RAISE 2026 was superbly run. Two days at the Carrousel du Louvre, 9,000 people, 350 speakers, and it was amazing. Credit to everyone who made it work. I had an incredible time, and it made me think an enormous amount about the moment we’re in right now.
Here’s my question.
How do you take the potential of AI, operationalize it, and deploy it into reality so that it’s adding value for users? The mess, cost, speculation, moneymaking and competition of that race-to-market is the big challenge and question we all have to answer, whether we’re building frontier AI models or implementing AI tools in our departments and personal life.
Here’s my thoughts on that theme, and the day more generally:
The investment energy was extraordinary – billions flooding into data centres, neoclouds racing to market. CoreWeave, Vultr, IREN. Rafay's Mohan Atreya made a smart point about customers wanting regular compute and AI in the same place. Real momentum, real engineering.
Here’s my question: at the moment, most of the capital is landing at the compute layer. That's the furthest point from the person actually using the AI. So, will the investments be symmetrical to the returns? If the intelligence-as-utility metaphor ends up being borne out, I wonder which investor set will make their money back… the outfit investing in more infrastructure, or the outfit investing in users, product use cases, and the detail of user problems?
That leads us to the bubble question. The dot-com comparison came up repeatedly at the RAISE summit. Are we in a bubble? Is it going to “pop”? Who will be left holding all the shares when the music stops? And per the above, will those shares be in the value-capture half? Or something else?
For me, when the correction comes, what survives is whatever can prove it has a real customer base willing to pay for it – because doing this well makes the shares valuable when the curtain is lifted. Helping businesses to deliver product/market fit around the world is where Global App Testing excels, of course – so we can help you there.
One of the great software clichés is that change happens slower than you expect, and then all at once. I think that's especially true of market penetration — and AI is about to learn it the hard way.
There's an idea I picked up from Nataly Kelly, an internationalization expert, that has stuck with me: a market isn't a switch you flip, it's a queue you join. Language, payment rails, device mix, regulation, trust, habit. Each one is a small delay. Together they're the reason your brilliant product does nothing in Brazil for eighteen months and then, suddenly, does everything. Commercial leaders don't like this. We prefer models where an action produces a proportional result — spend here, get customers there, draw the line, forecast the quarter.
But real-world adoption doesn't move in a line. It moves in a long flat stretch of accumulating detail, followed by a step change once enough of that detail resolves.
MACHINA, the Physical AI summit, was the best hour of my week — humanoid robotics, industrial automation, machines operating in the world. The room crackled.
There's an old essay by the programmer John Salvatier called Reality has a surprising amount of detail. His argument is simple: every plan looks clean until you try it, and then it explodes into a thousand tiny obstacles nobody could have listed in advance. That's what "change management" means. The undocumented workflow. The edge case that only shows up in Seoul. The user who holds their phone differently. It reminds me of the “deal-is-difficult” challenges we find when our customers try to launch their products globally.
It also proved Salvatier's point in the most literal way possible. Software can hide its unknown details behind a slightly wrong answer. A robot can't. Floor surface, lighting, the exact height of a doorway. Physical AI doesn't discover the gap in a dashboard. It walks into it.
Every conversation above points to the same conclusion. The hardest part is not proving that AI can work. It is proving that it works for real customers, in real environments, across real markets.
If you're deploying AI to real customers, you already suspect where your surprises are hiding — the market you haven't tested in, the device you don't own, the workflow nobody wrote down. We built the Readiness Gap scorecard to make that suspicion measurable: 12 questions, six dimensions, a number you can take to your board.