Saudi Arabia’s $10 Billion AI Megaproject: How New Deals With Nvidia and AMD Are Thawing the U.S. Chip Market - Michał Opalski / ai-agile.org
Introduction: From Petro‑State to Compute Superpower
In 2016 Crown Prince Mohammed bin Salman sketched an audacious roadmap—Vision 2030—that promised to pivot the Kingdom of Saudi Arabia from oil dependency to knowledge‑based prosperity. At the time, critics dismissed it as branding. Less than a decade later the kingdom is signing the biggest single order for artificial‑intelligence hardware the world has ever seen. The new state‑backed firm Humain—a joint venture between the Public Investment Fund (PIF) and energy‑tech arm Aramco Digital—has inked parallel agreements with Nvidia and AMD worth a combined US $10 billion. The contracts cover hundreds of thousands of AI accelerators, a network of three “AI factories” inside the Gulf, and a surprise fourth site near Houston, Texas.
Why does a desert monarchy suddenly want more compute capacity than most European countries combined? Riyadh’s answer is brutally simple: just as oil powered the 20th‑century economy, compute will power the 21st—and whoever owns surplus exaFLOPS will wield disproportionate influence over finance, media and security. For the United States, meanwhile, the deal is a real‑time experiment in whether export controls can be fine‑tuned to constrain China without starving allied demand. For chipmakers stressed by cyclical inventory corrections, the Saudi cheques arrive like rain after drought.
1. Anatomy of a $10 Billion Bet
1.1 Dual‑Track Supply Strategy
Humain’s procurement team structured its orders into two mirrored tranches. Nvidia’s leg starts with 18,000 Blackwell‑generation GPUs (principally B200 SXM modules paired with Grace CPUs) landing in Q4 2025, scaling to “several hundred thousand” by 2030. AMD’s track begins with 12,000 Instinct MI325X boards and migrates to the MI400 family once that silicon clears validation in early 2026. Both suppliers are guaranteed minimum annual volumes; above those floors Humain can flex between vendors to balance thermals, software stacks and political optics.
Underpinning the hardware flows are elaborate escrow ladders: phases unlock only after U.S. export licences are renewed, on‑site auditors sign compliance attestations, and Humain proves that at least 60 % of previous shipments are already installed. The design creates rolling liquidity for Nvidia and AMD—revenue is recognised at each ladder—while giving Washington monthly visibility on end use.
1.2 Financing & Governance Mechanics
Rather than dip into PIF’s foreign‑bond portfolio, Riyadh opted for a split structure: $4 billion in short‑dated sukuk (Islamic bonds) listed in Dubai, backed by anticipated data‑centre cash‑flows, and $6 billion in phased purchase orders (PPOs) denominated in dollars. The PPOs sit off PIF’s sovereign balance sheet, allowing Humain to borrow cheaply against them in the commercial paper market. A four‑bank syndicate led by JPMorgan arranged $1.2 billion of 364‑day revolving credit so that the project can pay suppliers within 30 days even if sukuk settlements lag.
Corporate governance is unusually transparent for the region. Independent directors from Singapore’s Temasek and Norway’s NBIM sit on Humain’s audit committee, and quarterly financials will be reported under IFRS. The discipline is partly self‑interest: to resell unused compute to Western AI startups, Humain must look like any other colocation vendor subject to Sarbanes‑Oxley disciplines.
1.3 Risk‑Sharing Clauses
Hidden in the term sheets are pages of event‑risk hedges. If U.S. regulators tighten performance caps on export‑controlled GPUs, Humain can switch the balance of its order book to AMD or even Intel without penalty. If oil falls below $50 a barrel for two consecutive quarters, Nvidia and AMD agree to stretch payment schedules by up to 270 days in return for 25 basis‑point coupons. There are also controversial kill‑switch triggers: firmware in every accelerator can throttle FP8 throughput if an unlicensed re‑export is detected.
