Eighteen months ago, HUMAIN did not exist. Today it is the single largest buyer of frontier AI compute outside the United States and China, the anchor tenant of a compute supply chain that stretches from TSMC in Hsinchu to a power substation under construction in Al Kharj, and the legal vehicle through which the Saudi Public Investment Fund intends to turn the Kingdom into what its chief executive calls "the third pole" of global artificial intelligence. This episode examines what HUMAIN has actually ordered, where the GPUs are landing, who is paying for the power, and why the 600,000 number that headlines every press release is both real and deeply misleading.
The Shape Of The Order Book
The headline figure originates from the 19 November 2025 US-Saudi Investment Forum in Washington, where HUMAIN, NVIDIA, xAI, Global AI, and AWS announced what was characterised as an expansion of a strategic partnership first unveiled at the May 2025 Riyadh summit. The precise text commits HUMAIN to deploying up to 600,000 NVIDIA GPUs over three years, including the GB300 platform, across data centres inside Saudi Arabia and an unspecified US footprint tied to the same corporate structure. The 600,000 figure therefore is a ceiling, not a floor, and it bundles Saudi and American sites into a single number for political convenience.
Unpacking the forum announcements produces a clearer picture of the individual commitments. xAI signed to anchor a 500 megawatt or larger facility in Saudi Arabia with an initial cluster of 18,000 GB300 GPUs, extending Elon Musk's Colossus supercluster footprint out of Memphis. AWS committed to deploy and operate up to 150,000 NVIDIA GPUs inside a dedicated "AI Zone" in Riyadh, a sovereign-cloud construct that gives Amazon a ring-fenced Saudi region without surrendering its operating model. Separately, AMD, Cisco, and HUMAIN filed on the same day for a joint venture designed to deliver up to one gigawatt of AI infrastructure by 2030, anchored by AMD Instinct MI450 GPUs and a 100 megawatt first phase scheduled to come online in the second quarter of 2026, most likely at one of the two Riyadh and Dammam sites HUMAIN is currently building.
Qualcomm, the final piece of the hardware stack, signed a parallel agreement covering edge and on-device AI compute, aimed at Saudi government deployments and the inference tier that sits below the GB300 training clusters. The Qualcomm deal is politically significant because it extends the HUMAIN supply chain beyond the NVIDIA and AMD duopoly, but the volumes are immaterial against the frontier-training numbers.
Add the commitments together and the three-year construction pipeline now visible inside HUMAIN is roughly this: 18,000 GB300s under xAI, up to 150,000 GPUs under AWS, 100 megawatts of MI450 under the AMD-Cisco JV, and an unspecified balance of NVIDIA inventory held directly by HUMAIN for its own sovereign workloads. The balance is where the 600,000 ceiling lives, and it is the portion that remains most exposed to US export-control enforcement.
How Riyadh Got The Licences
None of this compute moves without approval from the US Bureau of Industry and Security. The original Biden-era framework placed Saudi Arabia in the middle tier of a three-tier diffusion rule that limited the Kingdom to 1,700 advanced AI chips per year without individual licensing. That framework was rescinded in May 2025, replaced by a case-by-case negotiation model that required dollar-for-dollar US investment matching for each approved export. The immediate effect was to fold chip diplomacy into the broader US-Saudi investment relationship and to make HUMAIN's order book a direct function of PIF's willingness to commit capital inside the United States.
On 20 November 2025, one day after the Washington forum, the US Commerce Department formally approved exports of up to 35,000 advanced NVIDIA chips to HUMAIN and G42 combined, a tranche estimated at roughly one billion dollars at public list prices. The licensing came with what BIS characterised as "rigorous security and reporting requirements", understood in industry briefings to include continuous telemetry access for the US government, restrictions on co-location with Chinese-affiliated tenants, and end-use attestations covering both the Saudi and US portions of HUMAIN's network. A further 565,000 units of the 600,000 ceiling remain subject to future licensing rounds, which is the honest reason HUMAIN describes the programme as running through 2028 rather than 2026.
