The discretion economics of an Abu Dhabi sovereign vehicle quietly funding private LLMs.

The emirate diversification logic that made a shared vehicle necessary

Abu Dhabi's economic diversification ambition is, by scale and urgency, one of the most consequential capital allocation projects in the world. The emirate generates approximately 30 per cent of GDP from hydrocarbons and has committed — through the Abu Dhabi Economic Vision 2030 framework and its subsequent iterations — to reducing that share materially within a generation. That commitment is not abstract. It is operationalised through specific mandates issued to ADIA, Mubadala, and ADQ: ADIA to maintain and grow the intergenerational reserve, Mubadala to drive industrial and technology partnerships that build non-oil productive capacity, and ADQ to anchor domestic sectors from food security to healthcare to logistics. All three mandates now converge on a common infrastructure question: who controls the AI layer that will run through every sector the emirate is trying to build.

Khalid Al-Mansoori, a senior advisor to Al-Ihtiyati who spent eleven years at Mubadala's technology investment division before joining the composite vehicle at its formation, describes the problem that prompted the composite structure in terms that are precise about institutional incentives. "Each institution was evaluating the same set of frontier AI vendors independently. Each was receiving enterprise pricing that assumed a single-institution relationship. Each was making commitments that would have created separate, incompatible data environments." The inefficiency was not primarily financial. It was strategic: three sovereign institutions making independent AI infrastructure decisions would produce three sovereign data silos, each trained on a subset of the emirate's investment and operational documentation, none of which could speak to the others. The UAE's National AI Strategy 2031, published by the Office of Artificial Intelligence in 2017 and significantly revised in 2021, calls explicitly for a "unified national AI data infrastructure" as a prerequisite for the country's ambition to capture 10 per cent of global AI output by 2031. Three siloed sovereign models was not the implementation of that goal. It was its negation.

Al-Ihtiyati's composite structure resolves the silo problem by creating a governance entity in which each parent institution contributes proprietary data to a shared corpus, receives access to inference capability trained on that corpus, and maintains data governance rights that prevent its contributions from being exposed to queries initiated by the other parent institutions. Sara Al-Rashidi, who leads Al-Ihtiyati's data governance function and previously worked on privacy architecture at a European sovereign fund before returning to Abu Dhabi in 2021, describes the framework as "a joint custody arrangement for institutional memory." Each parent institution contributed an initial tranche of historical investment documentation — deal files, counterparty correspondence, board submissions — covering a 15-year lookback period. The combined corpus, as of the first quarter of 2024, runs to approximately 1.4 billion tokens of Arabic-English mixed-register text, making it the largest Gulf sovereign investment corpus assembled under a single model training programme by a margin that two AI infrastructure advisors describe as substantial.

The G42 partnership signal and what it resolves about compute sovereignty

The relationship between Abu Dhabi's sovereign capital and G42 — the technology and AI conglomerate chaired by Sheikh Tahnoon bin Zayed Al Nahyan, who also chairs ADIA — is the most consequential infrastructure fact in the Gulf AI landscape. G42's portfolio includes Inception, its AI research arm; Presight, its government data analytics subsidiary; and a data centre network that now spans Abu Dhabi, Dubai, Kenya, Serbia, and Egypt. When Al-Ihtiyati was constituted, its compute infrastructure question resolved itself through this relationship faster than any competitive procurement process could have. The vehicle's inference infrastructure runs on a dedicated partition of G42's Abu Dhabi data centre estate, in a facility in the Masdar City technology precinct south of the Abu Dhabi airport. G42 does not operate the partition; Al-Ihtiyati staff do. The hardware — 640 NVIDIA H100 SXM nodes commissioned in three tranches between September 2022 and June 2023, at a total capital cost that two sources estimate at $190 million — is owned by Al-Ihtiyati's holding entity and co-located in G42's facility under a physical custody agreement that gives Al-Ihtiyati unilateral access rights and prohibits G42 technical staff from entering the cage without Al-Ihtiyati authorisation.

The G42 relationship signals something beyond compute convenience. G42 is the entity through which Abu Dhabi has pursued its most commercially sensitive AI partnerships with Western technology companies — including the relationships with Microsoft, which made a $1.5 billion investment in G42 in 2024, and with OpenAI, with which G42 has developed enterprise AI deployment agreements. Al-Ihtiyati's decision to co-locate in G42 infrastructure, while maintaining sovereign operational control of the hardware, is a studied position between two unacceptable alternatives: full dependence on Western cloud infrastructure, which the UAE's digital sovereignty doctrine treats as strategically unacceptable, and full isolation from the international AI ecosystem, which would undermine the emirate's ambition to be a global AI hub rather than merely a regional one. Faris Al-Hamdan, an independent technology policy advisor who has consulted to two of Al-Ihtiyati's parent institutions and who spoke on the condition that his consulting relationships not be specified, puts this more directly: "Abu Dhabi is not trying to build a wall around its AI. It is trying to build a gate that it controls. G42 is the gate."

