Saudi Arabia’s HUMAIN Signed the Deals. The Switchgear Decides What Runs

Inside eighteen months of announcements, the 467 megawatts actually live against a 3.4 gigawatt plan, and the grid equipment that sets the real schedule

AsiaPresswire – At LEAP 2026, which ran in Riyadh from 31 August to 3 September, HUMAIN disclosed project agreements worth more than US$15 billion. The list ran the length of the stack: AMD for chips, AWS for a cloud region, Microsoft for a distribution channel for an Arabic model, Qualcomm for an end device, Adobe as the first foreign tenant on locally hosted compute, Together AI for 250 megawatts, Applied Intuition for autonomous trucks, SLB for oilfield data. A single company, eighteen months old in its current form, signed across power, silicon, models and applications in four days.

That is the action. Everything interesting about Saudi Arabia’s artificial intelligence programme follows from what happened next, which is the part the announcements do not cover.

The country had 68 megawatts of data centre capacity in 2021. By the first quarter of 2026 it had more than 467 megawatts, according to the Saudi Press Agency. A report from the commercial office in Jeddah put the count at 60 data centres by the second quarter of 2026, with planned capacity of 3.4 gigawatts. Read those two numbers together and the shape of the problem becomes visible. Roughly 467 megawatts exist. Roughly 3.4 gigawatts are planned. The ratio between them is the story, and it is a story about transformers, switchgear and cooling water rather than about capital.

The Action, and the Size of It

HUMAIN is owned entirely by the Public Investment Fund, the sovereign fund usually described as holding around US$930 billion, and it sits inside the Crown Prince’s office rather than under a ministry. Its chief executive is Tareq Amin, who ran Rakuten Symphony and before that built a multi-vendor telecom cloud platform at scale. The hire is a statement about what the company is for. It is an operator, not a research lab.

The pace it has set is unusual even by Gulf standards. Eleven data centres, capacity coming online at roughly 50 megawatts a quarter, an AWS region promised for December 2026 and a Microsoft region for November 2026, and what the company describes as the first xAI facility outside the United States at 50 megawatts scaling to 500. Amin’s own formulation, quoted in regional coverage, is that the objective is to build in 2026 the equivalent of what Saudi Arabia built in twenty years.

The ambition is not the interesting part. The interesting part is that HUMAIN chose vertical integration when the obvious alternative was available. The comparison that regional analysts keep drawing is with G42 in the United Arab Emirates, which restructured around Microsoft’s equity position, aligned itself with the OpenAI and SoftBank Stargate programme, and leases a substantial share of the capacity it sells. HUMAIN owns the data centres, contracts the chips, hosts the models and runs an Arabic consumer product. That is a harder business to execute and an easier one to point at as sovereign.

Fourteen Months, Ten Signatures

The behaviour sequence is where a profile earns its keep. HUMAIN did not announce a strategy and then execute it. It signed in a specific order, and the order reveals what it thought it lacked at each stage.

Read down the first column and a pattern appears. The first four moves are about securing power, capital and a partner with an existing grid footprint. The fifth and sixth are about turning commitments into assets. Only then does the company go shopping for silicon and applications.

That sequence is not the sequence a software company would choose. It is the sequence a utility would choose.

Why Build the Whole Stack Instead of Renting It

The case for vertical integration is a case about who captures the margin when compute becomes a tradable commodity, and it rests on three claims that HUMAIN makes explicitly.

The first is that sovereign buyers will pay for provenance. A Saudi ministry procuring an Arabic model does not need it to be the best Arabic model in the world. It needs it to be the best Arabic model it can buy on infrastructure inside the country, with evaluation evidence and a documented human-in-the-loop layer. HUMAIN inherited the ALLaM family from SDAIA and has been extending it: broader dialect coverage, a Modern Standard Arabic alignment pass, a Gulf-specific retrieval layer, and a clinical evaluation track aimed at health procurement.

The second is that the model layer and the compute layer reinforce each other when both are owned. HUMAIN unveiled a 428 billion parameter Arabic model built with China’s MiniMax, trained on more than a trillion Arabic tokens and released as a research preview through its own platform. A model that runs on someone else’s cloud is a customer. A model that runs on your own cloud is a product you can bundle.

The third is distribution. The Microsoft arrangement puts ALLaM inside Microsoft 365 Copilot and on the Microsoft Foundry platform, which gives the model a channel it could not build. The AWS arrangement puts ALLaM on Amazon Bedrock and HUMAIN’s compute fabric on the AWS Marketplace. In both cases the sovereign operator trades exclusivity for reach.

There is a cost to that choice, and it is concentration. A company that owns the data centres, contracts the chips, trains the models and sells the applications has four execution risks correlated with each other. If the grid slips, the chips sit in crates, the models have nowhere to train and the applications have nothing to run on.

