a16z：OpenAI 赢在创造新客户与分发，而非模型本身

OpenAI understands something special and rare, hiding in plain sight. Which is, as Peter Drucker put it a half century ago, the purpose of a business is to create a customer.

OpenAI is not going to win because they have the best models, or the most advanced chips, or the best cost-performance curve, or even the best product, although we think they do have those things. We think that OpenAI will win because they are so good at creating new kinds of customers, and because they have the most durable distribution strategy.

As we write this in September 2026, the intelligence-as-a-service product rankings aren’t the most important question anymore. Intelligence is everywhere. Anyone can do nearly anything; but most people aren’t, yet. Can you awaken the new behavior, and can you win at distribution, are what matters.

The case for OpenAI comes down to the following argument. To build a durable business at AI frontier scale, there are four levers you can pull:

Create new kinds of behavior
Distribute to lots of users
Price your product in a way that captures value
Have high switching costs

4 is basically out. The AI Frontier is a Red Queen’s Race, at least at the model layer. It’s so easy to swap between models, down to task-by-task switching, and everyone’s getting better all the time.

3 is what everyone hopes to achieve. But again, it’s a very competitive world out there & OpenAI (and others) are determined to lead at the cost-compute frontier, and pursue it via economics of scale, vertical integration, and other strategies to get token costs down. Whoever ends up leading on price-to-performance is going to be whoever’s winning at scale anyway.

Which means it comes down to 1 and 2. We think OpenAI is clearly the best at 1:

They have a well-established track record in AI of finding the simple, obvious-in-hindsight breakthroughs that unlock new kinds of consumption behavior

And they’re also the best at 2:

Every AI lab right now is trying to speedrun distribution, beyond their first-party products. And there are two ways you can do that: as a platform, or through partnerships. Doing it as a true platform is long-term preferable: the revenue comes a bit slower, but it’s much more durable and valuable when you get it.

There is an underlying reason why OpenAI is prevailing at 1 and 2, which is that they have some of the broadest and most well-rounded general intelligence, not only in their models but in their company and in their customer base. They are exposed to the broadest user base (across consumers, prosumers and enterprise), they have the deepest self-improving tech stack (down to their proprietary chips), and that’s why they consistently find the new patterns and primitives that work. They are taking the long view for how to build the winning AI platform, and it’s working.

OpenAI is really good at discovering the next thing

OpenAI did not set out to be a consumer company. They kind of fell into it by chance, with the release of ChatGPT. It’s a happy accident they did, because that moment is one of the turning points in technological history. Had they not done this, the world would look very different. AI would probably still be guarded in the hands of labs and close customers, it wouldn’t be improving at nearly the rate that it is, and we’d be in a “default closed” timeline, and not the infinitely richer one we have today.

Sam Altman recently told a story about the weeks after the release of ChatGPT, when most people at OpenAI had the attitude, “Well this was fun, but we should probably move on to building some real products. A chatbot can’t be it, can it?” It was specifically Peter Thiel who counseled Sam, no, this is it. You have created a new kind of customer. Nothing else is as important as this.

It’s not just ChatGPT, either: OpenAI deserves credit for being generally good at finding the breakthrough patterns and primitives that really matter, and awakening new kinds of user behavior. Ben Hylak articulated it well in a piece the other day: “OpenAI is the (mostly) undefeated king of finding the ‘next thing’. It’s always a lot simpler than you’d think and very obvious in hindsight.” The GPT chat interface, reasoning, tool calling, and Computer Use were all big breakthroughs that simplified and focused the direction of the technology, and made it usable. The one major exception where OpenAI was not the breakthrough leader was coding, and they caught up just fine.

Why are they so good at this? Perhaps in the early years they just had that special sauce; no one can exactly articulate the founding energy and magic that creates something like this. But over time, this has clearly become a skill they have cultivated. There are three critical inputs here:

They think and act like a company that wants to be a platform. “Platform” is a loaded term and we’ll get into it in detail in a second, but this is a company that actually believes its users are smart, and can make their own choices, within carefully crafted constraints.
They’re doubling down on technical depth, meaning they’re trying to own or control every layer deep into their stack, and continually refine the platform primitives their users need as building blocks.
Maybe most of all, they have tremendous consumer breadth. They remain the leader in sheer variety of kinds of consumers and use cases across people’s personal and work lives.

