Web Summit 2026 marked the end of the “demo culture” era. Twenty thousand people from over 100 countries filled the Vancouver Convention Centre, and the investor turnout was the largest the event has ever seen. But the mood had shifted in a way that was hard to miss by day two. The startup fantasy version of AI, the one where everything is possible and the only question is how fast, was not the dominant conversation. People were asking harder questions. And the answers, when they came, were uncomfortable enough to stay with you on the flight home.

The conference did not announce a single breakthrough technology. What it did, consistently and across sessions, was ask a different kind of question: Who controls AI? Who benefits from it? And who absorbs the cost?

That reframe tells you more about where the industry is right now than any product launch could.

The official AI Summit track was framed around “The great AI debate: Who decides our future?” That framing was not rhetorical. It structured real discussions around open versus closed AI systems, data sovereignty, national infrastructure investment, hyperscaler dominance, and whether countries outside the US-China axis can build anything independent at all.

The answer, for now, is: with difficulty.

The demo era is over

For the past three years, global technology events followed a predictable format. Keynotes built on technical potential. Founders showed what was possible in a controlled environment. Attendees left with excitement about a vague automated future and a bag full of branded merchandise.

Vancouver felt like the end of that cycle.

The people in that room have spent two years running pilots, sitting through presentations, and testing tools that worked beautifully until they had to work in production. The tolerance for polished demos without evidence has dropped. What replaced it in Q&As and corridor conversations was a much simpler demand: show us what it actually fixed, and tell us the number.

This is the shift we at Itera Research have been watching in our own client conversations for the past year. Most organizations are not early in their thinking about AI anymore. They have run the pilots. They have seen the demos. What they have not yet found, in many cases, is a system that solves a specific expensive problem reliably enough to replace the way they currently work. Until it does that, it is background noise.

The companies that are winning this moment are not leading with the fact that they use AI. They are leading with what they have stopped losing.

The session people kept bringing up

If there was one conversation from opening night that kept resurfacing in hallways and side sessions throughout the week, it was Sigrid Jin’s.

Jin, founding member of Sionic AI, described something that was either a legal nightmare or a preview of where software is heading, depending on who you asked. While on an airplane, he learned that Anthropic had accidentally exposed the source code for Claude Code. He moved immediately, burning through billions of tokens to reconstruct and translate the code across multiple programming languages. The result became Claw Code, an open-source alternative.

He was not apologetic. His position was that code is becoming a public good, that the focus should be on what software produces rather than who owns the lines, and that the legal concepts around copyright were written for a world that no longer exists. The irony was not lost on the audience. Anthropic itself trained on copyrighted material, and now its own code had been replicated through the same logic Jin was defending on stage.

That tension was still alive in conversations on day three. For anyone building on proprietary systems or licensing technology to clients, it is not a philosophical debate. It is a live business risk.

What the floor was actually talking about

The gap between what appears on stage and what people discuss in the corridors was wider than usual this year.

On stage: national strategy, infrastructure investment, open versus closed models. In the corridors, founder circles, and side meetups: something closer to organizational anxiety. The language we kept hearing, leaner teams, higher output per person, workforce redesign, was not the language of opportunity. It was the language of companies mid-restructure, figuring out how to rebuild around systems rather than headcount, and not entirely sure what to do with the people in the roles that are shifting fastest. Marketing, operations, recruiting, and coordination. Those came up repeatedly.

Nobody said “humans are becoming cheaper” in a keynote. The conversation was happening anyway.

There was also something else visible on the floor: exhaustion with companies that still lead with demos. The people in that room have spent two years building with these tools. They know the difference between a polished demo and something that actually runs in production. The tolerance for the former had dropped noticeably. What people wanted to see, and kept asking for in Q&As, was working proof. Real numbers. Evidence that the thing holds up when the demo ends.

The finding that surprised us most

The more capable AI becomes at producing content, code, and analysis at volume, the more the conference conversation kept returning to what does not transfer.

Trust. Relationships. Judgment built from experience in a specific domain. The kind of accountability that comes from someone putting their name behind a decision.

“The voice, the relationships, the judgment calls… those stay with the human.”

Shared widely from conference floor conversations, Vancouver 2026

When content becomes cheap to produce at volume, the things that cannot be generated, genuine credibility, real relationships, specific human judgment in high-stakes moments, become scarcer and so worth more. The organizations that understand this are not asking how to replace human work. They are asking which human work matters most, and whether the people doing it have what they need to do it well.

That question did not get a session. It probably should have.

Data sovereignty landed as a political argument, not a technology one

Paddy Cosgrave framed the opening night around a tension that shaped the rest of the conference: the battle between open-source and closed AI. On one side, trillions of dollars backing the belief that a small number of American firms will own the foundation layer of AI for everyone. On the other, open-source models, including Chinese ones, closing the performance gap fast.

Canada used the home-field advantage deliberately. Evan Solomon, the country’s first Minister of Artificial Intelligence and Digital Innovation, defined sovereign AI not as an aspiration but as a practical requirement: government and citizen data protected under Canadian law, not subject to external pressure from another country. He was honest about the difficulty. He acknowledged the gap between what AI can do and what the public actually trusts it to do, and said plainly that concerns about sustainability, power usage, and job training need to be spoken about openly rather than managed around.

One number spread quickly from those discussions: only four countries currently have large language models. Canada, France, China, and the USA. That is the actual competitive field. For every other country, and for most businesses, the practical question is not whether to build a model. It is who owns the infrastructure you depend on, and what happens if that relationship changes.

The business implication follows directly. The specific model a mid-market company uses is increasingly a commodity. Anyone can access the same models. The thing that cannot be replicated is the data a company owns, how it is structured, and how deeply it is woven into the way that company operates. Organizations whose business intelligence sits entirely inside a third-party tool have no real independence. Ownership of the data layer is where defensibility actually lives.

What we’re taking back

Vancouver 2026 did not produce a consensus. That may have been the point.

What it produced was a clearer picture of the arguments that will shape how organizations build, buy, and govern AI over the next few years. The open versus closed debate is not technical. It is political and economic. The sovereignty conversation is no longer abstract. Governments are spending real money and making real alliances. The shift from demo to proof is already happening in how serious buyers evaluate vendors.

For the organizations we work with, the practical implication is this: the choices being made right now, about data infrastructure, about which systems to build versus buy, about what human judgment to protect and what to hand off, will determine who has real capability in two or three years. That window is still open. It will not stay open indefinitely.

AI is no longer impressive by default. That is actually good news for anyone who has been doing the hard work instead of the easy announcements.

Itera Research works with organizations at the strategy and implementation level of AI adoption. If your team is working through the decisions raised at Vancouver, the AI Discovery service is a structured way to start.