Notes from LEAP 2026: The One Thing You Can't Mandate

I spent this week watching the announcements come out of LEAP. Fifteen billion dollars in commitments. HUMAIN launched a new Arabic model trained on over a trillion tokens. AMD brought its largest inference cluster outside the US online. xAI is scaling toward 500 megawatts. AWS confirmed its Saudi cloud region goes live in December. There were dancing robots, robotaxis, and an F1 simulator.

It's a serious show, and I don't want to take anything away from it. The capital is real. The partnerships are real. And this isn't just empty data centres waiting for someone to show up. HUMAIN is clearly pushing on the demand side too: a coding platform to grow local developers, an operating system that pushes AI into real workflows, a manufacturing deal to take AI out of the data centre and into physical, on-site use. They understand that buying compute is only half the job. Using it is the other half.

What struck me most wasn't a product. It was something Tareq Amin, HUMAIN's CEO, wrote going into the event. He said, more or less, that respect doesn't come from funding announcements or valuations or headlines. It comes from execution, from real products, and from delivering real value at scale. He said this chapter would be about results, not ambition.

I think that's exactly right, and I want to take him at his word, because it sets the correct bar. So the interesting question stops being how much was announced, and becomes the harder one he raised himself: how do you actually prove real value at scale?

Once you ask it seriously, you realise there are two different value questions here, not one. They sit at different layers, and they fail in different ways.

The first is at the top, the applications and tools. Here a state-led programme has an advantage no normal company has: it can create its own demand. It can put a platform into every ministry, make adoption a procurement requirement, roll something out across the public sector by decision rather than persuasion. That moves faster than any market. But it's also a trap, because it makes one number very easy to produce: usage. Mandate a tool and people will use it. Seats fill. Dashboards look busy. And usage is the easiest thing in the world to mistake for value.

I've seen this from the other side of the table for a long time. A large organisation buys a platform, adoption gets pushed from the top, and on paper it's a success, everyone's on it. Then you ask the quiet question, the only one that ever mattered, and the room goes still. What did it actually change? Did anything get safer, cheaper, faster, less risky? And can you show me, with a number that isn't just how many people logged in?

But that's the smaller question, because that's not where most of the money is.

Most of the money is at the bottom layer: the chips, the clusters, the gigawatts, the data centre capital. And down there the value test is completely different, and harder. A GPU cluster is only worth what it earns. It has a brutal clock on it, it depreciates fast, it burns power and cost whether it's busy or idle, and it goes obsolete in a couple of years when the next generation of silicon lands. So the question at the infrastructure layer isn't "is it being used." It's "is it loaded with paid, productive work, at high utilisation, from real demand, fast enough to justify the capital before the hardware becomes a liability."

And here's the sovereign trap, the bigger cousin of the mandate trap up top. A state can guarantee the supply of compute by writing a cheque. It cannot as easily guarantee the demand to fill it. If the developers and the real enterprise workloads don't arrive at the pace of the buildout, and Saudi's own numbers, a roughly 50% AI hiring gap, a target of 20,000 specialists it hasn't hit yet, suggest demand is the actual constraint, then you don't get a strategic asset. You get expensive capacity sitting warm and half-empty. Utilised sovereign compute is a national advantage. Underutilised sovereign compute is a stranded one. The only thing separating those two is, again, honest measurement, just a different measurement than the one at the top.

Now, in fairness, not all of this capital is trying to earn a market return, and it would be naive to pretend otherwise. Some of it is buying sovereignty, the ability to run your own AI without depending on US or Chinese infrastructure, control over where your data sits. That's a legitimate reason to build capacity that a pure spreadsheet would call idle. Strategic autonomy has a value that doesn't show up as revenue per GPU.

But that's exactly where I'd watch carefully, because sovereignty is also the perfect place to hide a demand shortfall. Every under-loaded cluster can be re-labelled a sovereign asset. Every workload that didn't materialise can be waved off as strategic patience. At some point someone has to be honest about which capacity is genuinely serving national autonomy, and which is just waiting for customers who were slower to arrive than the press release promised.

So I'd hold the whole thing to the standard Amin set himself. Not megawatts installed. Not seats deployed. Not models launched. But value delivered, adoption that changed how the work is actually done, and infrastructure that's earning its keep, measured plainly, including where it hasn't worked yet.

If the next LEAP is full of that, real utilisation, real outcomes, wins and misses shown honestly, then the story really will have moved from ambition to results. That's the one I'll be watching for.