For readers tracking the sovereignty story, this is the single most important structural fact: HUMAIN does not own its inference stack in the way an oil major owns a refinery. Every GB300 inside Saudi Arabia is licensed out of Santa Clara and reports home. The phrase SDAIA officials use in private is "strategic autonomy within managed dependency", and it applies more sharply to HUMAIN than to any other Saudi technology asset.
The Power Problem
A gigawatt of AI data centre demands roughly the same continuous power draw as the city of Jeddah in off-peak hours. HUMAIN's stated ambition, reiterated by Tareq Amin at the 2026 PIF Private Sector Forum in Riyadh in February, is 6.6 gigawatts of operational capacity within ten years. That figure is what the infrastructure side of HUMAIN must actually deliver, and it sits at the centre of every substantive bottleneck the programme faces.
Three financing and delivery structures are now visible. First, HUMAIN signed a framework agreement with Saudi Arabia's National Infrastructure Fund, known locally as Infra, at the World Economic Forum in Davos in January, providing up to 1.2 billion dollars to develop 250 megawatts of hyperscale AI capacity. Second, HUMAIN awarded Al Moammar Information Systems a December 2025 contract to design and build a dedicated AI data centre with 250 megawatts of initial capacity and a one gigawatt ceiling, subject to demand. Third, HUMAIN entered a joint venture with DataVolt to develop multi-gigawatt AI data centres in parallel with the Riyadh and Dammam builds, a move announced in the same PIF Forum window.
The power itself comes from two directions. Aramco, which contributed Aramco Digital assets into HUMAIN at formation, is supplying combined-cycle gas as a bridge, and Aramco chief executive Amin Nasser argued at COP30 in November 2025 that Saudi Arabia's long-run advantage is a marginal electricity cost of under two US cents per kilowatt hour. The renewable tier is being built around NEOM's solar capacity and the PIF-backed Sudair and Al Shuaibah plants, with Acwa Power serving as the dominant independent developer. The mix on paper is roughly 50-50 gas and renewables by 2030, but the 2026 and 2027 delivery windows are overwhelmingly gas-weighted because the solar plants have longer lead times than the GB300 deliveries they are supposed to power.
Cooling is the less-discussed constraint. The Riyadh and Dammam sites use hybrid air-liquid systems designed by the Cisco and HUMAIN engineering teams, but Saudi ambient temperatures make the delta relative to a Nordic hyperscaler facility roughly 40 per cent in cooling energy overhead, an acknowledged figure inside HUMAIN's own technical literature. That overhead is the ceiling on HUMAIN's claim to price-competitive sovereign compute against European or US alternatives, and it is the reason the pipeline favours coastal Dammam over interior Riyadh where cooling economics allow.
The Stack Above The Silicon
HUMAIN is not, and was never intended to be, a colocation business. The programme is organised as a four-layer stack: infrastructure, cloud, data and models, and applications, with the explicit goal of keeping Arabic-language inference on Saudi soil and Saudi government workloads on Saudi-controlled software. The two most visible products above the compute layer are ALLaM and HUMAIN OS.
ALLaM, the Arabic Large Language Model, launched its 34 billion parameter version in August 2025 alongside the HUMAIN Chat consumer application. The training corpus exceeds eight petabytes and is the largest declared Arabic-language dataset in existence. HUMAIN's own technical disclosures position the model against GPT-4 class systems on Arabic reasoning and dialect handling, and the Hugging Face release of an earlier seven billion parameter preview has attracted external evaluation. ALLaM is the clearest instance of HUMAIN using its own compute to produce a product that could not be procured from a Western provider without significant localisation work, and it is the argument HUMAIN makes when government counterparties ask why a sovereign model is worth the capital cost.
