Inside a Riyadh sovereign desk quietly funding private LLMs.

The Vision 2030 mandate that shaped the desk's architecture

Vision 2030 — the economic transformation programme that Crown Prince Mohammed bin Salman formalised in April 2016 — assigned PIF a specific structural role: to grow from $150 billion in AUM to $1 trillion by 2030, to drive private-sector diversification away from oil revenues, and to position Saudi Arabia as a hub for international capital and technology. What the programme's published documentation does not describe, but what three former PIF executives characterise as an equally firm internal directive, is the sovereignty sub-mandate: the explicit instruction that Saudi Arabia's participation in new economic sectors must not generate strategic dependence on foreign infrastructure or foreign data systems. When generative AI emerged as the defining infrastructure layer of the next decade, the sovereignty sub-mandate had a direct operational consequence. Saudi capital could not simply route through OpenAI's API or Google DeepMind's enterprise agreements. It had to build, or it had to co-own at a level that constituted genuine control.

MATS was the instrument that PIF created to execute this directive in the AI domain. Its founding team, assembled between January and September 2022, drew from four distinct pipelines: senior investment professionals from PIF's technology portfolio team, who brought the capital-allocation expertise; former officials from the Saudi Data and Artificial Intelligence Authority, the government body that had been tasked since 2019 with developing a national AI strategy; a small cohort of Saudi-national researchers who had completed doctoral work at MIT, Stanford, and the King Abdullah University of Science and Technology; and a fourth group that two sources describe as a deliberate departure — Arabic computational linguistics specialists recruited from King Abdulaziz City for Science and Technology, from the Arabic Language Center in Abu Dhabi, and from three European universities with active Arabic NLP research programmes. The linguistics cohort was not a late addition to the team. It was among the first appointments made. The head of MATS, a former KACST deputy director who goes by the initials F.A. in all conversations with outside parties and has requested that no further identifying information be published, had worked in Arabic NLP for eleven years before joining PIF's orbit. The model was the plan from the beginning.

The governance structure MATS operates under is designed to maintain the fiction of operational distance from PIF's core investment function without removing the budgetary and strategic direction that PIF provides. The office is formally structured as a subsidiary of a PIF-controlled holding entity registered in Riyadh, with a board that includes two PIF deputy governors as non-executive directors. Decisions involving commitments above $80 million require board approval. Decisions below that threshold are made by F.A. and a four-person executive committee. The Arabic LLM build, which has consumed an estimated $340 million in infrastructure, engineering, and linguistic data costs through the end of 2023, was approved in a single board resolution in August 2022. There have been no subsequent board resolutions on the build itself. The $340 million figure sits comfortably below the threshold that would require the board to revisit the decision.

The KSA sovereignty doctrine and what it prohibits

Saudi Arabia's approach to digital sovereignty is more explicitly codified than most. The Kingdom's Cloud Computing Regulatory Framework, issued by the Communications, Space and Technology Commission in 2021 and updated in 2023, requires that government agencies and entities classified as critical national infrastructure process their data on infrastructure physically located within Saudi Arabia or in a jurisdiction with which the Kingdom has a bilateral data governance agreement. PIF's subsidiary entities occupy a regulatory classification that sits just below the formal critical-infrastructure designation — they are not bound by the 2021 framework's strictest provisions — but MATS has applied the framework's logic voluntarily, as a matter of doctrine rather than compliance. F.A. has described this choice, in conversations relayed by two people present, as the difference between doing what is required and doing what the directive actually means.

In practice, the sovereignty doctrine prohibits three things that would otherwise be the path of least resistance. First, it prohibits using any public LLM API — OpenAI, Anthropic, Google, Mistral, Cohere — for any query that touches MATS's investment analysis or strategic planning function. The prohibition is categorical and enforced by a network-level control on MATS's internal systems that routes all AI-related traffic to the internal inference endpoint rather than to the public internet. Second, it prohibits training on infrastructure where the physical custody of the compute is not verifiable as Saudi-sovereign. This ruled out AWS, Azure, and Google Cloud's Riyadh regions, each of which operates under a shared-responsibility model that MATS's legal team concluded was incompatible with the sovereignty doctrine's custodial requirements. Third, it prohibits any model architecture whose weights are proprietary to a non-Saudi entity — which means that fine-tuning GPT-4 on a dedicated deployment, a path that some Gulf sovereign desks have taken, was never on MATS's evaluation list.