The breadth of people using them is the most underrated factor here. As Hylak puts it simply: “Computer use” [for example] is a very generic, and unspecific thing. To be better at it, you need to increase your intelligence in a general way. Across many domains, across many workflows.” People consistently underestimate how difficult it is to go from a specific solution to a general solution. But you usually only discover what’s simple and obvious (in hindsight), but generic and unspecified (as problems to tackle), by going after that general solution.

Why don’t more companies go after “general solutions”? Because it’s actually quite costly to do so, and goes against most of the best practices of running a focused product organization. You need to gather inputs from a huge range of customers and behaviors and actions, letting them do essentially anything they want to do, to make general progress on a domain, as compared to specific progress on a product.

That means that the ability to build “general knowledge” into your product is actually dependent on your distribution strategy. This is non-obvious but important, and OpenAI understands it.

How do you serve “exactly what I want” (which can be a very custom, specific, finicky thing) at massive scale, to everybody?

There are basically three ways. You can deliver it in a standalone product, you can deliver it through partnerships, or you can provide it as a platform.

Standalone Products have an advantageous constraint, which is that your product must, inherently, let the user “do whatever they want” (i.e. ask ChatGPT anything), so you are genuinely creating the customer experience, from getting it into their hands to watching them use it daily.
Distributing through partnerships has the advantage that partners give you a lot of reach, quickly, around a lot of use cases. But it is ultimately the partner that owns and directs the customer’s “anything I want” instinct. And therefore, over the long run, not only will they likely capture the attractive economics, they are also in control over what the user ultimately gets to do, and what gets learned. Furthermore, your customers may not be aware you are even powering it, so you don’t earn any trust or goodwill from them.
Distributing as a platform means people build whatever they want on top of you. This is slow, because you have to make all the new products supporting “anything I want” from scratch. But it’s very powerful once established, because even though revenue may come slower or at lower take rates, you have established a really powerful learning loop about the world, as users employ your primitives and patterns, that teach you about the world and about what people want from you. Not in the “training on customer data” sense, but in the, “the primitives become a better map of the territory” sense.

Over a long enough time, products come and go, partnership dynamics change, and their economics may disappoint relative to expectations. But OpenAI keeps coming back again and again with their drive to build the enduring platform. And what they’re doing is working.

True platforms are rare but worth it

Platforms are hard to build, and they’re hard to reason about. They’re even hard to define!

Generally we have a broad, fuzzy concept of platform that goes, “platforms let people build the thing for their specific need.”

In software, there are a few ways you can do that. Nineteen years ago, Marc wrote an essay called “Three kinds of platforms you meet on the internet” that remains useful today. Marc’s definition is: “A ‘platform’ is a system that can be programmed and therefore customized by outside developers—users—and in that way, adapted to countless needs and niches that the platform’s original developers could not have possibly contemplated, much less had time to accommodate.”

The “three kinds” correspond to three different ways you can arrange computers so that a user gets a custom outcome that they want. They are progressively harder to build (as you ascend from 1 to 3), but more magical in terms of what kind of behavior and building they unlock.

Type 1 is what we’d call “Headless” today - give users access to something useful through an API (like Flickr back then, or even some companies as valuable as Stripe would fall into this.)
Type 2 uses what you’d call “Plugins”: like the old Facebook platform, or most Shopify or HubSpot Apps today. You have a core service that provides the first 80% of users’ needs, and then custom apps that run on their own server, and probably managed through some kind of app store, that give users the custom experiences they want.
Type 3 is the critical one: it’s a genuine runtime environment that people can program to create whatever they want. These are really, really hard to build, but they are true magic when they work. iOS, AWS, Ethereum, and recently Cloudflare are some examples.