HUMAIN OS, unveiled by Tareq Amin at the February 2026 PIF Private Sector Forum, extends the same logic into an agentic operating system marketed to enterprise and government customers. The product runs natively on HUMAIN Core, the internal branding for the gigawatt-scale data centre infrastructure, and the commercial pitch is that Saudi ministries can run Arabic-first agentic workflows without either shipping data to a US hyperscaler or rebuilding an LLM stack in-house. The timing is deliberate: Saudi Arabia is positioning itself as the first country outside the US and China to commercialise a national operating system layer, and HUMAIN OS is the vehicle.
The honest question about the software stack is adoption. ALLaM and HUMAIN Chat have shipped, but the enterprise customers named by HUMAIN on stage at the PIF Forum include Aramco, stc, Saudi Electricity, and the Ministry of Finance, all of which are effectively arms of the same shareholder. The test case for HUMAIN's commercial viability will be whether it can book a material contract outside the PIF perimeter, and that is not yet demonstrated.
Who Actually Runs HUMAIN
Tareq Amin was appointed chief executive of HUMAIN in May 2025, the same month the company was formed. He joined from Aramco Digital, where he was the founding chief executive and where he built the carve-out that has since been folded into HUMAIN. His prior role, as chief technology officer of Rakuten Mobile in Japan, gave him direct operational experience with Open RAN deployments, and that background shapes HUMAIN's modular approach to data centre build-out.
The board is chaired by Crown Prince Mohammed bin Salman. That single reporting line is the reason HUMAIN can move at the velocity it has, and it is also the reason external investors are cautious about allocating capital that does not carry a PIF guarantee. The governance structure places HUMAIN in the same category as NEOM and the Red Sea Development Company: vehicles that execute national strategy, with political attention at the top and operational latitude underneath.
Beneath Amin, the senior technical bench is a mix of returning Saudi operators, Aramco Digital veterans, and international hires from NVIDIA, Microsoft, and the Indian hyperscaler ecosystem. The ALLaM team alone comprises more than 120 AI specialists, 35 of whom hold doctoral degrees, and the declared gender split on the technical side sits at 50-50, a number cited repeatedly in government communications because it supports the Vision 2030 talent narrative.
The Competitive Map
HUMAIN is not operating in isolation. G42 in Abu Dhabi, through its Khazna Data Centres subsidiary and its Microsoft partnership, controls a comparable footprint with a different political economy. Qatar is positioning for a narrower research and governance role through the Qatar Computing Research Institute. Egypt's Karnak project is smaller and operates under tighter financial constraints. The Kingdom's explicit ambition, stated by Amin on multiple occasions in 2025 and 2026, is to place Saudi Arabia third behind the United States and China in global AI capacity by 2030.
Whether HUMAIN closes the gap with G42 in the near term is the operational question. G42 has the earlier start, the deeper existing customer base, and a more developed software layer through Core42 and Inception. HUMAIN has the larger declared capital commitment, the backing of the Kingdom's sovereign wealth vehicle, and the better relationship with the current US political coalition. The competitive equilibrium inside the Gulf is likely to settle with both entities running large fleets on complementary workloads rather than one displacing the other, and some observers inside the regional venture community expect eventual joint ventures at the infrastructure layer.
Against Western and Chinese peers, HUMAIN's structural advantages are cheap power, deep sovereign capital, and regulatory flexibility at the national scale. Its structural disadvantages are chip-supply dependency, the cooling overhead described above, and the absence of a domestic semiconductor industry to hedge the import exposure. The HUMAIN executive team is candid about these limitations in briefings, which is a marked shift from the rhetoric of similar national programmes a decade ago.