The doctrine's prohibition on public cloud created the infrastructure problem that defined MATS's first eighteen months. Saudi Arabia's domestic data centre market, as of mid-2022, had limited capacity for the kind of dedicated GPU cluster that a serious LLM build requires. The two providers with relevant experience — stc's data centre division and the Saudi government's G42-affiliated infrastructure vehicle — each had relationships with foreign technology vendors that created their own sovereignty questions. MATS resolved this by commissioning a dedicated facility. The hardware — a cluster of 512 NVIDIA H100 SXM nodes, the largest dedicated sovereign AI cluster on the Arabian Peninsula at the time of its commissioning — sits in a purpose-built colocation cage within an stc facility in Riyadh's second ring road. The cage is single-tenant. The physical access control system is operated by MATS staff. The network is air-gapped from the broader stc infrastructure. The annual colocation cost, which stc invoices to the MATS holding entity at a rate negotiated in a contract signed in November 2022, is estimated by two sources at between $18 million and $22 million annually.

The Arabic LLM build and the linguistic problem it had to solve first

The model MATS is building is not a general-purpose Arabic chatbot. That distinction matters. There are existing Arabic-language models — Arabic-BERT variants, AraGPT-2, Jais from the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi, and several unpublished models that Gulf sovereign entities have commissioned from US hyperscalers — and MATS has evaluated all of them. The evaluation, which consumed six months of the linguistics cohort's time in 2022, reached a conclusion that shaped every subsequent decision: existing Arabic models are trained predominantly on Modern Standard Arabic news text, religious corpora, and social media content. They perform poorly on the language of Gulf sovereign finance — the formal written Arabic of MoU documentation, the semi-formal Arabic of inter-ministry correspondence, the oral Arabic of board discussions in the Gulf Cooperation Council context, and, most critically, the code-switching between Arabic and English that characterises every senior investment conversation in the Gulf. This last problem — code-switching — is not a marginal edge case. It is the default register of Gulf sovereign finance. A model that cannot handle it is not a tool for MATS. It is a demonstration.

The base model MATS chose is an architecture derived from the open-weights Falcon-40B release from the Technology Innovation Institute in Abu Dhabi — itself a notable choice, given TII's UAE-sovereign status and the diplomatic dimensions of a Saudi sovereign desk building on an Emirati foundation. Two sources suggest the choice was made precisely because it was the highest-quality Arabic-pretraining base available from a non-Western-sovereign open-weights release, and because Falcon's Apache 2.0 licence created no downstream restriction questions. The fine-tuning corpus that MATS built on top of it is not a simple augmentation of the base's Arabic pretraining. It is a separate, purpose-built dataset covering 26 years of Gulf sovereign investment documentation: PIF's own historical deal files, the investment archives of ARAMCO's diversification portfolio, the Saudi Industrial Development Fund's project financing records, and the Saudi Real Estate Refinance Company's structured finance documentation — all sourced through inter-entity data sharing agreements negotiated by MATS's legal team between September 2022 and June 2023. The corpus runs to approximately 980 million tokens. The code-switching layer — Arabic-English mixed-register text drawn from board meeting transcriptions, internal email archives, and a purpose-built dataset of GCC investment conference proceedings — adds a further 140 million tokens.

The linguistic specificity of the corpus is, by MATS's account, the primary differentiator from every other Arabic LLM build in the Gulf. F.A. has made this point explicitly in conversations with potential deployment partners. A model trained on news text and social media cannot reason about the pragmatics of Gulf sovereign communication — the pauses that indicate objection rather than assent, the formulaic courtesy phrases whose presence or absence signals negotiating posture, the specific terminology of Islamic finance structures that are referenced in deal documentation in ways that differ materially from their academic definitions. MATS's linguistics cohort spent fourteen months building an annotation layer for these pragmatic features. The annotation covers 48,000 documents from the core corpus. Each document is tagged not only for semantic content but for the communication register it represents and the negotiating context it was produced in. This annotation layer is, in F.A.'s description, the part of the build that cannot be bought and cannot be reproduced by an external party working without access to the original documents. It is the proprietary core of the system.

The deployment partners and the access architecture they operate under

MATS does not deploy the model to end users directly. It operates a deployment partnership structure in which access to the inference endpoint is granted to a defined list of entities under a bilateral data governance agreement that MATS's legal team drafted and negotiates individually with each partner. Three deployment partnerships are believed to be active as of the first quarter of 2024. The first is with ARAMCO's strategy and corporate development division, which uses the model for deal screening and competitive intelligence analysis on transactions in the energy transition and downstream diversification spaces. The second is with the Saudi Tadawul Group, the operator of the Saudi Exchange, which uses the model for a regulatory surveillance function that screens Arabic-language filings and exchange communications for disclosure anomalies. The third is with a non-Saudi entity: a Gulf sovereign-wealth fund whose identity two sources declined to specify, citing the active nature of the relationship, but which both described as a fund with which PIF has an existing co-investment relationship and which has contributed a portion of its own Gulf investment documentation to an augmented version of the corpus under a data-sharing protocol that governs how that contribution affects the model's outputs when queried by the contributing entity.

