The AI race is just getting started

At WWDC, Apple finally showed the smart Siri. The one it promised back in 2024 and pushed back for two years. It is a ground-up rewrite now: it reads the context on your screen, works with your personal data, and carries out actions inside apps. The Passwords app walks through websites on its own and changes the ones that aren’t secure. Visual Intelligence looks through the camera and reacts to what it sees.

The reaction was predictable. Finally. They caught up. Even though, asked directly from the floor about agents, Apple admitted the system is still built mostly around queries rather than actions. And the models underneath are trained with Google’s Gemini as a teacher, while the flagship cloud version runs on someone else’s hardware, on NVIDIA chips in Google’s cloud, for which Apple reportedly pays around a billion dollars a year.

The whole debate over whether Apple is late rests on a single assumption. That we already know what winning in AI looks like.

And that is far from obvious.

The market rarely misses the direction, but often misses the source of value

Over the past months it has become fashionable to argue about who is winning the race. The list of winners looks predictable. People look at model size, the quality of assistants, investment volume, GPU counts, data centers, market cap. Against that backdrop Apple keeps landing among the laggards. No public model on the scale of GPT or Gemini. No assistant considered the best. Apple Intelligence leans heavily on outside models. The conclusion writes itself: it’s losing.

The problem is that this whole assessment keeps score using metrics we ourselves declared final.

The history of technology shows one pattern. The market reads the direction well enough. The mistakes come later, when it’s time to figure out where the value is actually created and who gets to keep it. In the internet boom everyone watched user counts, but the money went to those who built infrastructure on top of the traffic. In the social media boom they counted audiences, and the winners were those who turned the audience into an advertising machine. The intermediate metric almost always turned out to be real, just not where it was posted on the scoreboard. Today the scoreboard reads models, compute, and GPUs. The question is exactly the same. Is that the source of value, or only the price of entry.

The approach itself isn’t wrong. Model size matters. Compute matters. Infrastructure matters. But there is a wide gap between an important factor and a source of durable advantage. The biggest data center doesn’t mean the best business. The biggest model doesn’t mean a new economy. The smartest assistant doesn’t mean a lasting edge.

Technological significance and economic significance come apart in time

It’s worth recalling Acemoglu here. His position is often flattened into AI skepticism, but it’s about something else. The question isn’t whether AI will change the economy. The question is how fast and at what scale. And his estimate of the scale diverges sharply from the market narrative. In his paper “The Simple Macroeconomics of AI” he puts the ten-year effect at roughly a 1.1 percent gain in GDP and about 0.05 percent of productivity growth a year. Not zero, but a long way from the promises capital is being raised against. And a year and a half later, now looking at the first agentic products, he hasn’t raised those numbers. The main missing signal, he says, isn’t the power of the models but the absence of applications that make AI simple to use and turn raw capability into everyday work. The same kind of applications that once made ordinary office software ubiquitous.

For an investor that’s an uncomfortable thought. It forces you to look not at what the technology can do but at what it returns. And the rhetoric on Wall Street is slowly shifting. Not long ago the conversation was simple: more compute is better, more GPUs are better, more investment is better. Now you hear something else more and more often. Where’s the productivity. Where’s the profit growth. Where’s the return on capital. Where’s the economic effect of the hundreds of billions poured into infrastructure. The market hasn’t stopped believing in AI. It’s moving from admiring the technology to pricing its economics.

We talk too much about the conversation

For now the industry mostly debates the quality of answers. How well the system writes, how it holds context, how it reasons, how naturally it carries a dialogue. All of that matters, but we are gradually starting to confuse intelligence with the conversation about intelligence. A chatbot will draft you an email. An agent will find the right thread itself, pull in details from your calendar, and send it. At some point the main question stops being how well the system answers and becomes how well it acts.

