If you aren’t familiar with the tech conference LEAP, you aren’t alone. Founded in 2022, this Saudi Arabia-based event draws roughly 200,000 attendees, 1,800-plus vendors, and over 1,900 investors, yet has received little coverage in the US. At this size, it’s fair to call it one of the premier global tech conferences, with a heavy focus on AI. Given the AI and cloud activity coming out of the Middle East, that tracks.

HUMAIN, the company that Saudi Arabia’s Public Investment Fund launched in May 2025, has been pretty open about what to expect. Its pre-LEAP messaging frames the event as the reveal of a full AI stack, built in Saudi Arabia and built for the world. Saudi Arabia, by the way, has designated 2026 its Year of AI. And AMD, AWS, Cisco, HPE, and Cerebras all hold main-stage slots.

The interesting question for me isn’t whether HUMAIN can build datacenters. It can, and it is. The question is what happens when the forces reshaping enterprise AI collide. Two of them point straight at Saudi Arabia. The third points away from it.

Take them in order, then the company.

First: Inference Is the Workload Now

Training was the story for three years. Inference is the story now. And if you want evidence, skip the forecasts. Look at how the hyperscalers describe their own spending.

In its fiscal Q1 2026, Microsoft disclosed that roughly half its capital expenditure went to short-lived assets, mostly GPUs and CPUs. By Q4, that share was two-thirds, on quarterly CapEx of $41 billion, itself up 70% year over year. Short-lived assets are chips, not shells.

The industry-wide picture matches. Deloitte’s 2026 TMT Predictions put inference at roughly two-thirds of all AI compute this year, up from a third in 2023, and expect most of it to run in large datacenters rather than at the edge. At gigawatt scale, that second point matters more than the first.

Here’s why I think that matters. Training is lumpy and capital-intensive. Inference is continuous, OpEx-driven, and latency-sensitive. Training gets budgeted once. Inference appears on the bill each month, and it grows with adoption. So enterprises start buying compute the way they buy electricity, and cost per token becomes a line item in the procurement model.

Second: Power Redraws the Map

For the last few years, the gating question has been who could get GPUs. While that supply chain challenge continues, a new pressing question has arisen. Who can power them? And soon it’ll be who can get the transformers and substations to deliver that power.

Compute follows electrons. And, electrons don’t move well.

But workloads do.

This can break the link between where a company sits and where its AI actually runs. For a batch job, the answer is wherever power is cheapest. For an interactive workload, it’s the cheapest power that still meets latency requirements. And those requirements are looser than most CIOs I talk to assume.

Western Europe is the clearest casualty of that shift. And what’s at risk are the workloads themselves. The demand is there, and so is the capital. What’s missing is spare generation and grid capacity on the timeline enterprise AI is arriving. Regulation compounds the problem rather than solving it. New capacity is measured in permitting cycles and regulatory reviews – perhaps years before ground is broken.

The workloads won’t wait for that.

Third: Sovereignty Pulls Workloads Home

Sovereignty stopped being a talking point and started being written into law. In June, the European Commission adopted its proposal for the Cloud and AI Development Act, or CADA, the centerpiece of its Tech Sovereignty Package.

So, what does CADA actually do? Two things. It pushes to triple European datacenter capacity within five to seven years, and it gives public bodies a way to buy compute without handing control to providers who answer to foreign law. The second part is what matters here.

The mechanism is a rating system. CADA scores providers based on where the infrastructure is located, who owns it, who operates it, and who staffs it. These scores then sort providers into four assurance levels – each harder to reach than the last. The Commission expects most contracts to clear only the lower ones: roughly 20% at Level 2, under 10% at Level 3, and about 1% at Level 4.

That’s the shift, and I don’t think enough people have registered it. Sovereignty is no longer a checkbox asking where data is stored. It’s a graded architectural test; every one of those criteria rewards keeping the workload close to home. A provider can satisfy data residency while still failing on ownership, operations, or staffing.

CADA is still a proposal, not law, so treat it as directional. But the direction is clear, and Europe is HUMAIN’s nearest big market.

And buyers are saying the same thing. Deloitte’s State of AI in the Enterprise 2026, which surveyed 3,235 leaders across 24 countries, found that 77% now weigh an AI solution’s country of origin when selecting vendors. They’re acting on it, too. Broadcom’s Private Cloud Outlook 2026 found that production inference running primarily on public cloud fell from 56% to 41% in a year.

