Distribution Beats the Better Model

On January 12, 2026, Apple and Google issued a joint statement: the next generation of Apple's own foundation models would be based on Google's Gemini. A few months earlier Microsoft had started letting customers run Anthropic's models inside parts of Microsoft 365 Copilot, a product built on OpenAI's.

Two giants with every reason to want their own AI treated the model as a part they could buy.

That tells you where competition in AI products has moved. When everyone can rent roughly the same model, the better model buys a lead that lasts about as long as the incumbent's next release. What usually decides who wins is access to the user at the moment the work happens, and incumbents already have it. Startups still win, and some have won spectacularly. But they win under specific conditions, and you can check for those conditions before you bet on them.

Why the model stopped deciding

Startups rent from the same short list of labs as the incumbents do. When a lab ships a better model, it reaches the incumbent's product and the startup's product within weeks of each other. Real differences remain in the harness, the ordinary software around the model that decides what it sees and what it may do, and those differences can be large (Same Model, Different Results). But a harness is code, and code has become cheap to copy (Where the Moat Moves).

What doesn't get cheaper is getting in front of the person doing the work. In an AI product, "distribution" means exactly that, and it has concrete parts:

• An installed base: people already signed in every day for other reasons.
• Default placement: the feature appears without anyone choosing it, the way an AI summary appears above search results.
• Contracts and trust: the AI lands on an invoice that already exists, from a vendor that has already passed security review and will still be around next year.
• Bundling: a price of zero at the margin, or a modest increase on a plan the customer already pays for.
• Data and permissions already granted: the incumbent's assistant can read the email, files and customer records the incumbent already hosts.

Each part removes a step a startup has to pay for: being discovered, tried, bought, approved by security, integrated, and turned into a habit. Add those up and a startup can easily spend hundreds of dollars to win one seat; one worked example puts it at $480 (Capability, Product and Business Model). The incumbent's cost of putting a button in front of a user it already has is close to nothing.

When everyone rents the model, the scarce input is the user's attention at the moment the work happens.

Two clocks

Alex Rampell of Andreessen Horowitz summed up every startup-versus-incumbent fight in 2015: the question is whether the startup "gets distribution before the incumbent gets innovation."

That gives you two clocks. The startup's clock measures how long it takes to collect enough users, contracts and habits that leaving would cost something (switching costs). The incumbent's clock measures how long until its built-in AI is good enough that the difference stops justifying a separate purchase.

AI made the incumbent's clock much faster. Clayton Christensen's research found that incumbents rarely lose to improvements of their existing product for their existing customers, which he called sustaining innovation. They lose to products that start where they aren't looking. An AI assistant inside an existing product is the textbook sustaining improvement, and renting a model makes it quick to ship. An incumbent that can't build can buy: Thomson Reuters agreed in 2023 to acquire the legal AI company Casetext for a reported $650 million, and its assistant was reported in early 2026 to have reached a million users.

The incumbent's clock now runs in quarters, not years.

A model of the race

To see how the two clocks interact, it helps to build a deliberately simple simulation. Picture a group of users who will each adopt one main AI tool for one job. Each quarter:

• The startup, which launches first with a better product, wins users two ways. Paid marketing reaches 1 percent of those who haven't adopted anything yet. Word of mouth adds more, in proportion to how many users it already has and to the quality gap: the share of uses on which its product is visibly better. A gap of 0.5 means the user notices the difference on one use in two.
• The incumbent, once it ships its good-enough version, reaches 12 percent of the remaining users each quarter through default placement.
• Incumbent users switch to the startup if the gap is big enough to bother, faster as the startup grows.
• The gap closes about 8 percent a quarter once the incumbent ships, because both ride the same model releases.

This is the standard way economists describe how products spread, with outside promotion plus imitation, with one twist: imitation only works if people can see the difference. The numbers are illustrative, not forecasts for any company. But the shape of the result is robust.

Scroll sideways to see the whole figure

Look at the medium case. The incumbent's first quarter after shipping reaches 11.2 percent of users. The startup's first five quarters of marketing and word of mouth together reached 10.4 percent. One quarter of default placement beat more than a year of a better product.

And paid marketing alone never catches up. The startup's lead, when it has one, comes almost entirely from word of mouth fed by a gap people can see. That's why the small-gap startup ends with 14 percent against the incumbent's 81, even though it started first.

About half. With a two-year head start, about a third is enough. With no head start, the startup needs to be visibly better on more than seven uses in ten.

Scroll sideways to see the whole figure

The boundary runs diagonally, and that's the main result: a head start and a visible gap substitute for each other, partly. A small gap needs years of incumbent delay. A large gap survives an incumbent that ships fast. With neither, the startup loses.

Two other dials matter as much. Halve the startup's cost of winning a customer and a losing case turns into a lead. And the incumbent's reach matters as much as its timing: if the incumbent reaches 2 percent of users a quarter instead of 12, even a startup with a tiny gap ends close to even. That second dial explains a lot about the real cases.

When distribution won

Google is default placement at its most powerful. It began rolling out AI Overviews to everyone in US search in May 2024, and by July 2025 reported more than two billion monthly users. Almost nobody chose them; anyone who searched and saw a summary counts. That makes the number the strongest form of distribution and the weakest evidence that people prefer the product. Both things matter.

Microsoft 365 Copilot shows what happens when the incumbent's reach is low. It launched to enterprises in November 2023 at $30 per user per month. By January 2026 Microsoft reported 15 million paid seats against a base of more than 450 million commercial seats, about 3 percent. By July it passed 30 million. Thirty million seats is more than almost any startup will sell. But a paid add-on travels like an incumbent with low reach, because someone still has to approve $30 per user. That leaves room for competitors with a modest gap, and it's exactly why Microsoft also gives a free version to its commercial users.

