Why Is OpenAI Still the Likely Winner in an AI-Saturated World?

If every company's models keep improving, today's leader could be caught tomorrow. So why might OpenAI keep winning?

David George of a16z points not to models, chips, or the cheapest tokens, but to two harder-to-copy abilities: getting people to adopt an entirely new way of using AI, then bringing that behavior to a vast audience.

ChatGPT is the clearest example. It did not merely improve the existing chatbot. For the first time, ordinary people got used to solving problems by describing what they wanted in an empty text box, without first finding the right software, learning its menus, or inventing search terms.

The argument turns the usual account of AI competition around. We normally think a better model brings more users. George's bet is that the opposite may increasingly be true: because OpenAI reaches more users facing a wider range of problems, it is better placed to learn what the next generation of models and products should do. More users help it build better tools; better tools bring in more users.

Four ways to turn a lead into a lasting advantage

1. Create a way of using the product that did not exist before.
2. Distribute it to a large number of users.
3. Find pricing that captures the value it creates.
4. Build switching costs that make customers reluctant to leave.

George discounts the fourth lever first—but specifically at the model API layer. A developer can use Claude for coding, GPT for analysis, and a cheaper model for batch processing, or choose a provider task by task. Integration and evaluation still take work, but changing models usually does not mean rebuilding the entire product. As every provider improves, a single lead is unlikely to lock customers in indefinitely. Whether products, working relationships, and platforms can create switching costs is a question the original essay leaves open.

The essay cites shifts in OpenRouter market share suggesting that OpenAI and Anthropic's shares of calls can swing sharply in a short period. That alone does not prove users have no loyalty. It does suggest something narrower and important: call volume is not the same as a durable customer relationship. Traffic may move when price or performance changes.

Now consider the third lever, lower prices. Price matters, but competitors are all pursuing scale, vertical integration, and custom chips to reduce token costs. George thinks the company that ultimately offers the best price-performance ratio may be the one that already has the most scale. Cheapness, on that reading, is more the reward for winning scale than the original reason for winning the market.

So switching costs are weak at the model layer, and price, while important, may follow scale. George places the source of a lasting gap in the first two levers:

1. Create new behavior: help people use AI for something they could not, or would not have thought to, do before.
2. Build distribution: put that behavior in more hands and keep seeing what people want to do next.

The rest of the case explains why those two abilities reinforce each other.

“Creating customers” means creating a new habit

When ChatGPT launched, OpenAI did not assume it had found a lasting consumer product. Sam Altman later recalled that, despite fast growth, many inside the company wondered whether people were only enjoying a novelty: surely a chatbot could not be the real product. The team even considered looking for another direction.

Peter Thiel saw a newly created kind of customer rather than a toy. He told Altman that this was the product. Like Google's search box, ChatGPT's blank prompt appeared to offer almost nothing—but anyone could arrive with a problem of their own.

“Creating customers” here does not mean taking existing customers from a rival. It means giving someone who previously had no reason to buy, and perhaps no idea they needed this, a new habit.

Before ChatGPT, a person wanting to analyze a file, change some code, or draft a contract would often need to know which software to use and how to operate it. ChatGPT changed the starting point: state the goal first, then ask AI to work out a way forward. What emerged was not just a new account, but the habit of asking AI first.

Several later breakthroughs have a similar quality: the chat interface, reasoning, tool calling, and computer use. Each looks obvious afterward; the hard part was recognizing in advance which simple form people would return to again and again.

Chat lets people express a goal directly. Reasoning lets a model think before responding. Tool calling turns answers into actions. Computer use lets AI work with software that has no dedicated interface for it. These are not merely incremental benchmark gains; each opens another mode of use. Coding is a notable exception to the pattern: OpenAI was not first to discover its breakthrough form, though it later caught up.

George traces OpenAI's apparent knack for finding “what comes next” to three conditions:

1. It does not decide everything on behalf of users. Instead of choosing only a fixed feature, users can specify what they want to accomplish within safety boundaries.
2. It controls enough of the technical stack. From models and tools to code execution and chips, deeper control gives it more ways to fill a missing capability when a new need appears.
3. It sees a wide variety of users. Consumers, specialists, and enterprises bring strikingly different problems every day, exposing recurring friction sooner.