Where HUMAIN Actually Sits

I see HUMAIN strong in two areas, challenged in one, and aspirational in two more.

Strong: power and land. This is the foundation, and it’s real. Saudi Arabia can offer grid capacity at a scale and price that European and most North American markets can’t approach on this timeline, alongside land, water, and connectivity. Capital buys chips and buildings. It can’t conjure a grid that doesn’t exist.

That’s HUMAIN’s most durable asset.

Strong: position relative to Europe. Geography is the second asset, and it’s underrated. Riyadh to Milan is roughly 4,060 km, a round trip of 55 to 60 milliseconds on real fiber routes. Saudi datacenter developer DataVolt puts this as half the world’s population within 120 milliseconds. That’s inside the envelope for most interactive inference, and it’s available where there are spare gigawatts. Europe has the demand but lacks the power. Saudi Arabia has both the power and the proximity.

I think that mismatch may be the most under-told part of the HUMAIN story.

Challenged: silicon supply and time to energized megawatt. The partnerships are there. NVIDIA, AMD, Qualcomm, Cisco, and Groq are all in, and an AMD-Cisco joint venture targets up to 1 GW by 2030. What HUMAIN shares with everyone else in the market is the physical bottleneck: everything from the chip to the substation, and the trades to install it. Allocated grid capacity and energized, chip-populated, contracted capacity are very different numbers.

The gap is where the execution risk lives.

Aspirational: the agentic platform. A data platform, voice agents, an enterprise agent suite, and an AI-native ERP are a coherent set of ambitions, and the logic holds: agentic systems shift what an enterprise buys from seats toward outcomes, and whoever supplies the infrastructure is placed to supply what runs on it, too. I like the aggression here. Going full-stack this early is the right instinct for a company built to be a national champion.

It’s also a tall order. Agentic AI is still an emerging market. But the competition isn’t. HUMAIN will be up against companies that have owned these enterprise workflows for decades. Their incumbency isn’t in agents. It’s in the customer relationship.

This is the part of HUMAIN’s strategic plan I’d expect to take the longest to achieve.

Aspirational: end-to-end transformation. The forward-deployed engineering instinct is right, and realistically, the only way agentic products reach production. Scaling through the big consulting firms is pragmatic, but there’s a tension worth naming. Part of HUMAIN’s mandate is that AI capability accumulates inside the Kingdom, and deployment work is where that capability gets built. If outside firms deliver most of it, the expertise accrues to them instead.

How HUMAIN staffs this work determines whether that capability stays in the Kingdom.

Saudi Arabia’s Advantage, or HUMAIN’s?

So this raises the question beneath it all. If cheap power, proximity to Europe, and a sovereignty framework are the advantages, why can’t AWS, Microsoft, Google, or Oracle build in Saudi Arabia and capture them? Partly they can, and partly they already are. HUMAIN’s partnerships with hyperscalers suggest it isn’t trying to stop them in any way.

What HUMAIN brings isn’t a bigger balance sheet or better technology. It’s alignment with the national strategy. Grid capacity is allocated through national planning, not bought on a spot market. The data embassy law, if it arrives, is national policy. HUMAIN was purpose-built as the vehicle for those decisions.

If this sounds unusual, it shouldn’t. National champions from Airbus to Samsung to Aramco were built the same way, and several became formidable competitors.

That advantage is strongest within a single jurisdiction. Saudi Arabia’s position relative to Germany is durable. So is HUMAIN’s relative to a hyperscaler operating in Saudi Arabia. Outside the Kingdom, it competes on conventional terms, where scale and distribution matter more, and a 15-month-old company faces incumbents with twenty years of relationships.

That’s a big ambition.

None of which means HUMAIN only builds at home. It partnered with New York-based Global AI in late 2025 to build datacenter and compute capacity in the United States, and it’ll put capacity wherever customers need it. But capacity built elsewhere is just capacity. The power economics don’t travel with it. Neither does the alignment.

So the anchor business is being the preferred vehicle for compute built in Saudi Arabia, which is substantial on its own. Enterprises are moving toward compute portfolios anyway, so nobody has to displace AWS to matter. The global business is the harder bet, and it comes down to the legal question below.

