I spent a couple of days in Paris this week — investor meetings and time with developers at Android Makers (by DroidCon). It’s a familiar scene: T-shirts, snacks, high energy, and a relentless focus on building and shipping faster.

One session stood out.

A brilliant Google session — “Build intelligent Android apps with Google’s AI” — by Jolanda Verhoef, Android Developer Advocate at Google. It looked simple. It wasn’t.

What she showed was a very clean model: some AI runs directly on the device (Gemini Nano), some runs in the cloud (Gemini Flash/Pro), and developers access everything through a single API. From their perspective, it’s almost trivial. You make the call, and the system decides where the work happens.

That’s the shift. AI is no longer something you integrate. It’s becoming infrastructure. This got me thinking…

From centralized AI to distributed intelligence

For the past few years, AI has been defined by scale — massive GPU clusters, hyperscalers, and centralized training. That layer still matters, but it’s no longer where the real change is happening.

The change is in inference, and inference is moving everywhere.

Into phones. Into laptops — we’ve already seen this with local models and tools like OpenClaw. Into edge environments. And increasingly, into networks. Models are getting smaller, devices are getting smarter, and the boundary between device and cloud is disappearing behind orchestration layers that dynamically decide what runs where.

This fundamentally changes the system. Latency, cost, privacy, and reliability are no longer fixed design choices. They become variables in a distributed environment.

The hard part is no longer building the model. The hard part is deciding where it runs.

Developers will abstract it away—but that doesn’t make it disappear

At the DroidCon event, none of this complexity was visible. Developers are already operating in a world where AI is simply “there.” You want voice, summarization, reasoning — you call an API, and it works.

And that’s exactly how it should be.

But abstraction doesn’t remove complexity. It relocates it. The decisions around cost, performance, compliance, and control don’t disappear — they move deeper into the stack, into infrastructure and orchestration layers that most developers will never see.

That’s where the real battle is being fought.

The real battle is orchestration

Every AI interaction now involves a set of decisions: where should this run, which model should be used, how do you balance cost against latency, and how do you stay within regulatory boundaries?

This is no longer a backend detail. This is the system.

We are moving into a world of continuous, real-time orchestration of intelligence across devices, edge, and cloud. And that orchestration has to optimize not just for performance and cost, but for compliance, control, and trust.

Telcos are back—just not in the way we expected

If you only looked at the developer conversations in Paris, telcos are irrelevant. There is no talk about Network APIs or CPaaS. Developers are building on Google, not on operators.

But that’s a surface-level conclusion.

Because underneath the abstraction, infrastructure suddenly matters again. Proximity matters. Control matters. Regulation matters. And that’s exactly where telcos sit.

Interestingly, this was the exact topic of a separate discussion I had in Paris with a French telco headquartered there. We were talking about sovereign AI, local infrastructure, and how networks need to evolve. On its own, that conversation sounds like a typical telco strategy discussion. Seen alongside what’s happening on the developer side, it becomes clear that both worlds are converging.

Voice, regulation, and reality

Two areas stand out.

The first is voice. As AI becomes conversational, real-time voice infrastructure becomes critical again. Phone calls are still one of the most reliable and ubiquitous communication channels in the world, and they run over telco networks. Embedding AI directly into that layer is an obvious next step. This was explored in depth in a recent CPaaS Acceleration Alliance research report on AI Voice, with plenty of real-world examples and market data.

The second is regulation — and this is where things get uncomfortable.

In Europe (and more and more in other parts of the world), anything that touches consumers — contact centers, customer service, regulated industries — has to meet very high standards. GDPR already sets a strict baseline. The EU AI Act raises it further. Enterprises need auditability, explainability, and control. They need to be able to comply with requirements like the “right to be forgotten.”

Now combine that with the current generation of large, centralized LLMs.

You quickly run into a problem.

These systems are difficult to audit, difficult to control, and difficult to localize. In many real-world, consumer-facing enterprise scenarios, they are close to unusable in their current form. Add to that the strong and growing requirement to keep data within regional boundaries, and the gap becomes obvious.

This is why AI moves back to the edge

This shift is not just about performance. It’s about control.

Enterprises need to know where their data is processed, how decisions are made, and whether outputs can be trusted and audited. That naturally leads to more local execution — on-device, in-country, or within regional infrastructure.

Suddenly, the architecture Google showed — device plus cloud, connected through orchestration — is not just elegant. It is necessary.

And this is where telcos have a real opportunity. They already operate within regulatory frameworks, they already manage data within borders, and they already have infrastructure that can be upgraded to support this new model.

Don’t ignore the hardware layer

There is another layer that becomes critical in this transition: hardware.

If AI is going everywhere, it needs to run everywhere. That means new chipsets, new architectures, and new supply chains. Companies like Intel and AMD play a crucial role here — not just as chip designers, but as part of the broader ecosystem enabling distributed AI across devices, networks, and data centers.

You don’t get distributed intelligence without distributed compute.

From building models to coordinating systems

The first phase of AI was about scale. Bigger models, bigger clusters, more compute.

In many ways, that was the easy part.

Now we are entering a much more complex phase. Intelligence is fragmented across devices, networks, edge environments, and cloud platforms. Coordinating that system — deciding what runs where, under which constraints, and at what cost — is a fundamentally different challenge.

That is where the next wave of companies will be built.

Bringing it all together

That’s also why the mix of conversations in Paris mattered.

On one side, developers are being handed incredibly powerful tools that make AI feel effortless. On the other side, infrastructure players—telcos included — are trying to figure out how to stay relevant in a world that is becoming more distributed, more regulated, and more complex.

And then there’s capital, looking for where real value will be created.

The opportunity sits right at the intersection of those three worlds.

We need to bring telcos, investors, and startups together. Because the future of AI will not be built by any one of them in isolation. It will be orchestrated across all of them.

That’s the real shift.

And that’s why this moment matters.