A map of Saudi AI infrastructure: chips, cloud, data centres and the power underneath

Saudi AI infrastructure is announced in fragments, one partnership at a time, which makes the shape of it hard to see. Laid out by layer, the map is clearer and so are the gaps. The silicon layer has two tracks running in parallel. AMD, Cisco and HUMAIN announced on the opening day of LEAP 2026 that AMD Instinct MI355X systems on Cisco networking had gone live and were serving HUMAIN customers, with a reported plan for 13,000 AMD GPUs this year, up to 250 megawatts of additional infrastructure from 2027, and as much as one gigawatt by 2030. Separately, HUMAIN launched an AI cloud on NVIDIA HGX B300 Blackwell Ultra systems, under a partnership announced in May 2025 targeting up to 500 megawatts powered by several hundred thousand NVIDIA GPUs over five years, with a first phase described as an 18,000-GPU GB300 supercomputer. Aramco has separately named NVIDIA, Groq and Qualcomm as collaborators on industrial and edge AI hardware. The cloud layer is where the international providers sit. AWS confirmed its first Saudi region for December 2026, supported by more than $5.3 billion, with an expanded HUMAIN partnership making up to 50 megawatts available in the country's first AI Zone by 2028, and has said it will make HUMAIN's ALLaM model available through Amazon Bedrock. Microsoft has a Saudi region and is bringing ALLaM to its own platforms. HUMAIN's own AI cloud is live with 1.1 megawatts and reported utilisation above 90 percent, targeting 35.1 megawatts by 2027. The data centre layer is the most crowded and the most varied in scale. HUMAIN and DataVolt have begun construction on 100 megawatts of a 360 megawatt first phase at Oxagon in NEOM, with the first 100 expected available in 2028. Al Moammar Information Systems announced a $1.2 billion programme to reach 192 megawatts of operational capacity. NHC Innovation is putting $880 million into data centres in the Khuzam Digital Valley targeting 65 megawatts by 2033. At the smaller end, Axelerated Solutions and BIM Capital are raising a SAR675 million fund, subject to Capital Market Authority clearance, for two Tier III sites starting at 10 megawatts each. The power layer is the one that decides all of it. At LEAP 2026 the energy sector signed three agreements to support data centre and AI power needs, the first between National Grid SA and HUMAIN covering electricity supply for AI data centres in Riyadh and the electrical infrastructure behind them. The Ministry of Investment and the Shareek programme signed agreements with Saudi Energy and Dawiyat around a Tier III facility. Three things become visible once the map is assembled. The first is concentration. HUMAIN is in almost every layer: operating both accelerator estates, partnering with AWS on the AI Zone, partnering with DataVolt on data centres, holding the grid agreement, and publishing its own models. This is a deliberate national champion structure and it has genuine advantages, principally speed and the ability to coordinate across layers that would otherwise negotiate with each other. It also means Saudi AI capability is substantially dependent on the execution of a single organisation, which is a different risk profile from a market with several competing domestic operators. The second is that the layers are not equally advanced. Silicon and cloud have live deployments. Data centres are mostly under construction or in financing. Power is at the agreement stage. That ordering is backwards from the dependency chain, since the power has to be there before the megawatts can be energised, and it is the normal way infrastructure gets announced rather than a warning sign. It does mean the dates in the later layers are the binding ones. The third is the gap between announced and live. HUMAIN's NVIDIA cloud reported 1.1 megawatts operational against a programme targeting up to 500. The AMD estate went live without a published megawatt figure. Everything else in the data centre layer carries a date between 2027 and 2033. Announced Saudi AI capacity is now measured in hundreds of megawatts with a gigawatt-scale target; disclosed live capacity is in single-digit megawatts plus an unquantified AMD deployment. What the map does not contain is as informative. There is no domestic chip design or fabrication of any kind, which is unsurprising and is a genuine dependency: every accelerator in every layer is designed and manufactured abroad under export control regimes that are outside Saudi influence. There is little visible domestic activity in the middle layer between infrastructure and applications, the model tooling, evaluation, observability and deployment platforms, with OmniOps and MOZN among the few exceptions. And there is no published national figure for total live AI compute capacity, which means nobody outside the operators can size what actually exists. The single most useful disclosure the Kingdom's AI programme could make is that last number: total energised AI capacity, updated periodically. Every oth