That last point matters especially. Expense software has to understand expenses; customer-support software has to understand support. General-purpose AI faces countless goals no one could list in advance. Without a broad user base, a company may not even see where people repeatedly get stuck, much less know what to build next.

Distribution is not just selling more; it is seeing more

We usually associate distribution with ads, channels, and preinstallation: get the same product in front of more people. In this argument it has a second role: showing a company more real problems, so it can decide what to build next.

OpenAI encounters consumers, heavy professional users, and enterprises. Some ask it to write, others analyze data, repair code, or classify support tickets. Each request may look idiosyncratic, but at sufficient scale common gaps emerge: many tasks need persistent memory, computer operation, safe code execution, or collaboration among people and agents.

A company can turn such recurring needs into reusable foundations. Engineers call them primitives; think of them as Lego pieces. They are not a complete support or finance application, but building blocks such as remembering context, operating a computer, and executing code safely. Developers can combine them into products OpenAI itself never imagined.

A possible flywheel follows: more users reveal more problems; recurring problems make useful general-purpose building blocks clearer; richer building blocks let outside developers create new applications; those applications bring more users.

“Learning” here does not mean OpenAI is free to train on customers' private data. It means broad usage can reveal the shape of demand: what people keep trying to do, where they get stuck, and which capabilities deserve to become part of the product.

Three routes to users, but only one keeps widening

An AI company can enter more contexts in three ways: build its own product, appear inside partners' products, or become a platform. All three can increase calls. The difference is who ultimately faces the user and hears what that user is really trying to accomplish.

Partnerships offer the fastest reach. Put a model inside office software, a phone, or an enterprise system, and calls can soar. But users may remember the partner's brand; the partner often controls the interface, pricing, and customer relationship. The model provider gets usage without necessarily learning why a request was made, making that demand harder to turn into its own product capability.

A platform is slower. It must supply not just models but tools, permissions, safety, data, and a runtime, and it may not earn the most money at first. Once others can build freely on it, however, thousands of developers help explore demand. The platform need not predict every vertical application itself to see where its ecosystem is headed.

That is why George favors the platform route. Partnerships borrow someone else's entrance to gain traffic. A platform invites others into its own environment to create something new. The former expands reach; the latter may expand both reach and the ability to discover new needs.

An open API is not yet a real platform

“Platform” is easy to overuse. An API that lets others call a model certainly exposes a capability, but it is still far from an iOS- or AWS-like platform.

Drawing on Marc Andreessen, George distinguishes three levels of technical platform. These are not the three business routes just discussed. They answer a different question: how deeply can outside developers build within your system?

• Type 1: Provide an API. The platform exposes a capability; developers supply their own code, databases, servers, networks, and security. This asks the least of the provider and the most of developers.
• Type 2: Allow plugins. A core product handles common needs; outsiders add specialized functions, as with extensions to existing software.
• Type 3: Provide a runtime. Developers can build and run an actual system inside the platform's compute, storage, permission, and safety boundaries. iOS, AWS, Ethereum, and Cloudflare illustrate this level.

What makes Type 3 different is not simply a longer feature list. The platform takes responsibility for the most difficult common layer. A mobile app developer does not need to build an operating system; a cloud developer does not need to maintain a data center. Likewise, an AI agent will struggle to spread if every developer must independently work out where its code runs, which data it can read, how failures roll back, and how overreach is stopped.

A model handles only part of this. A user writes a prompt and a model follows it, which can look like a kind of runtime. But models are relatively interchangeable. The harder thing to move may be the surrounding system of tools, data, code execution, permissions, and safety rules. George thinks that layer could become OpenAI's more durable advantage.

Why agents need to write code on the fly, not just click tools

Imagine asking an agent to find anomalous orders in dozens of spreadsheets, sort them under different conditions, and draft an email for each customer. No platform team can create a dedicated button for every such one-off task.

