Saudi Arabia Has Signed the Power Agreements Behind Its AI Data-Centre Push

Saudi Arabia’s artificial-intelligence build-out has acquired a less glamorous but more decisive partner: the electricity system. On 2 September, Saudi Energy announced three agreements linked to digital infrastructure. National Grid SA signed an electricity-provision agreement for HUMAIN’s AI data-centre project in Riyadh. Saudi Electricity Project Development Company, or PDC, signed a framework agreement with center3 for potential future data-centre and energy infrastructure projects. PDC also signed a memorandum of understanding with Huawei covering data and AI centres, related energy solutions and technical cooperation. [S1]

The announcements show that the Kingdom is coordinating grid, infrastructure and technology companies around a fast-growing sector. They do not disclose the Riyadh project’s contracted megawatts, a connection or energisation date, the value of the agreements, or how much electricity is available today. The center3 agreement is a framework for future projects; the Huawei instrument is an MoU. None should be described as a completed data centre or a new block of operating AI compute.

That distinction is important because capacity announcements are multiplying. HUMAIN and MIS disclosed a 250-megawatt data-centre design-and-build agreement in September, while other partnerships describe large multiyear ambitions. Those figures cannot simply be added together: they may refer to overlapping programmes, different phases or different definitions of site, facility and IT capacity. The power agreements could help make the pipeline deliverable. Their public text does not yet tell us how much of it is deliverable, or when.

Last verified: 27 September 2026.

Three agreements, three different levels of commitment

The first agreement is the most concrete. National Grid SA and HUMAIN agreed on the provision of electricity to HUMAIN’s AI data-centre project in Riyadh. Saudi Energy says the arrangement supports the project’s electrical infrastructure and future energy requirements. [S1] The disclosure confirms a named project and a named grid-side counterparty. It does not give a load, project phase, connection point, delivery schedule or tariff.

The second instrument links PDC, a Saudi Energy subsidiary, with center3, stc Group’s data-centre business. The companies signed a framework agreement to establish a strategic partnership in future data-centre and energy infrastructure projects. Its stated purpose is to support sector expansion and integrate digital infrastructure with energy infrastructure. [S1]

The third instrument is a memorandum of understanding between PDC and Huawei. It covers cooperation on data-centre and AI projects, digital infrastructure, energy solutions, technical expertise and possible future collaboration. [S1] The release does not describe a specific facility, a signed construction contract, a purchase order or an investment amount.

What each instrument establishes

All three can matter. A framework can remove coordination barriers before a project reaches procurement. An MoU can bring engineering and technical suppliers into early design. A grid agreement can be a necessary prerequisite for a facility. But these are not equivalent forms of evidence. The Riyadh electricity agreement is the clearest project-specific milestone; the other two remain enabling arrangements.

Why power has become a strategic AI input

An AI data centre is not just a warehouse filled with computers. It is a high-density electrical load with demanding requirements for continuity, cooling, network connectivity and operational security. A facility cannot sell reliable compute simply because a building is complete or a chip order has been announced. Power must be available at the right location, at the required scale, on the commissioning schedule, with redundancy sufficient for the service promised to customers.

The expansion of AI changes the shape of the problem. Large accelerator clusters can concentrate substantial demand in a small area. Their electrical load is more continuous and their cooling requirements more intensive than many conventional server deployments. Operators have to coordinate utility connections, transformers, substations, switchgear, backup systems, cooling plants and high-capacity networks. Each component has its own permitting, manufacturing and installation timetable.

The Saudi Energy announcement recognises this relationship in institutional form. The grid operator is directly involved with HUMAIN’s Riyadh project, while PDC is intended to connect energy development with data-centre expansion. Huawei’s MoU extends the cooperation layer into technology and engineering. It is a signal that the Kingdom sees power and digital capacity as interdependent infrastructure rather than separate industries. [S1]

The relevant policy question is not whether Saudi Arabia has abundant energy resources. It is whether power can be generated, transmitted, connected and delivered with the reliability and economics that a particular data-centre customer requires. A national electricity system may have generation capacity in aggregate while a particular urban node faces connection constraints or requires transmission upgrades. The public agreements do not provide enough detail to assess those local conditions.

The megawatt problem: one unit, several meanings

Data-centre disclosures use “MW” in different ways. A project may state the power capacity of its grid connection, the gross electrical supply to a campus, the critical load available to IT equipment, or a target for a future portfolio. Those quantities differ. Cooling, power conversion and other building systems consume part of a facility’s incoming electricity. The difference depends on design and operating conditions.

