What ISC 2026 Told Me About Where AI Infrastructure Is Actually Heading

I spend most of my week talking to teams about the practical side of running AI and HPC workloads: what they can get, when they can get it, and what it costs them to keep it running.

So I go to ISC less for the keynotes and more for the corridor conversations. This year, those conversations were sharper and more specific than I've heard them in a while.

If I had to sum up ISC 2026 in one line, it's this: the industry has stopped talking about AI infrastructure as an abstract ambition and started talking about it as a supply problem, a sovereignty problem, and a physics problem — all at once.

Here are the takeaways I keep coming back to, and why I think they matter if you're scaling AI for the next 18 months:

1. Sovereignty stopped being a compliance conversation and became a dependency one

The mood on this shifted noticeably, and I think everyone knows why. On 12 June, the US Commerce Department ordered a suspension of global access to two frontier models — including for foreign nationals — citing a national-security concern. No advance notice, no technical findings shared, and the practical effect was a worldwide shutoff for every customer, not a targeted block. Access has reportedly since been restored, but the point had already landed.

What made this different from the usual AI Act discussion is that it wasn't a regulatory story. It was a geopolitical one. This was widely described as the first time US export-control authority had been applied to a commercially deployed model's access controls rather than to a chip or a hardware sale. Which means, in principle, any frontier API could be subject to the same mechanism. One European parliamentarian put it more bluntly than I could:

"Europe cannot keep building its tech stack on access that can be switched off overnight by a foreign government."

I heard versions of that sentiment all week. And here's the nuance that mattered to the people I spoke with: sovereignty is no longer just about where your data sits. You can run a perfectly GDPR-compliant deployment and still have the model underneath it switched off by a decision made in another jurisdiction. The conversation has moved from "is my data resident here?" to "who can turn this off, and what happens to me when they do?" For organisations that can't tolerate that kind of interruption, depending on a single jurisdiction is now a risk to be planned for, not assumed away.

This is exactly what we built Hyperstack Secure Private Cloud for. It's dedicated, single-tenant infrastructure that can be sited in EU or Canadian jurisdictions, operated by a UK-based provider rather than a US hyperscaler. Pair it with Hyperstack AI Studio, where you can fine-tune and deploy open-weight models on infrastructure you control, and the “who can turn this off?” question stops being a live risk: the infrastructure sits in a jurisdiction that isn't exposed, and open weights can't be dark-switched by anyone's directive. That's the whole point — sovereignty across the stack, not just at the data-residency layer.

2. The hardware crunch is structural, not a blip

If you've felt like GPUs and memory have been harder to get hold of, you're not imagining it, and it isn't going to resolve itself quietly. The clearest number I took away: data centres are expected to consume up to 70% of all high-end memory chips produced in 2026. That's a complete reversal from the era when consumer electronics set the demand curve — and it's a measure of how much the world now wants AI compute.

The pressure runs down the memory and packaging supply chain that feeds the whole industry, not any one chipmaker. Blackwell-generation hardware is extraordinarily memory-dense, which is exactly what makes it so capable: a single NVIDIA B300 packs eight HBM stacks — 96 DRAM dies — so a fully configured 8-GPU system carries 768 DRAM dies for HBM alone. Multiply appetite like that across every hyperscaler building out at once — many of them locking up DRAM and NAND supply years ahead on long-term contracts — and you can see where the competition for parts comes from.

The pricing tells the same story. DDR5 contract prices have more than doubled, from roughly $7 to $19.50 a unit, and Samsung raised 32GB DDR5 module prices by 60% in a single move. Packaging is tight too: demand for the latest accelerators is running well ahead of what TSMC's CoWoS advanced-packaging lines can turn out, and that capacity is booked out through 2026. Even the consumer tier is feeling the memory squeeze, with RTX 50-series production reportedly trimmed by 30–40% in H1 2026 on GDDR7 shortages — which tells you this is a whole-market dynamic, not a data-centre quirk.

The honest takeaway is one I'd give any customer: constraints from HBM shortages and packaging bottlenecks are expected to persist through at least H1 2027, with meaningful new capacity not landing until late 2026 into 2027–2028. Lead times on data-centre GPUs are running 36–52 weeks, and even the hyperscalers are capacity-constrained. This is precisely the gap we're built to close. As an NVIDIA partner, we secure allocation ahead of demand and pass that access on — so you get current-generation compute in a fraction of the time it would take to source it yourself. Our NVIDIA B300s land on-demand in August, with reserved private clusters around a similar time. The scarcity is real; the point of working with us is that you don't have to feel it.

3. The real bottleneck now is speed to deployment — and cooling is a first-order problem

Here's the shift that surprised me most. Several vendors weren't leading with chips at all. They were leading with time. Conventional data-centre construction still takes 18–24 months, with power, mechanical, and cooling systems delivered by separate vendors working independently — which is where the coordination gaps, schedule slips, and cost overruns creep in.

The response you could see forming on the floor was modular. KAYTUS, for instance, launched a gigawatt-scale, fully prefabricated, containerised liquid-cooled design split into IT, Power, and Cooling “cubes,” scaling from a 3MW base unit up to 1GW — pitched to cut that 18–24 month cycle down to a matter of months. When everyone is waiting on the same chips, being able to stand capacity up faster becomes its own advantage.

Cooling, meanwhile, has stopped being an afterthought. Multiple exhibitors — KAYTUS, DDN, HPE Cray among them — led with liquid cooling as core to their AI factory pitch, tied directly to the power draw and density of the latest hardware. And sustainability was framed as inseparable from capacity itself; ISC's own programme chair called sustainability the cornerstone of future computing power. Power and cooling efficiency has quietly become a sales angle, not just an engineering footnote.

Here's why I find that reassuring rather than daunting: it's the part we've already solved for our customers. When you run on Hyperstack, you're not the one negotiating 18-month construction schedules or coordinating power, mechanical, and cooling vendors — you consume purpose-built, liquid-cooled, energy-efficient capacity. We've absorbed the physical build so you don't inherit the timeline, the coordination risk, or the sustainability overhead. And our data centres in Europe and Canada run on 100% renewable energy. For teams that would otherwise be staring down a two-year facilities project, that's often the difference between shipping this year and shipping in 2028.

The quieter theme: getting more useful work per GPU

Running underneath all of this was a shift I found genuinely encouraging. When hardware is scarce and expensive, the bragging rights move from who has the most GPUs to who gets the most useful work out of each one. Several vendors were pitching data-pipeline and KV-cache acceleration specifically to raise GPU utilisation and lower cost per token. For a provider like us, that's a far more honest value story than raw capacity numbers — and it's one we already care a lot about.

Where that leaves us

Put the three together — a dependency risk that can be triggered overnight, a hardware supply chain that's stretched for structural rather than cyclical reasons, and a deployment timeline that's become the real constraint — and a pattern emerges. Each of these problems has a concrete answer, and it happens to be what we've built.

If your concern is jurisdictional control, that's Hyperstack Secure Private Cloud: dedicated, single-tenant, EU- or Canada-sited infrastructure under EU law, with AI Studio on top so your models are open-weight and yours to run. If your concern is getting current-generation NVIDIA hardware at all, that's our secured allocation — NVIDIA B300s on-demand from August and reserved clusters too — so you skip the 36–52 week queue. And if your concern is the physical build, that's our purpose-built, liquid-cooled capacity, available on demand instead of as a two-year construction project. Three problems, three answers, one platform.

If any of this is on your roadmap for the next year or two, I'd rather have a straight conversation than send you a brochure. Book a cloud consultation and ask for Olly Brooks and I'd happily chat.