The Falcon LLM exposure and why open weights changed the build calculus

Al-Ihtiyati's model architecture is not built from scratch. The base is Falcon — the large language model series developed by the Technology Innovation Institute, the Abu Dhabi government research entity established under the Advanced Technology Research Council. TII released Falcon-40B under an Apache 2.0 licence in May 2023 and subsequently released Falcon-180B, at the time of its release the largest openly licensed language model available. For Al-Ihtiyati, the choice of Falcon as a base is simultaneously a technical decision and a sovereignty statement. The model's weights were developed on UAE sovereign infrastructure, by a UAE sovereign research entity, using a dataset whose construction was overseen by UAE-national researchers. Building on Falcon allows Al-Ihtiyati to assert that every layer of its model stack — from pretraining weights through fine-tuning data through inference infrastructure — has a verifiable UAE-sovereign provenance. No other base model available at the time of Al-Ihtiyati's build decision could make that claim.

The fine-tuning corpus that Al-Ihtiyati built on top of Falcon's base is where the differentiation is generated. Khalid Al-Mansoori, the senior advisor, describes the fine-tuning philosophy as "domain specificity at the emirate level, not the sector level." The corpus covers the full range of sovereign capital allocation activity across ADIA, Mubadala, and ADQ's operational domains: private equity deal analysis, infrastructure project financing, public equities portfolio construction, real estate investment underwriting, and the inter-entity correspondence that governs how capital flows between the three parent institutions and their portfolio companies. A dedicated sub-corpus — assembled over 14 months by a team of three Arabic computational linguists and two Gulf investment domain specialists — covers the specific register of Abu Dhabi sovereign finance: the formal written Arabic of investment committee submissions, the semi-formal Arabic of counterparty negotiations, and the Arabic-English code-switching that characterises every board-level conversation at the three parent institutions. The total fine-tuning corpus runs to 780 million tokens. The domain annotation layer — tagging each document for register, counterparty relationship type, and negotiating context — covers 62,000 documents and required, by Al-Mansoori's estimate, approximately 19 person-months of specialist annotation work.

The total committed capital to the Al-Ihtiyati build, across infrastructure, engineering, data, and operating costs through the end of 2023, is estimated by two independent sources at between $280 million and $340 million. The three parent institutions have not disclosed this figure. The split of costs between the three parent institutions follows the equity structure of the composite vehicle: ADIA holds 40 per cent, Mubadala holds 38 per cent, and ADQ holds 22 per cent. Each institution's financial commitment runs through a separate subsidiary entity to maintain the operational distance between the composite vehicle and the parent institutions' published financial statements. No line item attributable to Al-Ihtiyati appears in any publicly filed account of any of the three parent institutions.

The National AI Strategy alignment and the institutional cover it provides

The UAE's National AI Strategy 2031 is the most explicit government AI framework in the Gulf Cooperation Council, and possibly in any major economy. It names specific targets: 10 per cent of global AI output by 2031, contribution of AED 335 billion to the UAE economy by 2031, and government services 100 per cent AI-augmented by 2031. It identifies the sovereign wealth and investment sector as a priority domain for AI integration. It calls for Arabic-language AI capability developed on UAE infrastructure. Al-Ihtiyati, as a composite vehicle of the three sovereign institutions that together manage the largest share of Abu Dhabi's investable capital, aligns with every element of this strategy — and its existence, if publicly acknowledged, would represent the strategy's most consequential single implementation. The alignment is not coincidental. Two sources with knowledge of Al-Ihtiyati's formation describe conversations between the composite vehicle's founding team and officials of the UAE's Office of Artificial Intelligence during the vehicle's structuring phase in late 2021 and early 2022. The OAI did not co-design Al-Ihtiyati, but it was consulted. The national strategy framework is, among other things, the political legitimation structure for a significant capital allocation that none of the three parent institutions wants to defend publicly on purely financial terms.

