Artificial intelligence continues to advance at a pace that would have seemed unthinkable just two years ago. New models, new integrations, and new product layers appear almost every week. But beneath that dizzying rhythm lies a question the sector may not be able to avoid much longer: can the AI business really sustain itself?

The doubt does not come from a lack of innovation. On the contrary. One recent move that best illustrates where the industry is heading is Google’s decision to bring Gemma 4 into Android’s AI Core Developer Preview. The signal matters because it points to a more mature product evolution: bringing capable models closer to the user, reducing dependence on the cloud, lowering latency, and opening the door to multimodal experiences that feel more natural inside the phone.

In other words, Google is showing one possible partial answer to AI’s big economic problem: if part of the processing can move onto the device, the experience may improve and some infrastructure costs could ease. But that same news also highlights a larger tension. Because even as the industry finds new ways to distribute AI, it is still unclear whether the economic model behind it is as solid as the narrative around it.

Google’s bet on AI closer to the user

The integration of Gemma 4 into Android AI Core is not just news for developers. It is also a strategic statement. For a long time, the most advanced AI seemed fully dependent on giant data centers, remote compute, and rising infrastructure costs. That approach remains crucial for frontier models, but it may not be the only way to scale product.

Google appears to be betting that an important part of the future will live in on-device experiences: faster, more private, more integrated, and potentially more efficient. If a model can run closer to the user, understand local context better, and operate like a native function of the phone, then AI stops feeling like an external service and starts becoming part of the operating system itself.

That matters not only for user experience, but also for business. In an industry where the cost of serving every interaction matters, any step that improves usefulness without always depending on more remote compute becomes strategically valuable.

The core problem: building AI is still extremely expensive

But even with bets like this, the heart of the economic problem does not disappear. Training advanced models costs enormous amounts of money. And serving them in production, updating them, fine-tuning them, integrating them, and scaling them to millions of users is not cheap either.

Modern AI requires more than talent and ideas. It requires:
- increasingly powerful GPUs,
- high-performance memory,
- data centers,
- energy,
- distribution networks,
- and full teams dedicated to safety, product, and deployment.

That brutal cost puts companies under double pressure. On one side, they need to keep launching new models and new features because the competition never stops. On the other, they need to prove that all this spending can turn into real and sustainable revenue.

That is where the discomfort begins. Because the public narrative around AI is full of growth, expectations, and adoption, but the balance between income and cost remains far less clear than many headlines suggest.

The price war makes it even harder

As if the base cost were not enough, the market is filling with competitive pressure. It is not only OpenAI, Google, Anthropic, Meta, or Microsoft competing. There are also open-source models, cheaper alternatives, specialized companies, and vertical solutions that force prices down or demand better justification for why one option should cost more than another.

That erodes one of the most obvious routes to profitability. If each more powerful model forces the rest to respond, and each response pushes prices down or increases promotional and infrastructure spending, then the race may look spectacular from the outside while becoming much harder to sustain on the inside.

On top of that, many users have grown used to getting an enormous amount of value at a relatively low price. And in the enterprise segment, even though interest in AI is extremely high, the pressure to show immediate ROI is also growing. Nobody wants to pay indefinitely for an abstract promise. Companies want measurable results.

More usefulness does not automatically mean more profitability

This is one of the most delicate points of the current phase. AI can become a better product without automatically becoming a healthier business. A model that is more useful, faster, or better integrated can drive adoption, yes, but it can also increase expectations, operational complexity, and maintenance cost.

With its on-device AI push, Google seems to be exploring a route where part of the future value comes from integrating AI more naturally and less expensively into the ecosystem. But the industry as a whole still has to answer a much bigger question: whether the next generation of AI will be only more impressive, or truly more profitable.

Because it is one thing to build something millions want to use. It is another to build something millions use that also generates enough return to sustain the infrastructure that makes it possible.

The next war will not just be about intelligence, but efficiency

That may be the real turn we are starting to see. During the first big wave, the battle was dominated by an obsession with capability: who reasoned better, who generated better, who impressed more. Now the next war seems to be shifting toward another axis: who can turn AI into a truly useful product without letting costs destroy the business.

That means efficiency, integration, smart distribution, and clearer monetization models. And this is why stories like Gemma 4 inside Android matter: they are not only about technical progress, but about an attempt to find a more realistic way to scale AI.

Conclusion

AI remains one of the most powerful technological revolutions of the moment, but it is no longer enough to say the future will be huge. The industry has to prove that future can also sustain itself economically.

Google is showing an interesting direction with its bet on bringing open, multimodal models closer to the device. But the bigger debate remains open: whether this whole ecosystem can keep its current pace without depending forever on massive spending, cross-subsidies, or inflated expectations.

The next big test for AI will not be only technical. It will be financial. Because in the end, the winner is not just whoever builds the most powerful model, but whoever makes that power make sense as both a product — and a business.