The hard part begins where AI starts to act

Once AI starts doing things on its own, the cost of a mistake changes. A chatbot’s error is annoying. An agent’s error leads to lost money, leaked data, or a broken business process. The problem stops being about model quality and becomes about reliability. About a system’s ability to act under uncertainty without creating new risks. That’s exactly why nobody has a genuinely reliable general-purpose agent yet. Google shows an agentic Android update, Apple shows an agentic Siri, OpenAI has been in this race for a while. Plenty of attempts. The reliability ceiling hasn’t been broken by any of them.

The clearest case here is Anthropic, the company riding the highest wave right now. Claude already writes code across parallel repositories and runs autonomously for days on a single command. But when it came to acting on a user’s computer, the company itself wrote that this part is still raw compared to working with text, that the agent makes mistakes and the risks shift as you go. And the fix there isn’t to make the model smarter, it’s to lock it inside a sandbox the customer controls, within the boundaries of their own infrastructure. Reliability, again, is drawn not from the size of the model but from control over the environment where the model acts. The same logic Apple is working out, just from the other end.

The architecture of the new Siri shows where the center of gravity is moving. Underneath there isn’t one large neural network but five specialized models with an orchestrator over them. A light 3-billion-parameter model on the device for everyday tasks, a 20-billion model that loads only the fragments it needs for a given request, heavier cloud models for complex cases and image generation, and the most powerful one for long chains of reasoning. But the point here isn’t the models, it’s the orchestrator. A dispatcher inside the operating system that parses the request and decides what to run locally, what to send to the cloud, which app to reach. The hardest and most valuable part isn’t any single model, it’s the layer that connects them to real actions in the system. And that layer Apple keeps to itself.

The main asset may turn out to be actions, not queries

Most models learn from what people write, say, and ask. But there is another kind of data. What people do. The decisions they make, what they open and close, what they cancel, postpone, choose. This is data not about intentions expressed in words but about actual behavior. And Apple holds one of the largest positions in this kind of data, simply because billions of people live inside its devices every day. Except Apple has the dataset and deliberately left itself no access to it for training.

Apple has architecturally barred itself from training models on user data. Private Cloud Compute, which the heavy requests pass through, is built as a stateless system: data is processed and erased, never written to logs, never fed into training future models, and Apple publishes the source code so independent researchers can verify it. Even when the heaviest tasks run on someone else’s hardware, on Google’s servers with NVIDIA chips, the request goes into a cryptographically isolated container that Google has no access to. A competitor that collects user behavior for training grows its model on the real actions of millions of people. Apple gave up that resource on purpose. By the “most-trained model” metric it loses in advance, and by its own choice.

But the same architecture creates a different asset. An agent that reaches into your email, photos, and messages to act on your behalf doesn’t need the largest model, it needs a guarantee that it won’t leak or memorize your data. Apple controls the layer where an agent actually performs actions. The operating system, the chip, access rights, the sandbox, the interfaces between apps. The Siri orchestrator is precisely that in practice. A reliable agent doesn’t have to know everything about you in the cloud if it works right inside the system, where your apps, files, and permissions already sit. The section above ran into reliability as the industry’s main unsolved problem. Control over the execution layer is the lever that turns reliability from a property of the model into a property of the system. Here Apple holds a position almost no one else has. The only question is whether it can use it, or whether it stalls on execution again, the way it stalled on Siri for two years.

What earns respect is that Apple didn’t take the standard path. Catching up with the competition would have been easier if it dropped the constraint on privacy. Apple didn’t drop it.

The winner may be decided by entirely different metrics

The story that Apple is losing the race looks too simple. The model race, maybe it’s losing. The assistant race too. The flashy-demo race, definitely. But all those conclusions rest on the assumption that these are the contests that will decide the winner. And that is far from settled.

In five or ten years the deciding factor may turn out to be not the largest model or the prettiest interface but the ability to embed AI reliably into the real processes and everyday actions of billions of people. The main asset may turn out to be behavioral data rather than query data. And the winner may well be decided by metrics almost no one is discussing today.

The market is already naming winners with some confidence. It’s just not clear it understands the rules of the game.