In this articleSections · 9
Executive summary
Saudi Arabia has an opportunity to turn its energy expertise and investment capacity into a competitive compute industry, but the advantage depends on delivered power, utilization and measured service quality. This is a strategic argument, not a forecast of guaranteed cost or emissions savings.
Why this matters: Reliable, affordable power is a serious constraint on AI infrastructure. Saudi Arabia has an opportunity to turn its energy expertise and investment capacity into a competitive compute industry. This is a strategic argument, not a forecast of guaranteed cost or emissions savings.
What the announcements establish
PIF launched HUMAIN in May 2025 to invest across the AI stack. NVIDIA's 13 May 2025 announcement described planned AI factories with up to 500 MW of capacity over five years. AWS and HUMAIN announced plans for more than $5 billion of investment on the same date. These statements establish investment intent and announced scale; they do not show that every planned facility is commissioned or that customers receive cheaper service. PIF, NVIDIA, AWS.
Groq's 10 February 2025 announcement also described a Saudi inference expansion and reported an operating cluster in Dammam. That is evidence for a specific deployment from the provider, not a measurement of the whole market. Groq announcement.
The argument that follows is my inference from these announcements and from the engineering constraints. Announced capacity, delivered power and commercially available GPU hours are different milestones.
Replace the oil-tanker comparison with a real grid comparison
The first version imagined imported Saudi oil being shipped, refined and burned to power a European or American GPU farm. That can illustrate a hypothetical fuel chain, but it is a poor general description of competing data centers. Electricity comes from the site's actual grid or contracted generation mix. The EIA reports that petroleum supplied about 0.7% of US electricity generation in 2025. EIA electricity generation by source.
A useful comparison asks what each candidate site can deliver: firm power, connection time, redundancy, electricity price and carbon intensity. A Saudi gas-adjacent campus should be compared with real alternatives, including other gas grids and renewable, hydro or nuclear supply. Avoid charging every alternative an imagined tanker and refinery bill.
The IEA's 2025 Energy and AI report treats grid access and electricity supply as important constraints on data-center growth. It also discusses locating facilities where power and grid capacity are available. That supports an energy-aware siting strategy; it does not establish that any country will always be cheapest. IEA analysis.
Scenario A: build near available power
Imagine a future campus near Saudi generation and grid infrastructure. The opportunity is to shorten connection delays, plan cooling around the hardware and climate, and secure a contract suited to a continuous load.
This scenario still needs a generation design, fuel agreement, permits, grid studies and an operating model. On-site generation requires capital, maintenance, redundancy and fuel supply. Solar generation needs a separate plan for night-time and interrupted output. A nearby pipeline is not a completed power project.
The comparison should also include water availability, cooling choices and seasonal performance. Waste-heat use may be practical in a suitable design, but its temperature, nearby demand and integration costs must be demonstrated. Desert nights do not remove cooling costs throughout the year.
Calculate the whole GPU-hour
Start with the load definition:
Facility energy = IT load in MW × 1,000 × PUE × operating hours.
For an illustrative 1 MW IT load, PUE 1.2 and 720 hours, total facility energy is 864,000 kWh. PUE 1.2 means supporting systems use energy equal to 20% of IT energy; there is no universal rule that every processor watt requires another half-watt.
Multiply that energy by an eligible, dated tariff. Then add capacity charges, taxes and commitments separately. For the tariff example and corrected arithmetic, see Saudi Data Center Advantage.
A complete GPU-hour model also includes hardware depreciation, utilization, maintenance, network capacity, staff, financing and software efficiency. Lower electricity costs affect only part of that total.
For example, if energy is 20% of total cost and its cost falls 30%, the total falls 6%, assuming everything else stays constant. Those percentages are explicit scenario assumptions, not observed Saudi market figures. There is no verified model here that supports the earlier claim of a 30–50% reduction in GPU-hour cost.
Scenario B: use an existing compute hub
An established region may have a higher energy rate and still deliver a lower cost per useful result. Available hardware, stronger utilization, faster deployment, mature operations or proximity to customers can outweigh part of the power-price difference.
Compare the same workload and service level across both options. Measure cost per completed training run or per successful inference at the required throughput and latency. Include data-transfer charges, operational risk and any migration cost.
For training, distinguish moving a batch job between regions from splitting a tightly coupled training cluster across long-distance links. The network inside the cluster has its own bandwidth and latency requirements. An application that tolerates a distant batch job may still need inference near its users.
Exporting a digital service still has physical costs
A completed computation can be delivered over a network without shipping its fuel to the customer. That makes compute a plausible digital export. It does not make fiber transfer free, bandwidth unlimited or latency uniform around the world.
Routing, peering, congestion, data volume and the application's interaction pattern all matter. Benchmark the real customer paths; do not promise a global sub-200 ms round trip from geographic position alone.
| Decision | Evidence to request |
|---|---|
| Power availability | Committed capacity, connection date and redundancy design |
| Tariff | Provider, category, period, eligibility and all bill charges |
| Compute economics | Hardware cost, utilization and cost per useful result |
| Cooling | Seasonal PUE, water requirements and failure-mode testing |
| Network | Measured customer latency, throughput and transfer costs |
| Data governance | Storage location, access controls and contractual responsibilities |
Carbon savings need their own comparison
Energy adjacency does not automatically lower emissions. The result depends on the electricity displaced, generation efficiency, fuel lifecycle, methane leakage, cooling energy and additional renewable supply. Gas-based compute may compare differently with a coal-heavy grid than with hydro or nuclear supply.
Claims about carbon capture and waste-heat recovery need project evidence: installed systems, actual performance, energy penalties and an accounting boundary. Certificates or offsets should be identified separately from a facility's physical electricity mix.
The earlier estimate that moving a tenth of AI compute would avoid tens of millions of tonnes of CO₂ had no traceable model. It has been withdrawn, along with the illustrative 280-versus-95 cost cascade. A numerical comparison needs disclosed inputs and a defensible baseline.
The strategic opportunity
Saudi Arabia can build local capability in data-center operations, cooling, grid engineering, software and AI services. How much value the country captures depends on execution, customers, skills and sustainable operating economics. Local hosting can support data-residency requirements, while access, security and compliance still need their own controls.
Energy-linked capacity contracts, compute exchanges and other financial structures are possibilities, not established predictions. The stronger near-term task is concrete: deliver reliable power, usable compute and a service customers can evaluate.
The opportunity remains compelling. Treat power as a design input, prove the whole workload economics, and measure the resulting service. That is how an energy advantage becomes a compute business.