The term “AI capacity” adds another layer. Two campuses with the same electrical load can deliver very different computing output depending on the type and number of accelerators, memory configuration, network topology, storage, software and utilisation. A power commitment does not reveal how many GPUs will be installed or how much useful training and inference they can provide. The September electricity announcement does not state a MW figure at all. [S1]

This is why the 250 MW figure associated with HUMAIN and MIS should not be casually mapped onto the National Grid SA agreement. MIS’s Saudi Exchange disclosure describes a design-and-build agreement for AI data-centre facilities, increased from an earlier 50 MW plan to 250 MW. It does not, in the public summary, define the 250 MW as grid intake or critical IT load, identify phases or give an energisation timetable. The electricity agreement names HUMAIN’s Riyadh AI project but does not explicitly state that it is the same 250 MW scope or quantify the load. [S2]

The two announcements may be related, but the public documents do not provide a complete project map. Until HUMAIN, MIS or the grid operator identifies the relationship, analysts should not add their figures or present the power agreement as proof that all 250 MW has a firm supply date. That is not pedantry. It is the difference between a project pipeline and a financed, connected, commissioned service.

A still larger number entered the Saudi press in September. Saudi Gazette reported Communications Minister Abdullah Alswaha saying the Kingdom was working with the private sector to develop more than 14 GW of computing capacity. [S4] The report does not identify the sites, phases, power definition or delivery dates behind that figure. By contrast, Saudi Press Agency reported that operational data-centre capacity had reached 440 MW in 2025. [S5] Those numbers cannot be subtracted to calculate a construction backlog: one is a ministerial development claim without a published project ledger, the other a historical operating-capacity figure. The gap is precisely why the unit and development stage must accompany every headline.

What a grid agreement does—and does not—solve

A supply agreement can give developers a route to plan around a utility connection. It may establish a process, allocate responsibilities and identify future requirements. But a complete power-delivery record usually needs more: a defined load profile, connection voltage, substation or transmission works, construction milestones, energisation date, reliability arrangements and commercial terms. The September release does not publish those details.

There is also a sequence of distinct milestones. A customer can request capacity; the utility can study the load; parties can agree on connection terms; equipment and network upgrades can be designed; construction can start; the site can be energised; and the facility can pass commissioning tests. Even after that, servers must be installed and integrated before customers receive a service. A release that says “provision of electricity” should not be silently converted into “powered and operational”.

The grid side has to manage more than one project. Data centres compete for connection resources with housing, manufacturing, transport, public services and existing businesses. Concentrated loads can require investment in substations and transmission. If several facilities are built in the same growth corridor, the order and timing of their connections may affect each other. A power system that plans early can turn large digital loads into predictable anchor demand; a system that underestimates timing can leave completed buildings waiting for supply.

The September agreements appear designed to bring parties together before every detail is public. That can be a sensible development model. Early coordination helps identify suitable sites, equipment needs and funding. Yet an enabling partnership should remain labelled as enabling until a project-specific contract fills in the missing terms.

Energy cost and reliability shape the business case

For a data-centre operator, electricity is both a major operating cost and a service-quality condition. A low-cost supply that is interrupted can damage customer workloads. Highly reliable power generally requires redundant paths, backup generation, storage or other contingency systems. Those systems add capital and operating costs, and they affect emissions, fuel use and maintenance needs.

The power profile also changes by workload. Training runs can operate at very high utilisation for long periods, while inference demand can fluctuate with user activity and business hours. Some jobs can be shifted in time or location; real-time services cannot be delayed without affecting customers. A data-centre platform serving government, finance and industrial users needs to understand which workloads are flexible and which require continuous service.

Renewable power introduces another measurement challenge. A renewable project’s annual nameplate capacity does not automatically provide firm electricity at every hour to a data centre. Matching supply and demand depends on generation timing, storage, grid transmission, contracts and the operating profile of the facility. Operators can use power-purchase agreements or certificates to account for renewable electricity, but these do not by themselves describe physical delivery at the server rack.

The official September announcement does not disclose an energy source, emissions profile, tariff or efficiency target. That is normal for an initial infrastructure announcement. It means there is not yet enough evidence to say whether the agreements will create a low-carbon compute platform, a cost advantage or a defined reliability improvement. Those questions belong in subsequent project and operating disclosures.

