Saudi AI announcements are now denominated in megawatts, not chips

Read the LEAP 2026 announcements by unit rather than by company and the same measure keeps appearing. HUMAIN's new NVIDIA-powered AI cloud is live with 1.1 megawatts and targets 35.1 megawatts by 2027. AMD, Cisco and HUMAIN plan up to 250 megawatts starting in 2027 and as much as a gigawatt by 2030. HUMAIN and DataVolt have begun 100 megawatts of a 360 megawatt phase at Oxagon in NEOM, with the first 100 expected to be available in 2028. Al Moammar Information Systems is spending $1.2 billion to reach 192 megawatts. NHC Innovation is putting $880 million into data centres in the Khuzam Digital Valley targeting 65 megawatts by 2033. AWS says its expanded HUMAIN partnership will make up to 50 megawatts available in the country's first AI Zone by 2028. Two years ago the equivalent list would have been denominated in GPUs. The change in unit is the change in the industry. The clearest evidence is not in the compute announcements at all. It is in who Saudi Energy signed with. At LEAP 2026 the energy sector signed three agreements to support digital infrastructure for data and AI centres, the first between National Grid SA and HUMAIN, covering electricity supply for AI data centres in Riyadh and the electrical infrastructure behind them. When a national grid operator is signing at a technology trade show, the bottleneck has moved. The reason is arithmetic. A rack of current-generation accelerators draws in the tens of kilowatts, an order of magnitude more than the enterprise servers data centres were designed around, and the density is still rising. That changes what a data centre is. Power delivery, not floor space, sets the ceiling. Cooling stops being an afterthought and becomes a design constraint that determines whether the site works at all. And the lead time for the thing that matters most, a grid connection at the required capacity, is measured in years and is not controlled by the company making the announcement. This is why the megawatt figures deserve more trust than the dollar figures. A company can revise a capital number by revising its own plan. A connection date depends on generation capacity, transmission build and the queue, and those are outside any single developer's control. When an operator publishes a date for energised capacity, it is making a claim that a third party has to honour. Saudi Arabia's position on this is better than most markets and not unconstrained. The Kingdom has generation capacity, a build programme and the ability to plan generation and demand together in a way liberalised markets cannot. Land is available at scale and industrial power is comparatively cheap. Against that, the load being added is very large relative to anything the grid has previously absorbed for a single sector, cooling in this climate imposes a real efficiency penalty compared with northern latitudes, and the demand is concentrating geographically around Riyadh, NEOM and a small number of industrial sites rather than spreading across the network. The date structure in the announcements tells you how the queue is being managed. Almost nothing large lands in 2026. The dense cluster is 2027 and 2028, which is roughly the time required to move from signature to energised capacity if the connection work starts now. The gigawatt figure sits at 2030 and the NHC Innovation capacity target runs to 2033. Those are not vague long-term ambitions. They are the honest lead times of electrical infrastructure, stated plainly. This reframes how the week's announcements should be read against each other. The comparison that matters between two AI infrastructure projects is not which has more announced money or a bigger partner. It is which has a signed power agreement, at what capacity, with what energisation date. On that test, the HUMAIN projects are unusually well positioned, because the grid agreement was signed alongside them. Projects announced without a corresponding power arrangement are earlier in the sequence than their dollar figures suggest. It also explains a set of announcements that otherwise look peripheral. Work on cooling, on power efficiency and on where compute physically sits is now core rather than adjacent. One of the LEAP Rocket Fuel winners was an MIT spin-out working on immersion cooling for AI chips, which in a market where power and heat are the constraint is a considerably more central technology than it would have appeared in 2023. There is a competitive dimension across the Gulf. The UAE has been building AI capacity on the same logic, and both markets are selling a similar proposition to international AI companies: sovereign-adjacent capacity, at scale, with power available on terms that are difficult to obtain in Europe or in constrained parts of the United States. The credible differentiator between them will not be announced megawatts, since both can announce. It will be energised megawatts, and the ratio between the two. That ratio is the number worth tracking. Announced capacity ac