On the Label “Neocloud”

Is HUMAIN a neocloud? I think the label’s wrong, though maybe for different reasons than others.

A neocloud is a bet on scarcity. Rent GPUs into a shortage, capture the spread, and hope the shortage outlasts the depreciation schedule. It’s cyclical arbitrage, and can be a great trade for 36 months and a challenging one after.

HUMAIN is the opposite trade. It’s a bet on abundance: power, land, and investment that doesn’t require a quick return, expressed as a cloud company. The risk is industrial and political rather than cyclical, and the horizon is decades. Neoclouds are challenged when GPU supply loosens. HUMAIN fails if the power advantage never converts into contracted, energized capacity, or if the sovereignty mechanism doesn’t hold.

So the honest framing isn’t “more than a neocloud.” Neocloud describes one of HUMAIN’s business areas – probably the one closest to working today. The full-stack claim spans infrastructure, models, agentic products, and transformation services. My read is that infrastructure is a strategy, and the other three are optionality.

This is perfectly reasonable for a company that’s 15 months old.

The Crux

Here’s the bet. The third force described earlier (sovereignty) pulls workloads home. But HUMAIN’s global thesis needs sovereignty to become portable instead. It needs a German bank or a US health system to accept that a workload in Riyadh is, in the legally relevant ways, operating under their own law.

And the mechanism is the data embassy: a facility physically in Saudi Arabia but operating under the customer’s home jurisdiction. Estonia and Luxembourg set the precedent in 2017. And Saudi Arabia’s draft Global AI Hub Law, out for consultation, would be the first comprehensive G20 framework of its kind.

But there’s a timing challenge. A facility in Riyadh can’t reach CADA’s upper tiers, regardless of what the legal wrapper says. HUMAIN needs sovereignty to be portable at the exact moment its nearest large market is defining it as local.

And of course, there’s a live counter-signal. Mistral and HUMAIN announced a collaboration worth hundreds of millions of euros, with Mistral saying it will explore using HUMAIN’s datacenters to support growing local compute needs.

Think about who that is. Mistral is the company Europe points to when it talks about not being overly dependent (or dependent at all) on American AI. And it just put real money behind Saudi infrastructure—specifically, HUMAIN.

The first phase is regional, serving the Kingdom and its neighbors rather than Rotterdam. But this is a starting point, not a ceiling. What matters here is that a Western frontier lab looked at HUMAIN and decided this is a place where it should build.

To me, this is the strongest evidence so far that HUMAIN’s strategy can work.

What I’m Watching at LEAP

Let’s keep this simple. I’m tracking six things related to HUMAIN:

1. Does the cost advantage get quantified? Cheap power is the whole thesis, and TCO is why a multinational would even look. A number for cost per token, with a method behind it, makes it real.
2. Where does the Global AI Hub Law stand? An enactment timeline, or a first bilateral agreement, would tell me a lot about how quickly the sovereignty story can move.
3. How many megawatts are energized? Announced capacity makes headlines. Energized, chip-populated capacity shows how the build is actually tracking.
4. What does contracted demand look like? Named customers and terms would be great. Even directional color would indicate how much of the buildout is already spoken for.
5. How far along is the agentic portfolio? Evidence of real deployments matters more than a polished demo. But as I noted previously, I see this as a longer-horizon aspiration.
6. Does the Mistral partnership get more specific? Named workloads and committed capacity would be a strong signal. Anything served from Riyadh beyond the region would be a bigger one.

The HUMAIN infrastructure story is credible today, and it’s a bigger business than most people outside the region realize.

There’s also more to like here than concrete and megawatts. The backing is patient in a way publicly traded competitors can’t easily match, which counts for a lot when payback runs a decade. The vision holds together.

I don’t see HUMAIN as a company chasing whatever’s hot this quarter. It’s far more strategic in playing the long game. The silicon strategy is deliberately agnostic across NVIDIA, AMD, Qualcomm, and Groq, smarter than betting a national program on one roadmap. And the willingness to partner on what it can’t yet build suggests a company that knows exactly what stage it’s at. Fifteen months in, HUMAIN has moved faster than most companies with twice the history.

The full-stack story is still a promise, and the global story still depends on a law that hasn’t been enacted. But the assets are real, and the direction is right. LEAP is where we find out how much of that promise HUMAIN can start converting.