a bit too early, though startups should ask how they would evaluate each platform if it launched. Some early-stage companies will be able to cultivate preferred relationships with OpenAI, and he thinks those who can should play that card. Otherwise, the main thing is to understand the value exchange once the platform launches and be ready to redirect strategy quickly.

That speed is hard for leaders who fear giving teams whiplash. Rachitsky adds an example: ChatGPT now sends his newsletter more traffic than Twitter, and he could block it, but he would rather it recommend his work than let someone else take that share. Balfour agrees that if you don't do it, somebody else will, which is the same dilemma media publishers face.

Rachitsky steers to a different topic: Reforge now builds software, starting with Reforge Insights, an AI product researcher that aggregates feedback and identifies gaps. Balfour says that through selling tools and years of transformation projects, they see a wide gap between CEO "AI manifesto" memos and what actually backs them up. In his view the most impactful step is to impose hard constraints.

He gives examples. One client benchmarked against peers and decided each function would be one-fifth the size, which blocked hiring and forced AI adoption. Another approach, possibly Shopify, is that you get no new headcount until you prove AI cannot do the work. Some executives refuse to review a PRD unless it comes with three prototypes.

The biggest separator is making the hardest decision, which is exiting people. Balfour describes catalysts who lead the charge, converts who need structure and permission, and anchors who silently drag their feet. Companies that set a deadline to make the change or exit are, he argues, treating this as a deep culture change, since "cultures thrive on density." He says fewer than 10% of companies take this stance, but they are the furthest along.

Balfour says most executives are disconnected from how AI adoption is going on the ground. When Reforge asks users of prototyping tools how many others on the product and design team use them, the answer is almost 90% of the time just "me and one other person." He tells of a principal PM at a well-known tech company whose prototype stalled for a month until the CEO happened to hear about it at a happy hour and fixed it the next day.

His advice is to get to the ground floor and measure actual usage, as Shopify does. He also quotes Fareed Mosavat: output is constrained by the slowest part of the system. In AI adoption that is often IT, legal or procurement, and in product teams it is the parts that have not sped up. Accelerating only engineering just moves the bottleneck to PMs or design, so leaders should attack the slowest part ruthlessly.