The broader Vision 2030 trade-off is straightforward: AI infrastructure can support diversification, but it adds a large new electricity demand. The net value depends on the compute services and economic activity produced per unit of power, the reliability and cost of the system, and the extent to which local firms and workers capture the associated value. Megawatts delivered are an input metric; useful workloads, productivity and export revenue are the outcomes.

Coordination across companies and government entities

The three agreements draw in different capabilities. National Grid SA is the network counterparty for the named HUMAIN project. PDC is positioned as a development vehicle at the boundary of power and digital infrastructure. center3 operates in the domestic data-centre market through stc Group. Huawei supplies digital and energy technologies and engineering experience. HUMAIN is the AI platform seeking to build compute and applications. The structure could help line up infrastructure decisions that normally sit across separate organizations.

Coordination is valuable because project delays often occur at interfaces. A data-centre developer may have land and construction contractors but lack a timely grid connection. A utility may plan network capacity without a confirmed customer schedule. Equipment can arrive before a site is ready, or a building can finish before the network and cooling systems are commissioned. A shared development framework can identify these interdependencies earlier.

The institutional question is how commitments are governed. A public announcement provides the parties and general purpose, but not the decision rights, risk sharing, procurement method or dispute process. It does not say who finances network upgrades, who carries cost overruns, or how capacity is allocated if several users seek power at the same time. These matters may be confidential; they are nevertheless central to execution.

The Huawei MoU adds a technology partnership to the mix, but it should not be read as proof that Huawei will supply equipment to the Riyadh project. The company is named as a cooperation counterparty for possible data-centre and AI work, with an exchange of engineering expertise and future collaboration opportunities. [S1] Procurement and project delivery would require additional evidence.

Likewise, the PDC–center3 framework may create a pipeline for future joint projects. The release does not name locations or capacity. It may be strategically useful before deals are public, but readers should look for follow-on announcements that name the asset, scope and stage. The next useful evidence would be a site-specific agreement or a construction and connection milestone.

How this fits the Saudi AI capacity ledger

Saudi Arabia is assembling a broad AI infrastructure portfolio. The public record includes HUMAIN’s strategic relationships with chip and cloud providers, the MIS 250 MW design-and-build scope, planned hyperscale projects, cloud-region timelines and other proposed campuses. AMD, Cisco and HUMAIN said in August that MI355X-based systems were live and serving customers, while describing a separate, planned build-out of up to 250 MW beginning in 2027. [S3] Some announcements describe operating services; some describe construction; others are ambitions, frameworks or financing plans. A single national total that sums all these figures risks counting overlapping programmes and mixing stages.

Local business reporting supplies a second distinction. Al-Eqtisadiah reported center3’s ambition to develop AI data-centre infrastructure starting at 250 MW alongside stc’s HUMAIN partnership and connectivity agreements. [S6] That statement is a company plan, not an operating-capacity disclosure, and the report does not establish whether its 250 MW overlaps the MIS or AMD-led programmes. It belongs in a ledger as a separately attributed ambition until project ownership, sites and phases are reconciled.

The September power agreements should be inserted into that ledger as an infrastructure-enabling milestone. The HUMAIN–National Grid arrangement is project-linked but unquantified. The PDC–center3 framework and PDC–Huawei MoU are prospective partnerships. They can support later capacity, but neither has a public megawatt figure. Treating them as new supply would inflate the apparent pipeline.

An analytical tracker should record at least six fields for every project: named site or geography; project owner and delivery partners; stated capacity and its definition; development stage; grid agreement and energisation status; and customer-service status. It should separately identify whether the announced capacity overlaps another programme. Where a company does not publish a field, the tracker should say “not disclosed” rather than infer it from another press release.

This approach lets an analyst recognise genuine progress without confusing announcements with operating assets. A signed connection agreement is a step beyond an aspiration. A connected site is a further step. A commissioned facility serving paying users is stronger evidence still. Each milestone reduces uncertainty, but none substitutes for the next.

What to watch next

For the HUMAIN–grid agreement, the critical next disclosures are the amount and phase of capacity, the planned connection date and whether required network works have started. A site location, substation plan or phased load schedule would help determine whether the grid arrangement covers the 250 MW MIS programme or a different Riyadh project. The exact relation should be confirmed, not assumed.

For PDC and center3, watch for a named development with an ownership structure, capital commitment, land and power scope. A framework can mature into a joint venture, a design contract or a project-specific agreement. Those instruments would tell us whether this is an operating platform for investment or an exploratory cooperation channel.