2. Why Riyadh Wants—and Needs—ExaFLOPS
2.1 The Arabic Data Dividend
Only about 1 % of the internet’s text is in Arabic, and even that tiny slice is riven by dialectical variance. Global large language models (LLMs) thus underperform on Gulf Arabic, let alone Hijazi or Najdi dialects. By training sovereign models on vertically integrated Saudi datasets—legal rulings, energy contracts, historical archives—Humain hopes to create a linguistic moat analogous to China’s Mandarin corpora. The plan is to license Arabic LLM APIs to fintechs in Cairo, media companies in Casablanca and e‑commerce platforms across the Levant, turning Saudi Arabia into the de facto language infrastructure provider for 400 million speakers.
2.2 Decarbonising Oil With Data
Aramco already operates the world’s largest industrial IoT network, streaming 20 billion sensor points per day from wells, pipelines and refineries. Humain’s compute clusters will crunch that firehose to predict equipment failures, optimise enhanced‑oil‑recovery chemistry and model carbon‑capture sequestration layers beneath the Empty Quarter. Company engineers estimate that every 1 % gain in extraction efficiency is worth $3 billion annually—enough to amortise an entire AI factory in two years.
2.3 Soft‑Power Compute Diplomacy
In a region where influence once flowed through petrodollars and religious custodianship, Riyadh now wants to rent exaFLOPS to allies from Rabat to Jakarta. Since many Muslim‑majority economies lack sovereign clouds, anchoring their AI workloads in Saudi data centres creates long‑term diplomatic leverage while aligning with cyber‑sovereignty norms. Humain has already offered subsidised training credits to African Union research labs—paid for, in effect, by oversupply clauses in its GPU contracts.
3. The U.S. Angle: Commerce, Congress and CHIPS
3.1 Anatomy of Export Licences
Last October the U.S. Bureau of Industry and Security (BIS) raised the performance ceiling for GPUs exported to “countries of concern”, a category that excludes Saudi Arabia but includes China and Russia. Nvidia’s special‑order B200‑SA and AMD’s MI325X‑SA are therefore licensable, provided firmware enforces cluster size and secure‑boot attestation. Each export licence covers a 90‑day window and a specific shipment lot; Humain must file end‑use reports within 30 days of installation, including telemetry hashes confirming geofencing.
3.2 Capitol Hill’s Balancing Act
Politically the deal is awkward. Senate Majority Leader Chuck Schumer supports allied‑only fast‑track licences but warns that “the Gulf must not become Shenzhen‑by‑stealth.” Meanwhile House Republicans blast the Biden and Trump administrations alike for creating an inspection regime “too cosy” with a kingdom still under scrutiny for human‑rights abuses. A bipartisan Secure Allied AI Act, due this summer, will almost certainly bake Humain’s audit terms into statute, making future compliance failures potential felony offences.
3.3 CHIPS Act Guardrails: A Saudi Loophole?
Washington’s $52 billion CHIPS Act offers fabs subsidies on condition they avoid “material expansion” in China for ten years. Some lobbyists argue that Saudi co‑investments could let AMD finance U.S. expansions while keeping the letter of the guardrails: cash from Riyadh goes into New York or Arizona fabs, not Shanghai. Critics warn this sets a precedent for other petro‑states to buy their way into domestic U.S. industrial policy.
4. Market Impact: A Lifeline for Semiconductor Supply Chains
4.1 Clearing the Inventory Glut
In early 2025 analyst firm TrendForce estimated that 380,000 high‑end GPUs were sitting in inventories at OEMs and cloud providers, victims of double‑ordering during the pandemic and a sudden lull in crypto demand. The Saudi order alone absorbs nearly half that glut over two years. Taiwan’s OSAT houses, whose utilisation fell below 60 % in Q1, have already recalled furloughed workers.