The agent could repeatedly call tools to read files, filter data, and write emails. But as the chain grows, it may miss a condition, duplicate an action, or lose track of an earlier step. Often a more reliable approach is to write a small program: express the steps, conditions, and loops together, run it, inspect the result, and revise only what went wrong.

This is why code matters to a general-purpose agent. It expresses goals the platform developer did not foresee and makes a complex process repeatable, inspectable, and editable. Code is not merely a feature for programmers; it is a way for agents to organize complex work.

But teaching a model to write code is just the start. The platform must also answer hard questions. Which files may the code read? Which tools may it call? Who may receive the result? How does it recover from failure? Must it stop and get human approval before a payment or deletion?

That is the heavy work of Type 3—and a possible source of defensibility. A useful agent platform is not just an intelligent model; it is an operating environment in which the model can act safely.

The biggest market may be awakened, not captured

Suppose model A cuts its price and a customer switches from model B. A gains market share, but the AI market gains no new user; existing demand has moved from one provider to another. Economists call this substitution elasticity.

The bigger opportunity is whether cheaper, smarter, easier-to-use models cause people to do things they previously would not have done at all. Someone who never programmed might have AI build a personal tool; a company that never automated its work might hand a multi-step knowledge task to an agent. This is total demand elasticity: growing the pie rather than redistributing it.

So the most visible short-term number may be whose API traffic came from whom. The more consequential long-term question is who gets the most people who never used AI to start using it—and gets existing users to trust AI with more kinds of work. ChatGPT's empty text box created one such market. George is betting OpenAI can do it again.

Not strong therefore widely used, but widely used therefore stronger

The pieces of the argument now fit together.

The usual explanation is that OpenAI built a good model, obtained enough compute, and made a compelling product, therefore attracting a large user base. George reverses the arrow: OpenAI may keep making good models and products, and achieving better price-performance, precisely because it already reaches many different kinds of users.

The proposed loop is:

1. More users bring more varieties of real problems.
2. Recurring problems expose missing general-purpose capabilities.
3. OpenAI incorporates those capabilities into its models, tools, and platform.
4. Better products attract more users and enable outside developers to build more applications.
5. Greater scale supports more investment in compute, chips, and the broader technology stack, helping bring costs down.

Models, chips, and costs still matter. In this telling, they are simply not the first domino. That is the users and demand made visible by distribution.

This is also the boldest step, and the one that calls for the most distance. Available observations show ChatGPT spreading quickly, general-purpose AI entering many workflows, and agent usage rising. They do not establish that broad distribution necessarily produces the best model or chip. This is an investor's explanation for a possible future OpenAI advantage, not a proven law of business.

What this argument changes about what to watch

First, selecting today's best model is not a complete product strategy. Models matter, but the top benchmark position changes. A more lasting question is whether your product gets people to do something they previously did not do.

Second, a million calls do not necessarily mean a million insights about users. If every call comes through a partner, you may see requests without seeing the goals behind them. Alongside traffic, ask who directly faces the customer and who can see the original need.

Third, a platform is not an API with a new label. If others can genuinely build things on your system that you did not foresee, you need to solve code execution, data access, permissions, safety, and failure recovery for them. The moat may lie in precisely that unglamorous engineering.

Fourth, a larger opportunity than taking existing customers is awakening demand that does not yet exist. Making old work cheaper is valuable; enabling people to do things they could not, would not, or could not justify doing before is how customers are created.

Sam Altman's vision of future products can be compressed into two entry points: people who know what they want can have the system build their programs and tools, while those who do not know what to build can open ChatGPT and start with a sentence.

Both lower the hurdle of figuring out how to use AI. A small number of heavy users have already woven it through their lives and work. Many more people still open it only occasionally for a handful of tasks. The largest market has not fully awakened.

The article's real bet, then, is not that OpenAI will forever have the strongest model. It is that when intelligence becomes abundant, the scarce capability will be discovering what people want to do next, helping more of them actually do it, and turning their recurring needs into an ever-stronger platform.

OpenAI may not win. But George believes it comes closest today to combining those three steps: creating behavior, gaining distribution, and using that distribution to improve its platform. That, rather than any one benchmark lead, is why he thinks OpenAI's advantage may be hard to catch.