RAISE 2026 Insights: How AI-Native Teams Are Buying Infrastructure Now

I spent last week on the Hyperstack booth at RAISE in Paris, and it turned out to be one of the better rooms I've been in for reading where AI-native companies actually are right now. Less lecture hall, more founders shipping products at speed and watching the bill.

One shift came up again and again: almost nobody's asking whether the models are good enough anymore — that's mostly settled. The questions have turned sharper and more commercial. Can I get the compute? Can I run it economically? Can I keep control of it? Here's what I took from those conversations, and why it matters if you're scaling AI over the next year.

1. The conversation has moved to inference and token economics

Last year, the energy was still around model development. At RAISE, it had clearly moved downstream. Inference, token economics, and capacity were the three things I heard competitors and customers circling back to, over and over, usually in that order.

That's a healthy sign, honestly. It means the market is maturing past the "who has the biggest model" phase and into the "who can actually serve this profitably" phase. When you're running production AI at real volume, the question that keeps you up at night isn't benchmark scores. Rather, it's cost per token, and how much useful work you're getting out of each GPU you're paying for. Several vendors were pitching data-pipeline and inference-acceleration tooling for exactly this reason, because that's where the money now leaks or gets saved.

This is the part I find genuinely energising, because it's the territory we've been building in for a while. The story stopped being about who can stack the most GPUs and became about who gets the most out of each one. Hyperstack AI Studio is built around that — deploying, fine-tuning, and serving open-weight models efficiently, with the inference optimisation work (things like KV-cache handling) done where it actually moves cost per token. For a team watching its inference bill climb, that's a more honest value story than raw capacity bragging, and it's the one more people at RAISE were ready to have.

2. Capacity is a "now" problem — and the one-year-term playbook is breaking

The single most consistent thread on the booth was urgency. People aren't planning capacity for some abstract future; they want it soon, and they're increasingly aware that the supply situation means they have to plan further ahead than their procurement cycles are used to.

Here's the pattern I saw most often, and it's worth being blunt about: teams would come in still expecting to buy the very latest architecture on a one-year term, the way they always have — and then learn, sometimes mid-conversation, that this just isn't how the market works right now. Demand for current-generation hardware is running well ahead of what the supply chain can turn out, and lead times on data-centre GPUs are stretching to the better part of a year. That's not a Hyperstack problem or an NVIDIA problem; it's a whole-industry demand signal, and it's reshaping how the smart teams buy.

The ones getting ahead of it are the ones working with a partner who's already secured allocation rather than trying to source it cold. That's a big part of what we do — we commit to capacity ahead of demand so you don't have to join the back of a 9-to-12-month queue. Our NVIDIA B300s land on-demand in August, with reserved private clusters following in Q4, and that near-term window is a real opening: if you know you'll need serious compute this year, this is the moment to line it up rather than hope it's there when your procurement cycle finally catches up.

3. Sovereignty grew up

Sovereign AI came up constantly, and it's clearly graduated from a compliance checkbox to a live strategic use case. I even ended up in conversations about ventures being built specifically around sovereign AI. What struck me is how the framing has shifted for this audience. For a regulated enterprise, sovereignty is about compliance and data control. For an AI-native company, it's increasingly about independence — not wanting your product's future to hinge on a single provider's jurisdiction, pricing, or roadmap.

That worry isn't abstract anymore. Earlier this summer a single government directive reportedly cut global access to two frontier models overnight, foreign customers included, with no real notice. Access came back, but the lesson landed: if your product runs on someone else's closed API, you inherit their availability and their rules. I want to be fair here — that's not a knock on the labs building those models. They're doing exceptional work and for plenty of use cases they're the right call. The point is narrower: depending on a single jurisdiction is now a risk you plan for rather than assume away.

The other half of this, and it came up in almost every sovereignty conversation, is that open-weight models are improving fast. The gap to the frontier has closed enough that running open weights on infrastructure you control is a real option now, not a compromise. That's the combination we've built for: Hyperstack Secure Private Cloud gives you dedicated, single-tenant infrastructure sited in EU or Canadian jurisdictions under EU law. Alternatively, AI Studio through the Hyperstack platform allows you to run open-weight models you control through serverless inferencing.

Where that leaves us

Pull those three threads together and they're really one story. The AI-native companies I met in Paris are scaling past the point where the hyperscalers comfortably serve them — on economics, on capacity access, and on control — and they're looking for an infrastructure partner built specifically for that stage. That's the position we're leaning into, and RAISE made me more convinced it's the right one.

Concretely: if your pain is inference economics, that's AI Studio and efficient, utilisation-focused infrastructure. If it's getting current-generation NVIDIA hardware in a realistic timeframe, that's our secured allocation — NVIDIA B300s on-demand from August, reserved clusters in Q4. And if it's control and compliance, that's Hyperstack Secure Private Cloud with open-weight models you own. Three problems, one platform.

If any of that is live for you right now reach out to me directly. Tell me what you're trying to run, your timeline and how you're thinking about control, and I'll give you an honest read on what's realistic and what it would take.