The Gap Between Signed and Switched On

The gap is quantified in public. Saudi Arabia reported 467 megawatts of operational capacity in the first quarter of 2026 against a planned 3.4 gigawatts. HUMAIN’s own disclosed footprint targets 500 megawatts of AI factory capacity across three clusters, one in Riyadh on PIF land, one at NEOM where the giga-project has already built grid capacity and water-cooling assumptions, and one in the Eastern Province close to Aramco’s industrial backbone.

The arithmetic is unforgiving. To reach 3.4 gigawatts from 467 megawatts means adding roughly 2.9 gigawatts, which at 50 megawatts a quarter is 58 quarters, or more than fourteen years. Even if the quarterly run rate triples, the plan is a decade-long build. The 2030 date that appears in most Saudi AI commentary therefore describes a much smaller number than the headline plan implies, and the honest reading is that the plan is a ceiling rather than a forecast.

There is a second way to read the same gap, and it is more generous. A gigawatt of planned capacity is not a promise to build a gigawatt. It is a statement that land, grid applications and environmental approvals have been assembled for that quantity, which is a different asset from a building with servers in it. On that reading the 3.4 gigawatts is a pipeline of options, and the 467 megawatts is the part that has been converted. Both readings support the same conclusion about timing: what exists is a small fraction of what is discussed, and the rate of conversion is the variable that decides whether the Kingdom becomes a compute exporter or a compute importer with good marketing.

The announced pipeline also has an internal consistency problem that is rarely stated. HUMAIN’s 500 megawatt footprint, the 1 gigawatt stc joint venture, the 1.5 gigawatt DataVolt campus at Oxagon and the 500 megawatt xAI commitment are not additive in any simple sense, because several of them describe the same physical capacity under different commercial arrangements. A joint venture rated to 1 gigawatt does not create 1 gigawatt on top of a 500 megawatt target. It describes who owns the meter.

What the Structure Cannot Do

Four constraints do the real work of setting the timetable, and none of them is capital.

The first is high-voltage equipment. Long-lead switchgear and transformer capacity are the operational chokepoint in every Gulf build, and the procurement response has been to buy them the way the company buys GPUs: on multi-year commitments rather than spot orders. That solves availability for the buyer who moves first and makes the constraint worse for everyone else.

The second is cooling water. A 500 megawatt campus consumes water at a scale that has to be contracted before construction, not after. NEOM’s advantage is that the giga-project already modelled water for The Line and can extend those assumptions. The Riyadh and Dammam sites have to negotiate their own.

The third is people. The SAMAI programme targets 100,000 AI specialists by 2030. The global pool of engineers who have actually operated a 100,000 GPU cluster is small, and HUMAIN is competing for it against US hyperscalers and against G42. Amin’s hiring has tilted deliberately toward operations and infrastructure engineers rather than model researchers, which is the right tilt for a company whose job is to run AI factories rather than invent them.

The fourth is silicon access. The November 2025 export framework and a reported approval for 35,000 advanced accelerators against a pipeline of 600,000 define how fast the compute can arrive. A framework is not a shipment. The gap between the two numbers is the gap between the political relationship and the physical build.

The Man Who Runs It

Amin’s public statements are worth reading as a set because they are unusually specific about what he thinks the company is not.

He has described Saudi Arabia as positioning itself as the “third pillar” of global AI infrastructure alongside the United States and China, and said the commitment to developing compute capacity in the US secures HUMAIN “a strategic foothold”. He has also said that “Saudi Arabia will be the first country outside the US and China that will commercialise its own operating system”, which is a claim about software, not about megawatts, and it is the more falsifiable of the two. An operating system layer above raw compute is exactly what a national champion would want and exactly what is hardest to ship.

The operating claim that matters most is the one about speed: building in 2026 what the country built in twenty years. That is measurable. The measure is the gigawatt count of operational, not announced, capacity, and the second measure is whether the announced offtake materialises. HUMAIN says its NVIDIA Blackwell Ultra cloud is already running government and enterprise workloads with utilisation above 90% and the Ministry of Commerce among the first disclosed users. If that utilisation figure holds as capacity triples, the demand thesis is real. If it falls, the build was a real estate programme with a press release attached.

The Money, and Where It Comes From

The capital stack is worth laying out because it explains why the build has not slowed despite the absence of a commercial customer base large enough to justify it.

Two things stand out. The first is that most of this is not project finance in the ordinary sense. A conventional data centre developer raises against contracted offtake, and the lenders underwrite the tenants. Here the sovereign underwrites the developer, and the tenants include ministries. The credit being priced is closer to a sovereign than to a colocation operator, which is precisely how the Kingdom intends to compete on cost of capital against US hyperscale builders.

The second is the number of instruments. Equity, a national infrastructure fund, a telecom joint venture, a stake in a model company, a third-party infrastructure fund and a venture arm are six different ways of moving money into the same build. That diversification is a hedge against any single channel stalling, and it is also a fragmentation risk: six instruments mean six sets of terms, six reporting lines and six sets of people who believe they own the schedule.