Salesforce's agents work on records customers already store there, under permissions already set, which a startup would have to earn first. Its agent business passed half a billion dollars of annual recurring revenue by December 2025.

When the startup won anyway

ChatGPT launched in November 2022 while Google owned the search box, and by October 2025 had more than 800 million weekly users. Google's AI answers reached everyone in US search about a year and a half later. Three things favored the startup.

The work was new. Long conversations, drafting, coding help and analysis were jobs the search results page didn't do, so Google's default placement didn't reach them. The incumbent's business pulled against it: an answer that replaces a page of links competes with the ads on that page, a conflict the strategist Hamilton Helmer calls counter-positioning. That's a hypothesis about Google's incentives, and so far its search revenue hasn't shown the damage. And Google's users expected search to be right, so a wrong answer cost Google more than it cost a new product people were trying out of curiosity.

Then the incumbent's clock ran. The Gemini app passed 750 million monthly users by February 2026. Both companies ended up large, which is the medium-gap outcome in the model: the startup wins its category, and the incumbent keeps a big share.

Cursor is the cleaner case. GitHub Copilot launched in 2022 from a Microsoft company, running inside Microsoft's own code editor. Cursor, an editor built from a copy of that same Microsoft editor, passed $1 billion in annualized revenue in November 2025. Three conditions lined up.

A developer sees the quality of a coding tool many times an hour, on every suggestion and every edit, so a gap is visible constantly. Agentic editing across many files was a new workflow, and Cursor controlled its whole editor while Copilot shipped as an extension that had to work across several. And developers paid with a card, so Cursor reached users before procurement did, which is where incumbents are strongest.

The incumbent grew too: 4.7 million paid GitHub Copilot subscribers by January 2026, up 75 percent in a year. And the story kept moving. By 2026 Cursor's main rival was reported to be Anthropic's Claude Code, not Microsoft, and in August 2026 SpaceX, by then merged with the Grok developer xAI, completed its acquisition of Cursor. Winning distribution made Cursor something a company with a model wanted to own (Why Labs Build Products).

The three conditions

Put the cases next to the model and the conditions fall out. A startup wins when at least one of these sits outside the incumbent's reach:

The work. The job moves to a surface the incumbent doesn't own: a new kind of conversation, a new agent workflow. Default placement only works where the incumbent already sits.

The gap. The difference is large, visible on each use, and uses are frequent. A gap that shows once a quarter, in a report nobody reads closely, never generates word of mouth. A gap that shows forty times a day does. It helps if the incumbent's business or its users slow its response.

The purchase. Adoption bypasses procurement: individual sign-ups, free tiers, credit-card plans. Or the price is one the incumbent can't match without hurting itself, such as usage pricing against a vendor that sells per seat.

A startup wins where the work, the gap or the purchase sits outside the incumbent's reach. Where all three sit inside it, the better model usually loses.

The reverse describes where incumbents win, and it covers most enterprise software: the job happens inside the incumbent's product, the gap is modest or occasional, the purchase goes through procurement, and the data is already there.

"But bundled seats don't get used"

The strongest objection is that incumbents push AI into every license, report big seat counts, and people ignore tools they didn't choose.

The first half is true. When the UK Department for Business and Trade gave 1,000 Copilot licenses to staff in late 2024, about 64 percent used it at least weekly and about 30 percent at least once per working day. The evaluation didn't find evidence that the time saved turned into higher productivity. Seat counts overstate adoption.

The objection is still weaker than it looks. Light use occupies the budget and the default: a finance team already paying for an assistant inside its office suite needs a strong reason to buy a second one, which raises the gap a startup has to show. And the bundled product improves on the same schedule as the startup's, because both rent from the same labs. Low use today doesn't mean low use next year.

Where the objection bites is renewal. Bundled seats get decided again every year, and a seat used once a day is the most exposed line in the contract. Nobody publishes renewal rates for AI add-ons yet. When someone does, we'll know how sticky this distribution really is.

What I'd conclude

My view is that the worst place to build an AI startup is also a common one: a modest gap, inside a workflow an incumbent already owns, sold through procurement. All three conditions sit inside the incumbent's reach there, and a better model doesn't get you out of it, because the incumbent rents the next model too.

The interesting places are where the work itself is moving. The labs are betting heavily that work will increasingly start in an assistant that reaches into other applications, instead of inside the application that holds the data. If that happens, default placement inside the old applications loses value, and distribution moves to whoever owns the assistant. That's the shift I'd watch, more than any benchmark (What a Startup Can Still Win).

Humility is warranted on one point: if one lab opened a lasting quality gap and kept it for its own products, the model would decide again. As of October 2026 the leading models are close together.

What to do

If you're building or investing, ask five questions before anything about the model:

1. Where does the work happen, and who owns that surface?
2. How many times a day does the user see the difference?
3. How long before the incumbent can rent something good enough? Assume quarters, not years.
4. Can the product reach users before procurement does?
5. What can it offer that the incumbent can't copy without damaging its own revenue?

If you're buying, the same logic runs the other way. The assistant bundled into the suite you already pay for will usually be good enough, and it will improve with every model release. Pay for a separate tool where your people see the difference many times a day, and nowhere else.

Check yourself

Why does distribution beat the better model when everyone rents the same models?

A better model reaches the incumbent's product within weeks, but access to the user's attention, contracts and data does not get cheaper to copy.

Why did Cursor win against GitHub Copilot's head start?

A developer sees the quality gap many times an hour, agentic editing was a new workflow Cursor controlled, and developers paid by card before procurement got involved.

What does the race model say about a head start and a visible gap?

They partly substitute for each other. A small gap needs years of incumbent delay, while a large gap survives a fast incumbent.