4.2 Competitive Dynamics: Nvidia vs AMD vs Intel
Nvidia retains a 78 % share of data‑centre accelerator revenue, but AMD’s ROCm stack matured dramatically with the MI300 launch. By guaranteeing AMD a floor of 40 % share inside Humain’s clusters, Riyadh gives Lisa Su’s team the volume needed to keep pace. Intel sits on the sidelines for now, yet Humain’s request for Gaudi3 inference quotes could bring Pat Gelsinger a badly needed flagship win.
4.3 Secondary Waves: Cooling, Optics, Real Estate
Liquid‑cooling vendors such as Vertiv and CoolIT have already reported Saudi RFQs worth $600 million. Corning is shipping 800‑gig optical transceivers to Riyadh next spring. Even local real estate benefits: desert land within Neom’s industrial zone jumped 35 % after the AI factory plot was announced.
5. Technology Deep Dive
5.1 Blackwell Architecture & NVLink‑5
Nvidia’s Blackwell B200 pairs 208 billion transistors with on‑package HBM3e delivering 8 TB/s aggregate bandwidth. New FP8 and FP4 data types cut energy per training token by 40 % relative to Hopper, while NVLink‑5 pushes inter‑GPU bandwidth to 1.8 TB/s—critical for the 16‑GPU “SuperPOD‑SA” boards custom‑designed for Humain.
5.2 AMD’s Memory‑Centric Vision
AMD’s MI400 roadmap, previewed at Hot Chips, leans into a modular fabric where each GPU can address 576 GB of HBM4 as a shared pool. Early benchmarks on GPT‑4‑class workloads show parity with Nvidia in token‑per‑second but 25 % lower wattage thanks to a chiplet architecture borrowed from Epyc server CPUs.
5.3 Data‑Centre Design in the Desert
Keeping 100 megawatts of GPUs cool in 48 °C ambient heat is non‑trivial. Humain will deploy two‑phase coolant baths using a fluoroketone dielectric fluid that vaporises at 40 °C, condenses against chilled coils and recirculates without water loss—a crucial feature in a region where desalination already consumes 20 % of grid power. Solar arrays built by Masdar in neighbouring UAE will offset 45 % of the factories’ electricity draw during daylight, smoothing the Kingdom’s peak‑load curve.
6. Regional Chessboard: Gulf AI Arms Race
Abu Dhabi’s G42 has 24,000 GPUs online today and a memorandum with OpenAI. Qatar’s Lusail Compute Hub plans 30,000 Nvidia accelerators. Israel is training sovereign LLMs on AWS. Against this backdrop Riyadh’s target of 600,000 GPUs by 2030 is both defensive and aspirational: defensive because it prevents regional isolation, aspirational because it positions Saudi Arabia as the default AI export hub to Africa and South Asia.
7. ESG & Cultural Dimensions
Riyadh pledges a Power Usage Effectiveness (PUE) below 1.10—aggressive even by Scandinavian standards. Waste heat will warm greenhouses supplying Neom’s vertical farms. Workforce policy is equally radical: Humain must hit 45 % female technical staff by 2028, up from 22 % today, and every foreign engineer is paired with a Saudi graduate in a compulsory mentorship rotation.
8. Future Scenarios & What Could Go Wrong
Conclusion: Bytes Are the New Barrels
Saudi Arabia’s $10 billion hardware splurge is far more than a vanity project. It is the clearest signal yet that the race for artificial‑intelligence advantage is tilting from talent to raw compute—and that petrostates able to parlay energy wealth into silicon will matter long after oil demand plateaus. For Nvidia and AMD the contracts offer multi‑year revenue visibility and political insulation from China‑centric scrutiny. For Washington they serve as a proof‑of‑concept that export‑control regimes can support, rather than stifle, allied innovation.
If Humain brings its first 100‑megawatt factory online by 2026, Riyadh will operate the densest cluster of leading‑edge GPUs outside the United States and perhaps greater‑than‑Beijing. In an era where every industry—finance, biotech, entertainment—competes for scarce training hours, the ability to rent a slice of the Saudi desert may become as strategically valuable as the right to lift a barrel of crude from it a generation ago.