For PDC and Huawei, look for a procurement, pilot or engineering scope. If the collaboration results in energy-management systems, cooling improvements, high-voltage equipment or monitoring technology, the practical contribution should be described with operational metrics. A general statement of expertise exchange is not yet evidence of deployment.

For the wider sector, follow physical milestones: land preparation, building permits, substation construction, transformer delivery, grid energisation, installation of server equipment and launch of customer services. Each one should have a dated source. Do not treat “groundbreaking” as construction completion or “ready” as commercial availability without checking what the company means.

Power data should also include the long-run system effects. How much additional demand is attributable to data centres? What grid investments are required? How are reliability and emissions measured? Are large loads located where transmission and generation can support them? These questions affect both project returns and the electricity system’s ability to serve the rest of the economy.

From connection agreement to customer compute

The next set of milestones should be read as a sequence, because each removes a different uncertainty. A grid study establishes what the network can support and what upgrades are needed. A connection contract assigns responsibilities and commercial terms. Construction of substations and feeder infrastructure makes the connection physically possible. Energisation supplies power to the site. Commissioning tests the facility’s electrical and cooling systems under load. Only after servers, networking and software are installed can an operator launch a usable service.

This sequence has a financial dimension. Developers may commit capital to land and buildings before every grid upgrade is complete. Utilities may need to order long-lead transformers and switchgear before the customer’s final load profile is fixed. If projects are coordinated early, those orders can match the data-centre schedule. If the parties proceed on different assumptions, the result can be a finished building waiting for power or an upgraded network with no near-term customer load.

The parties’ new agreements could reduce that interface risk, but only a dated project record will show whether they do. Useful disclosures would include the requested load by phase, the expected date for each grid connection, the status of network works, the facility’s critical IT load once defined, and the date the first customer workloads begin. These figures can be published without revealing security-sensitive site plans or customer identities.

It is also worth distinguishing a grid reservation from actual electricity consumption. A project may secure capacity larger than its initial server load, then increase consumption as equipment arrives. Tracking both contracted capacity and metered load would show whether the build-out is progressing and how quickly customers are using it. This matters for system planning as well as commercial performance: a reserved megawatt is a claim on future infrastructure, while a consumed megawatt is an operating load.

The assessment

The 2 September agreements are a meaningful coordination signal. Saudi Energy has linked the National Grid directly with HUMAIN’s Riyadh AI data-centre project and created channels for PDC to work with center3 and Huawei on future digital and energy infrastructure. The agreements acknowledge that AI capacity depends on the power system and on partnerships that connect it to construction and technology.

The evidence remains early. Only the National Grid–HUMAIN instrument is tied to a named project; the public release does not disclose capacity, timing or cost. The center3 arrangement is a framework for future projects. The Huawei instrument is an MoU. No new operating megawatts can be inferred from the announcement.

For Vision 2030, the test is whether this coordination shortens the path from project announcement to reliable, usable compute. The first public proof points should be specific: defined load, connection works, energisation, commissioned IT capacity, and customers using the service. Saudi Arabia has announced a power partnership for the AI build-out. The next measure is how much electricity reaches which facility, on what date, and what productive digital services it enables.

Related Vision 2030 context

• HUMAIN and MIS expand planned AI data-centre capacity to 250 MW
• Mistral and HUMAIN’s Saudi compute commitment remains optional
• Saudi Arabia’s operating AI compute has crossed an important threshold

Sources

1. [S1] Saudi Press Agency, “LEAP 2026: Saudi Energy Signs Deals to Support Digital Infrastructure for Data, AI Centers,” 2 September 2026. SPA.
2. [S2] Saudi Exchange, MIS issuer announcement on the HUMAIN AI data-centre design-and-build agreement, 20 September 2026. Saudi Exchange.
3. [S3] AMD, Cisco and HUMAIN, “AMD, Cisco and HUMAIN Expand Saudi Arabia’s AI Infrastructure as AMD Instinct Systems Go Live,” 31 August 2026. AMD.
4. [S4] Saudi Gazette, “Saudi Arabia to develop over 14 GW of computing capacity to accelerate AI growth,” 12 September 2026, reporting Minister Abdullah Alswaha's remarks. Saudi Gazette.
5. [S5] Saudi Press Agency, “Saudi Arabia Ranks Second Globally in Data Center Market Attractiveness,” 2 May 2026, reporting 440 MW operational capacity in 2025. SPA.
6. [S6] Al-Eqtisadiah, report on stc Group's LEAP 2026 projects and center3 capacity plans, 3 September 2026. Al-Eqtisadiah (Arabic).