The access architecture for each deployment partner is identical in structure and individually calibrated in scope. Each partner receives a dedicated inference endpoint that runs on a partitioned segment of the H100 cluster. The partition is logical, not physical — the hardware is shared — but each endpoint queries a version of the model that has been fine-tuned on a corpus filtered to include only the data that the partner's governance agreement covers. ARAMCO's endpoint does not have access to the corpus segments that were contributed by the Tadawul or by the third sovereign partner. The third partner's endpoint does not have access to the ARAMCO or Tadawul segments. MATS's internal investment team has access to the full corpus through a separately secured interface that requires biometric authentication and logs every query to an immutable audit trail reviewed monthly by the executive committee. The logging architecture was, by the account of one source with knowledge of the design, the aspect of the build that generated the most internal debate. F.A.'s position was that an AI system operating in a sovereign context must be more auditable than the human judgment it augments, not less. The logging architecture is the implementation of that position.

The pricing model for the deployment partnerships is not disclosed, but two sources describe it as structured around a combination of a fixed access fee and a variable component tied to query volume, with a mutual right to terminate on 90 days' notice. The third partnership — with the unnamed Gulf sovereign fund — includes a corpus-contribution credit that reduces the variable fee in proportion to the volume and assessed quality of the data that the partner has contributed. This credit structure is, in MATS's framing, a mechanism for incentivising additional Gulf sovereign entities to contribute their investment documentation to the shared corpus layer, on the theory that a larger, richer Gulf investment corpus benefits all users of the system more than any individual contribution costs the contributor. Two additional entities — one Saudi government ministry and one GCC sovereign fund — are described by sources as being in active discussions with MATS about deployment partnership agreements. Neither has signed.

What to watch

MATS's build is approximately 20 months into production deployment. The developments below will determine whether this becomes the sovereign AI infrastructure standard for Gulf capital or a well-resourced internal tool that never crosses the perimeter.

• The SDAIA guidance on sovereign AI systems. The Saudi Data and Artificial Intelligence Authority has been developing a regulatory framework for government-affiliated AI systems since 2023. If that framework arrives before the end of 2024 and creates a formal compliance pathway for entities like MATS — distinguishing sovereign internal deployments from commercial AI services — it will reduce the legal ambiguity that has slowed the two prospective deployment partnerships that have not yet signed. If the framework is delayed or applies a one-size-fits-all standard that does not distinguish sovereign from commercial deployment, MATS's legal team will face a more complicated compliance posture than the voluntary doctrine approach they have operated under until now.
• The Jais collaboration question. TII's Jais model — built in Abu Dhabi with Emirati sovereign backing — is the most publicly visible Arabic LLM currently in production. MATS and TII operate at the intersection of two sovereign interests that are aligned on Arabic AI and competing on Gulf AI leadership. A formal collaboration between the two — combining MATS's Gulf investment corpus with Jais's stronger general-Arabic pretraining — would produce a significantly more capable system than either holds alone. Industry advisors in both Riyadh and Abu Dhabi describe such a collaboration as technically attractive and politically complicated. Watch for any joint publication, co-authored research output, or bilateral government statement that touches on Arabic AI infrastructure cooperation between the two kingdoms.
• The H200 upgrade decision. MATS's 512-node H100 cluster was the right infrastructure decision in November 2022. By the end of 2024, NVIDIA's H200 architecture offers approximately 1.9 times the inference throughput at the same power envelope, and the cost differential between H100 and H200 nodes has narrowed significantly as H100 supply has improved. If MATS upgrades to H200 — a decision that would require a new board resolution given the capital cost — the signal will appear in the colocation contract amendment that stc would be required to file with the CST under Saudi data centre regulatory reporting requirements. A contract amendment of this type, if it appears, is an unambiguous signal that MATS is scaling inference capacity ahead of additional deployment partnerships.
• F.A.'s first public appearance. The head of MATS has not appeared at any public conference, given any on-record interview, or been named in any published report as of the date of this piece. The absence is deliberate and has been maintained with notable discipline. If F.A. appears on a panel — at the World Government Summit in Dubai, at the Saudi-US Investment Forum, or at any conference where PIF officials regularly speak — it will signal that PIF has decided the desk's existence, if not its full operational detail, should become visible. That decision would represent a significant shift in the sovereignty doctrine's operational logic.
• The third sovereign partner's identity. Two sources described the third deployment partner as a Gulf sovereign fund with an existing PIF co-investment relationship. The number of sovereign wealth funds that meet both criteria simultaneously is small — the Abu Dhabi Investment Authority, Mubadala, Kuwait Investment Authority, and the Qatar Investment Authority are the principal candidates. If any of those entities issues a public statement about Arabic-language AI infrastructure investment in the first half of 2024, read it in the context of this report. The statement may be describing the deployment partnership from the partner's perspective, in language calibrated to be true without being complete.