What’s interesting about AI today is that all model companies and inference providers, in one sense, have a claim to be contributing “Type 3 platform” work. (Models are a kind of runtime, prompted with custom instructions!) But models alone do not unlock new behavior, nor are they all that defensible. Like we saw with coding, where the harness matters just as much, it’s the superstructure of constraints and primitives that actually matter here. We’ll give you a theoretical reason and a practical reason why.

The theoretical reason is that, if you want to discover the obvious-in-hindsight patterns and primitives that matter, you need exposure to as wide as possible range of use and tacit knowledge and trial-and-error from your users, so that you can evolve from a specific product (a local breakthrough) to a general product (a global breakthrough). And the purest way to do that is to let them write arbitrary instructions (i.e. code) for what they want to do. Getting the security right and the access and data structures and constraints right is a massive undertaking that “product-oriented” companies have a hard time prioritizing; only the truly committed will reach the promised land.

The practical reason, for AI right now, is that for agents to be useful, they need to write code. This is something that is perhaps non-obvious, but that we’ve learned in the past year. Given how amazing AI models are at writing code, they’re surprisingly bad at calling tools and carrying out instructions in everyday knowledge work - unless you flip the script around and say, “write a program that does these tasks”, and then they’re great at it. An open question in enterprise AI adoption right now is “how are people and organizations going to support this, safely?” and you should probably bet on the platform-minded people to get it right.

One of the big open questions right now is what kind of network effects, switching costs and other moats will emerge at the various layers of the AI stack. As of now, the model layer remains shockingly substitutable. There’s a general sense that “multiplayer mode” (e.g. collaboration on projects, agents being trusted by friends’ agents) has some good ingredients for business model defensibility, but it hasn’t yet really impacted the overall brawl around “usage up, costs down” that we see with token spend. The primary question that matters is not “is your model sticky”, it’s “are you creating new behavior?” As Steve Hou put it well: “The elasticity that matters is not the elasticity of substitution between models but the elasticity of aggregate demand for AI.”

Pattern-matching is never perfect, but if the past four years of AI is any indication, the next big breakthroughs in defensible, multiplayer usability of AI by regular people are going to emerge from epiphany moments where we find optimal layering of intelligent models, computing primitives, and behavior patterns, and something simple and obvious emerges. It could be from startups building on the model companies as platforms; it could be from the model companies themselves. Either way, OpenAI obviously wants to power it, and deliver it to users.

In a world of abundance, the causality goes backwards

Old habits die hard. In the regular ways that you might analyze an AI Frontier lab as a business, you might look at all kinds of obvious things to assess competitive advantages: who has the best models? Who has the best chips? Who has the best cost-performance? Who has the best products?

Obviously these all still matter. But in a world of abundant intelligence, on-demand anywhere, we’re probably getting the causality backwards if we think that having all of those things means you win. We think that, in the future, OpenAI will have the best models, chips and products because they have the best distribution, not the other way around. They’ve made great decisions thus far around securing enough compute, training models really effectively, developing Jalapeno as an optimized chip, and other practical things that can make a competitive difference within the set of AI options. But in the future, we think they will keep winning because they will stay at the frontier of gathering information about the world, from their own newly-awakened users, and self-improving from there.

New product and model offerings can grab everyone’s attention for a week or two, but over the long run we don’t really see OpenAI relinquishing their leadership position across what matters:

They have retained the dominant consumer brand and a wide user breadth
They have awakened new kinds of user behavior, that are obvious in hindsight and change the game of what’s possible
They have pushed the frontier of “best models, best price” in their model offerings
Their “flywheel” never lets up, and everything improves together

Sam describes the future product state of OpenAI as having only two offerings: either you can make anything you want, or else you can just talk to ChatGPT. This is what real abundance looks like in a product. You can either describe exactly what you want that will help you and get it, or else you can fall back to, “or I can just talk to this text box and figure out wherever I’m going.”

That abundance is generally available; today. There is a pretty small group of people that are using AI rampantly for everything in their lives; there is a somewhat larger and broader group of people that are using ChatGPT on a regular basis. They haven’t been awakened as customers yet. Who will do it?

We think OpenAI understands how.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.