Humain announced plans with the US firm Together AI for 55 megawatts of artificial intelligence capacity and a new data centre capable of scaling to 250 megawatts, alongside a Saudi AI cloud running NVIDIA Blackwell Ultra hardware.
Humain used the opening of LEAP 2026 to set out its next tranche of capacity: 55 megawatts of artificial intelligence capacity with the United States firm Together AI, and a new data centre capable of scaling to 250 megawatts. The company also launched a Saudi artificial intelligence cloud running on NVIDIA Blackwell Ultra hardware.
The two numbers describe the way this market now contracts. Fifty-five megawatts is capacity with a customer attached; 250 megawatts is what the site can grow to. Between them sits the design decision that determines whether the second figure is achievable: whether the electrical infrastructure, the cooling distribution and the structural provision were sized for the end state or for the first tranche.
Building for the end state costs money years before it earns any. Building for the first tranche is cheaper and risks a campus that cannot take the growth it was announced with, because the incoming supply, the riser space and the plant yard were fixed at a smaller number. Operators in this market almost always pay for the larger substation and the larger pipework early, and phase the equipment inside it. That is why capacity announcements pair a contracted figure with a scaling figure rather than giving one number.
The hardware choice matters less to the building than it sounds. Blackwell Ultra parts are dense, and density is a thermal problem before it is a computing one — a rack that draws tens of kilowatts cannot be cooled by moving air across it at any sensible velocity, which is why liquid distribution to the rack has moved from exotic to ordinary in new-build specifications. But accelerators turn over on a cycle of a few years while a data hall lasts decades. The design has to accept a generation of hardware that has not been announced yet, which in practice means oversized coolant loops, spare riser capacity and floor loading margins that look wasteful on day one.
Together AI is a compute platform rather than a hyperscaler, and its presence in the announcement is a signal about the customer mix. The Saudi build-out is not solely aimed at the three or four largest cloud providers. Platform companies that sell access to accelerators, and that need capacity in a region without wanting to hold the property risk, are a distinct tenant class, and one that signs for tranches rather than whole campuses.
For the domestic supply chain the relevant number is neither 55 nor 250 but the equipment list underneath them. A campus of this size takes bulk supply at transmission voltage, medium-voltage switchgear, transformers, uninterruptible power supplies, standby generation, chillers or liquid cooling plant, containment, and a structural frame designed around rack loads. Most of that is procured internationally on lead times that have been lengthening, and a growing share of it is what Saudi manufacturers are now positioning to supply — the reasoning behind Alfanar's $150m commitment to data centre components announced at the same show.
Humain has not disclosed the site, the contractor or the delivery date for the campus, and the 55 megawatts has not been given a commissioning date either.