Riyadh’s computing contracts push major delivery dates to 2028, undermining billion-dollar announcements

Riyadh concluded the “LEAP 2026” conference in Malham, which was postponed several months due to the regional conflict, by recording AI investment commitments close to $15 billion, raising the conference's cumulative total from previous editions to $44.2 billion, a roughly one-third increase in a single week, after the second day alone recorded more than $2.5 billion in data-center and advanced-technology deals. However, a close reading of delivery ledgers and operational capacity reveals a temporal and technical gap between contract signing and the actual hardware entering service.

Hiumain announced the launch of its AI cloud powered by Nvidia Hgx300 systems with an operational capacity of 1.1 megawatts and utilization exceeding 90 percent, and a plan to expand to 35.1 megawatts by 2027.The pledged capacity represents only half a percent of the previously announced commitment.It stipulated operating 18,000 GB 300 accelerators with a capacity of 100 megawatts in both Riyadh and Dammam by the end of the second quarter, a deadline that passed about thirteen weeks ago, and the processors currently available belong to the independent Blackwell Ultra class rather than the high-end Grace Blackwell tier promised to the market.

The project map shows that most of the massive commitments are deferred until 2028 or later. Although construction has already begun at the Oxagon complex jointly operated by DataVolt and Hiumain to build the first phase with a capacity of 100 megawatts out of 360 megawatts, delivery will not start before 2028. AMD and Cisco also agreed to add up to 250 megawatts starting in 2027 as part of a pathway targeting gigawatts by 2030, while Amazon Web Services confirmed the launch of its first cloud region in December 2026 with up to 50 megawatts allocated in its first AI region by 2028, alongside contracts in which Together AI booked 250 megawatts, and XAI secured its first data centre outside the United States with capacity starting at 50 megawatts and expanding to 500 megawatts.

On the model front, Hiumain unveiled the research preview of Hiumain M 3, an advanced Arabic language model developed by the Chinese MiniMax lab in Hangzhou under a direct mandate. It uses an expert-mixture architecture with 428 billion parameters, of which about 23 billion are active per token, and was pre-trained on more than one trillion Arabic tokens, achieving 89.37 percent across seven Arabic benchmarks according to the company’s unaudited self-evaluations. Its weights are expected to be released under the MiniMax community license next month, alongside a demonstration of the Horizon Ultra computer in partnership with Qualcomm in preparation for its corporate rollout on 20 September, with a promise to launch a dedicated operating system in 2027.

These shifts coincide with a marked regional divergence, as the Egyptian Ministry of Communications and Information Technology under Raafat Hindy limited its activity to exploratory talks with Hika Data and an American partnership to build a compute complex without disclosed numbers or timelines, while the Emirati side remains silent on updates to the first phase of the StarGate project with a 200-megawatt capacity, amid slower licensing procedures at the U.S. Office of Industry and Security.

This postponement forces technology leaders and infrastructure teams in the Gulf and the region to reset their expansion plans; relying on locally hosted high-performance compute will not be available at the targeted scale before 2028, meaning continued dependence on global clouds for inference and heavy training, or adaptation to the limited compute capacity currently available, while noting that tech-sovereignty strategies have become effectively reliant on repurposing advanced Chinese model weights rather than building them entirely from scratch locally.