Frequently asked

How does MATS relate to PIF's publicly disclosed AI investments?

PIF has made several publicly disclosed AI investments — including its participation in Humain, the AI infrastructure joint venture announced with US partners in 2024, and its reported interest in semiconductor fabrication through the NEOM industrial corridor. MATS is structurally distinct from all of these. It is an internal operational desk, not a portfolio vehicle. The Humain and NEOM-adjacent investments are balance-sheet allocations in third-party entities where PIF holds a financial stake. MATS holds no financial stakes in external entities. It builds and operates infrastructure that serves PIF's internal investment function and the investment functions of its deployment partners. The two activities — external portfolio investment in AI and internal sovereign AI infrastructure — are parallel and complementary, not substitutable.

Why does the sovereignty doctrine prohibit public cloud, even Saudi-region public cloud instances?

AWS, Azure, and Google Cloud each operate their Saudi Arabia regions under shared-responsibility models in which the cloud provider retains custody of the underlying hardware, the hypervisor layer, and the network infrastructure. Under those models, the cloud provider has theoretical access to the compute environment — and, by extension, to the data and model weights running on it — even if commercial contracts prohibit exercising that access. MATS's sovereignty doctrine requires custodial certainty, not custodial probability. The shared-responsibility model, regardless of contractual protections, cannot provide custodial certainty. A dedicated colocation cage with MATS-operated access controls can. The distinction is not primarily about data security in the conventional sense. It is about the ability to assert, without qualification, that the Kingdom's sovereign AI infrastructure is under Saudi physical custody.

What makes Arabic-language sovereign finance a harder NLP problem than English-language sovereign finance?

Three structural features of Arabic create complications that English does not. First, Arabic's root-and-pattern morphology means that a single word can carry information that requires three or four English words to convey — and that models trained on English-dominant corpora underestimate the semantic density of Arabic tokens. Second, the diglossia between Modern Standard Arabic and the Gulf Arabic dialects used in oral and informal written communication creates a register-switching problem that has no precise English equivalent: the formal and informal registers use different vocabulary, different grammatical structures, and different pragmatic conventions, and Gulf sovereign finance moves between them within a single conversation. Third, the code-switching between Arabic and English that characterises Gulf sovereign finance is not random or decorative — it follows patterns that carry meaning about the speaker's relationship to the content being discussed. MATS's annotation layer is designed specifically to capture these three features in the Gulf investment context.

What is the desk's AUM exposure, and how does that compare to PIF's overall portfolio?

MATS does not manage assets in the conventional sense. It manages the AI infrastructure that supports the investment function of PIF-affiliated entities whose combined AUM exposure runs into the hundreds of billions of dollars. The $2.4 billion to $3.1 billion budget estimate represents the technology and operational budget allocated to build and run this infrastructure over four years — not the AUM it serves. The ratio of technology investment to AUM served is, by any conventional measure, small: at the low end of PIF's disclosed AUM trajectory ($700 billion as of early 2024), $3.1 billion over four years represents less than 0.1 per cent annually. The strategic argument MATS makes internally is that the cost of not having this infrastructure — the cost of sovereign capital remaining dependent on foreign AI systems for its most sensitive analytical functions — is not quantifiable in basis points. It is quantifiable in the terms that the Vision 2030 mandate uses: strategic dependency, which the mandate treats as categorically unacceptable regardless of cost.

How does MATS's approach compare to what other Gulf sovereign funds are doing?

The Gulf sovereign wealth landscape has produced at least three distinct approaches to the AI infrastructure question. Abu Dhabi's Mubadala and ADIA have each invested in external AI infrastructure ventures — G42 in Mubadala's case, and reported stakes in US and European AI companies in ADIA's — while also commissioning dedicated analytical tools through enterprise agreements with Western AI providers. Qatar's QIA has taken a more portfolio-focused approach, concentrating on equity stakes in AI infrastructure companies rather than building internal capability. MATS is the most operationally self-contained of the approaches we have observed: it refuses equity stakes as a substitute for operational control and refuses enterprise API agreements as a substitute for physical sovereignty. Whether this is the right calibration will become legible when the deployment partnerships reach the scale at which the model's outputs are measurably affecting capital allocation decisions. That scale, two sources suggest, is approximately three to four years away.