In the first week of April, I spent days immersed in conversations with some of the sharpest minds in artificial intelligence — researchers from Google DeepMind, NVIDIA executives, founders redefining entire categories, investors betting on what comes next. It happened at Brazil Silicon Valley, in San Francisco, and at the Google AI Summit. What I'm sharing here isn't a recap of talks. It's a distillation of what lingered after the noise faded — the ideas that kept echoing, connecting with each other, forming something that felt urgent to put into words.

The cost of not starting

There's a rare consensus among people who normally disagree on everything: this is the most exciting moment to build products, companies, and new categories. Possibly the most exciting in the last thirty years.

This isn't generic optimism. It's math. AI models improve every week. They get cheaper every month. Someone from DeepMind said something that stuck with me: "If you froze all model development today, for two years, you'd still have more than enough room to build incredible things with what already exists." And the models aren't stopping.

"We now live in a world where you can build anything you want."

The cost of starting has never been lower. But what few people account for is the cost of not starting. And that cost has never been higher.

‍

Infrastructure isn't used — it's inhabited

This accessibility changes the game in a way most people haven't fully internalized yet. "The what and how of software have fundamentally changed." Simply "using AI" as a tool is no longer enough — it has become infrastructure. And you don't use infrastructure. You inhabit it. The entire company needs to be redesigned from the premise that AI agents will be part of the team — not as assistants, but as operators.

Companies that try to fit AI into processes designed for humans see modest gains, around 5%. Those that rethink the organization from scratch — data, workflows, hierarchies, decision-making — see gains of 100 to 300%. That's not an exaggeration. Those numbers came up across multiple conversations, from multiple sources, with concrete examples.

At one panel, the founder of one of the most relevant AI agent companies in Silicon Valley mentioned, almost casually, that three of his best "engineers" are already AI agents. Not people using AI. Autonomous agents that write, test, and deploy code. He found out when he saw two of them talking to each other in the company Slack.

If your company is AI-native in the narrative but not in the architecture, the market will expose that gap — and faster and faster.

And here's the reframing I consider the most important thing I heard: the greatest benefit of artificial intelligence isn't just cutting costs. It's doing what wasn't possible before.

Most companies still look at AI and think efficiency — doing the same things, faster and cheaper. That's real, and it has value. But those capturing disproportionate value think differently. They think in terms of possibility.

I live this every day. A global brand needs to maintain visual and tonal consistency across thousands of pieces, for dozens of markets, simultaneously — each one adapted to the channel, the audience, and the local cultural context. The speed and scale the market demands made this simply impossible. It wasn't a cost question. It was a constraint question. With artificial intelligence orchestrating creation, consistency, and personalization, that constraint no longer exists.

‍

What stage is your company at?

The framework that emerged from conversations divides companies into four maturity stages:

Stage 1: Getting out of denial. Accepting that AI isn't passing hype.

Stage 2: Individual adoption. People using AI tools day to day, each in their own way.

Stage 3: AI embedded in culture, data, and organizational processes. This is about efficiency.

Stage 4: Reimagining the entire sector. This is about possibility — and almost no one is here.

Most companies are stuck between stages 2 and 3. But stage 4 is where disproportionate value lives.

But possibility without design is noise. And here comes a principle that sounds simple but changes everything when building product: inject intelligence into the interfaces people already use. Don't force a new behavior. Slot into the existing flow.

The product that requires habit migration loses. The one that fits invisibly into the existing workflow wins.

For those building AI products, the lesson is direct: the best artificial intelligence is the one the user doesn't notice they're using. It simply makes everything better.

‍

Generic is the new invisible

And "better" today means personalized. Not as a differentiator — as a baseline requirement for existence.

Hyperpersonalization is no longer a premium feature. It's table stakes. If your product treats everyone the same, you're already behind. The conversation isn't "let's offer personalization." It's "without personalization, there is no product." As models get cheaper and more capable with each cycle, personalizing at scale stops being a technical luxury. It becomes economic viability.

But personalizing without understanding culture is just sophisticated translation. And here is perhaps the most underestimated insight from everything I heard: real localization isn't linguistic. It's cultural. The magic is in the vibe check.

Translating to English isn't localizing for the US. Adapting tone, aesthetics, references, timing, emotional context — that's localization. The world is increasingly flooded with generic AI-generated content. Everything looks the same. Everything sounds the same. What will separate signal from noise is the ability to be culturally precise. Those who master that don't compete on price. They compete on relevance.

‍

Data is your moat — but trust comes first

And relevance, in the long run, is built on data.

I heard this in at least four different conversations, from completely different profiles: data is your moat. The model isn't your differentiator — it's rented intelligence, available to anyone. Your data flywheel is what makes you hard to replicate.

This means startups that treat infrastructure as an IT decision — rather than a product decision — break down when it's time to scale. Cloud bills explode. Cost per token destroys unit economics. Token economy is the new survival metric.

The long-term defense rests on three pillars: proprietary data that improves with use, communities that reduce acquisition costs and create organic distribution, and rigorous control of how much each intelligence operation costs. Without that, you're a commodity with negative margins.

But to have data, you need something that comes before it: trust.

Trust is non-negotiable. Not as an abstract principle — as a pre-condition of business. If the client doesn't trust that their assets, their information, their identity are safe and handled with care, there's no relationship. And without a relationship, there's no flywheel. No data. No moat.

Leadership is doing the right thing when it costs money. It's turning down investments in lucrative sectors because they don't generate value for the world. It's maintaining ethics when no one is watching. The trust that builds is the hardest asset to construct — and the hardest to copy.

‍

Play where you win

Trust, data, cultural intelligence, AI-native architecture — all of this requires one fundamental decision: where to play.

Don't try to be everything to everyone. Don't try to be in every market. Play where you win. Focus on the highest-value segments where you have a real advantage — context, data, depth of problem understanding. NVIDIA, with all its scale, explicitly leaves the greenfield of applications to startups. Because it knows it can't have depth in every vertical. No one can.

A lot of people have ideas. That's relatively easy. The challenge is execution. And execution requires focus.

It also requires a healthy relationship with time. Think short-term. Be conscious of the long. But don't paralyze yourself trying to predict what comes in five years — you'll drive yourself crazy and it won't help. The competitive advantage cycle has compressed to eight weeks. What works at eighty percent today will work at ninety-nine percent in six months. Build for the future, because by the time you finish building, the new model will make it work.

You never know how close or how far you are from success. But you know that stopping has a certain cost.

I left that week with a lot on my mind. But one conviction rose above everything else.

We are living the most exciting moment to build.

The technology is accessible. Models improve every week. The barrier between having an idea and putting it into the world has never been thinner. But the window isn't infinite. Thirty years have passed since the last wave of this magnitude. Those who act now build. Those who wait, watch.

I'm lucky to be living this from the inside. At Pupila, I'm close to some of the brightest minds in the world, partnering with the largest companies on the planet, transforming the way brands are built with artificial intelligence. There's no better moment to be exactly here.

And the only conversation that truly matters isn't about price. It's about value. About what becomes possible — and what becomes impossible without it.

Everything is hard. But you have to do it.

‍