AI IndustryOctober 9, 20268 min read

ALL IN 2026: Lego Blocks, a Strained AI Stack, and a Lot of Sovereignty

By Jacob Cossette

Notes from ALL IN 2026 in Montreal: Factiverse, Mistral, the AI infrastructure stack, and why Canada needs to go big on sovereign AI.

I spent a day at Canada's biggest AI conference, and the loudest message was not about smarter models. The hard problems now sit around the model: the data, the power, the people who have to connect it all, and the question of who controls it.

A day at the Palais des congrès

ALL IN ran on September 16 and 17 at the Palais des congrès de Montréal, and I was there on the first day, September 16. The organizers counted more than 6,500 attendees from over 40 countries, with 200+ speakers. Walking the floor, I passed booths and stages from NVIDIA, Google, Cohere and plenty of other model builders and infrastructure players.

I bounced between talks all day, and a handful of ideas stuck with me. I'm writing them down while they're fresh, in the order they hit me.

Factiverse: 22 minutes instead of 20 days

The talk I keep coming back to came from Factiverse. The company is Norwegian, based in Stavanger, and grew out of fact-checking research at the University of Stavanger. Its CEO showed a platform that does in 22 minutes an analysis that used to take 20 days, with secure data and a decision you can defend at the end of it. The numbers come from their presentation, so read them as the company's claim.

What makes it interesting is how it gets there. Factiverse does not lean on a giant general model to guess the truth. It uses a model built around information retrieval and trained on curated sources and fact-checkers' work. The company says it identifies claims worth checking in 114 languages, with a success rate around 80%, and that it beats general-purpose models at that specific task. It has also done live fact-checking during the 2024 U.S. presidential debates, and it won best pitch in its category at TechCrunch Disrupt Battlefield 200.

My takeaway: a narrow tool built on trustworthy data can compress weeks of human work without asking anyone to trust a black box. Speed only matters if you can stand behind the answer.

Mistral: applied AI should be built like Lego

Marjorie Janiewicz, Mistral's Chief Revenue Officer, gave the clearest framing of enterprise AI I heard all day. Applied AI, she argued, should be built like Lego: modular blocks you assemble for a specific job. What closes the gap between a demo and a working system is customizing the model to the IP of the business. That is where Mistral says it is heading in the coming months, with smaller models that are deeply specialized rather than one giant model for everything.

Her example stuck with me: if you are a bank, you need a great OCR model, because reading documents correctly is the job. The shift she described was striking. About half of Mistral's conversations with clients are no longer about the limits of the foundation model. They are about how to customize the model to the client's needs.

This lines up with what Mistral announced in March at NVIDIA GTC. Its Forge platform lets enterprises and governments build models on their own data, and Mistral says it supports training from scratch rather than only fine-tuning or retrieval layers. Mistral's head of product put the trade-off simply: smaller models cannot be as good at everything, so customization lets you choose what to emphasize and what to drop.

The moat is moving from the model to the data, the workflows and the institutional knowledge wrapped around it.

The AI infrastructure race is a stack, and the stack is strained

The session that changed how I picture AI was about infrastructure. Think of AI as a stack, each layer depending on the one below:

Energy
Cooling
Chips
Servers
Models
Applied AI

Cooling is the layer I'd underline. Nearly all the electricity a chip draws leaves as heat, and that heat has to go somewhere or the hardware slows down and fails. As chips get denser, air stops being enough: TrendForce estimates liquid-cooled racks will reach roughly 47% of deployments this year, and Futurum ranks energy and cooling ahead of chip supply as the main constraint on AI growth. Mayfield's infrastructure analysis calls power and cooling the gate everything else has to pass through. A data center with the best chips and no way to cool them is an expensive heater.

NVIDIA's Jensen Huang has described a similar five-layer cake of energy, chips, computing infrastructure, models and applications. The speakers' point was that each layer is manageable alone. Having one or two of them is relatively easy. Aligning the whole stack so that no single layer becomes the bottleneck is really hard, and the current bottleneck is often the missing connection between people from different industries who rarely sit in the same room: utilities, cooling specialists, chipmakers, model labs and the companies deploying the result.

Outside analysts say the same thing. Futurum's 2026 outlook concluded that energy and cooling constraints have overtaken silicon availability as the main brake on AI expansion. TrendForce estimates that liquid-cooled racks will make up roughly 47% of deployments this year. Software moves in months, while power plants and data centers move in years.

One more challenge came up that I don't have a neat answer to: how to distribute AI into physical AI, meaning the machines, vehicles and devices that act in the real world. Getting intelligence out of the data center and into those systems is the next version of the same alignment problem.

Sovereignty, sovereignty, sovereignty

If the day had a single word, it was sovereignty. All day, the conversation kept coming back to who owns the compute, the models and the data. My view after a day there: Canada needs to go big on this, and the next year will show whether it does.

The groundwork is real. The federal Sovereign AI Compute Strategy commits about C$2 billion, split between C$700 million to attract private data centres, C$1 billion for a national supercomputing facility, and C$300 million to give smaller businesses affordable compute access. In June, Ottawa released its refreshed national AI strategy, AI for All, whose sovereignty pillar rests on the idea that sovereign AI starts with sovereign infrastructure. Plans include a public supercomputer by 2031 and a proposed C$500 million Canadian Tech Growth Fund so promising AI companies don't scale up elsewhere.

What the conference added for me is that sovereignty is more than data centres. It connects to the other talks. Customizing a model to your own IP, as Mistral described, is a form of sovereignty at the company level. Verifiable, curated data, as Factiverse showed, is sovereignty over what you believe. And every layer of that infrastructure stack needs power and cooling that has to come from somewhere. People can reasonably disagree on how much to build at home versus partner abroad. But if the stack is going to be aligned, someone has to decide where Canada plugs in, and the sooner the better.

AI is not a magic tool

The last idea is the one I most want to leave you with. AI is not a magic tool, and just because you can solve something with it doesn't mean you should.

Every example I liked at ALL IN had a clear problem, a narrow scope and a reason to trust the output: a verification platform grounded in curated sources, a specialized model for a bank's documents, an infrastructure plan that admits the physical limits. Each AI project also draws on that strained stack of power, cooling and compute, so the real question is whether the problem is worth the resources.

So here is what I'm taking back to my own work: pick the problem first, then ask whether AI is the right block to build with, and know whose rules the system runs under. It's less glamorous than a demo, but I think it's how this technology earns trust.