By The Numbers
- 600,000 NVIDIA GPUs, the three-year ceiling declared in Washington in November 2025
- 35,000 advanced NVIDIA chips approved for HUMAIN and G42 combined in the first tranche of post-diffusion-rule US export licences
- 18,000 NVIDIA GB300s committed to the xAI anchor cluster inside the 500 megawatt flagship site
- 150,000 NVIDIA GPUs committed under the AWS Riyadh AI Zone partnership
- 1 gigawatt AI infrastructure ceiling under the AMD-Cisco-HUMAIN joint venture, with a 100 megawatt first phase in Q2 2026
- 6.6 gigawatts total HUMAIN data centre capacity targeted over ten years
- 250 megawatts of hyperscale capacity backed by the 1.2 billion dollar Infra Fund framework signed at Davos in January 2026
- 1.8 trillion parameters claimed for the ALLaM flagship model class, against an eight-petabyte training corpus
- Under 2 US cents per kWh the marginal electricity cost Aramco cites as Saudi Arabia's structural AI advantage
- 40 per cent cooling energy overhead of Saudi data centres relative to a Nordic benchmark
HUMAIN is the clearest test yet of the thesis that petrodollar capital can buy a frontier AI position. The order book is real, the construction is underway, and the political relationships have cleared a path through US export controls that no other non-Western buyer has matched. The 600,000 GPU headline is a useful symbol but it disguises what is actually three distinct programmes, each with its own risk profile. The xAI anchor is a contract position on Musk's willingness to keep his compute commitments, and that is not a risk HUMAIN can control. The AWS Riyadh zone is structurally more durable but represents a smaller share of headline capacity. The AMD-Cisco joint venture is the most commercially conventional piece and the one most likely to deliver on schedule, which is why HUMAIN is quietly using it as the operational baseline in internal planning.
The sovereignty claim, however, remains qualified. HUMAIN's hardware is US-sourced under licences that carry reporting obligations. Its cooling overhead is a permanent structural cost. Its commercial customer base outside the PIF perimeter is unproven. The honest read is that the Kingdom is building an industrial-scale Arabic-language AI capacity that it controls operationally but does not yet own at the component layer, and that the political bet is on keeping US export controls permissive for long enough to convert the hardware advantage into a software moat through ALLaM and HUMAIN OS. If that conversion works by 2028, HUMAIN becomes a genuine third pole. If it does not, HUMAIN becomes a very large, very expensive captive cloud for the Saudi state, which would still be a material outcome but a less transformative one.
What To Watch Next
Three developments in the next six months will test the trajectory. First, the Q2 2026 commissioning of the 100 megawatt AMD-Cisco phase is the first real delivery milestone on the post-November commitments, and schedule slippage there would be the clearest negative signal. Second, the next BIS export licensing round, expected in the third quarter, will reveal whether the initial 35,000 chip approval was a ceiling or a floor for Saudi and Emirati access in the current political environment. Third, the first external commercial contract for HUMAIN OS outside the PIF perimeter, if it materialises, will be the signal that the stack above the silicon is actually sellable rather than simply deployable inside captive Saudi entities.
Beyond those near-term tests, the larger question is whether HUMAIN can execute the 2028 target of a genuine third-pole position in global AI compute without a material disruption in US-Saudi relations or a competing Chinese offer that forces a hedge. The infrastructure investment is already sunk. The strategic orientation is not.
Sources for this episode include NVIDIA, AMD, and HUMAIN press releases from the 19 November 2025 US-Saudi Investment Forum; Commerce Department licence disclosures; the February 2026 PIF Private Sector Forum; the January 2026 World Economic Forum framework agreement with Infra; Arab News, The National, Bloomberg, Reuters, CNBC, Data Centre Dynamics, and Saudi Gazette reporting; and Aramco and Acwa Power public disclosures.
A large language model, meaning software trained on massive text data to generate human-like text.
AI that can independently take actions and make decisions to complete tasks.
When an AI model processes input and produces output. The actual 'thinking' step.
The internal settings an AI model learns during training. More parameters generally means more capable.
Graphics Processing Unit, the powerful chips that AI models run on.
A standardized test used to compare AI model performance.