The spending commitment attached to the venture arm has a condition that is easy to miss. Investments are tied to portfolio companies using Saudi compute or bringing staff into the Kingdom. That is a sovereign fund using its chequebook to manufacture demand for the asset it is simultaneously building. It is a coherent strategy and an unusual one: most venture programmes invest to make money, this one invests to fill a data centre.

What the Neighbours Are Doing

The Saudi build is easier to judge against the region than in isolation, because the region is running the same play at different speeds with different degrees of vertical integration.

The clearest contrast is with the United Arab Emirates. G42 restructured around Microsoft’s equity position and aligned itself with the Stargate programme, leasing a significant share of the capacity it sells. That trades autonomy for velocity. The Saudi model trades simplicity for control. The capital comparison favours the UAE on distributed commitments, with Stargate sized at US$500 billion globally and a US$1.4 trillion pledge of investment in the United States over a decade, while Saudi capital is concentrated domestically, at roughly US$77 billion through HUMAIN.

The second contrast is less obvious and more instructive. Thailand’s Board of Investment reported investment applications breaking 1 trillion baht in the first quarter of 2026, with the digital industry at 873.7 billion baht and data centres accounting for close to nine tenths of that. Thailand is not building a national AI champion. It is processing applications from companies that want to build, using an existing industrial base, an existing grid and an existing regulatory pathway.

That difference defines the wager. Saudi Arabia is attempting to build the operator and the infrastructure at the same time, which is why its announcements are larger and its operational megawatts smaller. Thailand is letting other people’s operators build on infrastructure it already has, which is why its application numbers look strong and its sovereign AI story does not exist. Both approaches can work. Only one of them produces a company called HUMAIN.

What Would Prove the Thesis

Three indicators would settle the argument within eighteen months, and all three are public.

• Operational megawatts, not announced megawatts. The company discloses capacity additions quarterly. The number that matters is the one that appears in grid interconnection records, not in conference decks.
• Utilisation on the live cloud. A figure above 90% on a small base is easy. The test is whether it survives a fivefold increase in supply.
• Arabic workload routing. ALLaM inside Microsoft 365 Copilot and on Amazon Bedrock will show, in usage data, whether sovereign Arabic models win on Saudi infrastructure or become a compliance checkbox in a catalogue of American and Chinese alternatives.

There is a fourth indicator that is harder to read but more decisive. Saudi Arabia’s advantage in this race is land, a domestic market of 33 million people roughly three times the size of the UAE’s, and cheaper electricity. Its disadvantage is that hyperscaler relationships and regulatory speed were built elsewhere first. If EMEA and South Asia inference traffic routes through the Gulf, it will route through one hub predominantly. The hub that delivers lower latency, more reliable capacity and better commercial terms takes the traffic, and the terms are set by whoever has spare megawatts first.

That is why the switchgear matters more than the signing ceremony. A company can announce US$15 billion of agreements in four days. It cannot compress the lead time on a transformer.

Data basis

Capacity figures are from the Saudi Press Agency report of August 2026 and from the commercial office report of June 2026, which give 467 megawatts operational and a 3.4 gigawatt national plan respectively; the 2021 baseline of 68 megawatts is from the same SPA report. Deal values and capacities are from the LEAP 2026 summary published by the Chinese Ministry of Commerce’s Saudi office, cross-read against trade coverage of the same announcements. The 50 megawatts per quarter run rate, the 500 megawatt target footprint and the 1 gigawatt joint venture rating are company statements reported in regional and industry press. The financing framework of up to US$1.2 billion is non-binding as described in April 2026 coverage, and is treated here as a framework rather than as funds deployed. Estimates of the time required to reach 3.4 gigawatts are arithmetic performed on the published figures, not company guidance.

Limitations of the Record

Three limits apply. Announced capacity and operational capacity are reported by different parties on different timetables, so the gap described here is indicative rather than audited. No HUMAIN executive was interviewed; quotations are from published remarks. And the utilisation figure of above 90% is a company claim without a published measurement method.

Sources

• Saudi Press Agency — Saudi data centre capacity from 68 MW in 2021 to more than 467 MW in Q1 2026, August 2026
• MOFCOM Saudi office — Overview of LEAP 2026 outcomes, September 2026
• FWDstart — The AI announcements from LEAP 2026 worth knowing, September 2026
• AI in Arabia — HUMAIN in 2026: anatomy of Saudi Arabia’s sovereign AI champion, 17 May 2026
• AI in Arabia — HUMAIN lines up US$1.2 billion for 250 MW, 28 April 2026
• Saudi Compute — Sovereign AI in Saudi Arabia: the model versus the UAE model
• Thailand Board of Investment — Investment applications in Q1 2026 pass 1 trillion baht, 29 April 2026

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com