The alignment also resolves a regulatory ambiguity that would otherwise complicate Al-Ihtiyati's data governance arrangements. The UAE's data protection law — Federal Decree-Law No. 45 of 2021 — regulates the cross-entity sharing of personal data but does not specifically address the aggregation of institutional investment data across sovereign entities in a shared AI training environment. Al-Ihtiyati's legal team, which includes two secondees from Mubadala's general counsel office and an external UAE law firm whose name three sources declined to specify, has taken the position that the composite vehicle's data governance framework — which gives each parent institution control over how its data contributions affect outputs queried by the other parent institutions — satisfies the spirit of the data protection framework even in the absence of specific legislative guidance. The OAI's informal endorsement of the vehicle's structure provides, in the words of one of the seconded lawyers, "the regulatory context that allows us to proceed without waiting for a regulation that may never arrive in precisely the form we would have written it."

What to watch

Al-Ihtiyati is approximately 24 months into its operational phase. The following developments will determine whether the composite vehicle's model becomes an enduring piece of emirate AI infrastructure or a well-funded internal capability that never scales beyond its founding institutions.

• TII's Falcon 3 release and the upgrade decision it forces. The Technology Innovation Institute released Falcon 3 in December 2024, with a substantially improved architecture and a richer Arabic pretraining dataset. Al-Ihtiyati's fine-tuning investment was made on the Falcon-40B and Falcon-180B bases. Migrating to Falcon 3 would require rebuilding the fine-tuning corpus annotation layer — a 14-to-19-month project at current team scale. The decision to upgrade or to remain on the existing base is, according to one source with knowledge of the vehicle's technical roadmap, the most consequential near-term choice the team faces. Watch for any TII-Al-Ihtiyati joint research output, which would signal that the migration is underway.
• The ADQ corpus expansion. ADQ's 22 per cent equity stake in Al-Ihtiyati reflects its smaller initial data contribution relative to ADIA and Mubadala — its investment documentation archive is shorter by institutional age and narrower by asset class. As ADQ's portfolio matures and its own deal history deepens, the value of its corpus contribution to the shared model increases. Watch for any revision to Al-Ihtiyati's equity structure or any ADQ announcement about AI-augmented investment operations. Either would signal that ADQ's contribution weight is being renegotiated.
• The G42-Microsoft data governance question. Microsoft's $1.5 billion investment in G42, announced in April 2024, included provisions governing how G42's data — and the data of entities co-located on G42 infrastructure — would be treated under US data governance frameworks. Al-Ihtiyati's physical custody agreement was designed to insulate its corpus from this question. Whether that insulation is legally sufficient under US export control law — which applies to certain AI hardware and, potentially, to training data that flows through US-connected infrastructure — is a question that has not been publicly resolved. Watch for any US Commerce Department guidance on AI training data and Gulf sovereign infrastructure, which would force Al-Ihtiyati's legal team into a more exposed position than their current framework anticipates.
• A third-party deployment partnership announcement. Al-Ihtiyati's inference capability currently serves only its three parent institutions. The vehicle's governance board has discussed, according to two sources, the possibility of extending access to a small number of Abu Dhabi government entities — specifically the Abu Dhabi Department of Finance and the Abu Dhabi Investment Office, which manages the emirate's foreign direct investment promotion function. If either entity begins describing AI-augmented investment screening capabilities in public communications, it is reasonable to read that description as a reference to Al-Ihtiyati's deployment endpoint.
• Sara Al-Rashidi's first public-facing role. The data governance lead has, like the vehicle itself, maintained a profile below any public threshold. She holds no LinkedIn presence under her name and has not spoken at any conference as of the date of this piece. If she appears — at the Abu Dhabi AI Forum, at the World Government Summit, or in any published research — it will signal that Al-Ihtiyati's founding institutions have decided the vehicle's existence warrants a degree of institutional visibility. That decision, when it comes, will represent the end of the discretion economics this piece describes.

Frequently asked

Why structure this as a composite vehicle rather than assigning the mandate to one fund?

The three parent institutions — ADIA, Mubadala, and ADQ — have distinct mandates, different counterparty relationships, and, critically, different data that the model needs in order to be useful across the full range of Abu Dhabi sovereign capital activity. Assigning the build to a single institution would have produced a model optimised for that institution's data and use cases, with the other two institutions either contributing data under governance arrangements they did not control or building independent models. The composite structure keeps all three institutions inside the same model while giving each institution governance rights over how its data contributions affect outputs queried by the others. The cost is shared; the capability is unified; the sovereignty question is resolved at the emirate level rather than at the institutional level.

How does Al-Ihtiyati's build differ from simply buying enterprise AI access from a Western provider?

Enterprise AI agreements with Western providers — OpenAI, Anthropic, Google DeepMind — give a sovereign institution access to frontier model capability under contractual privacy protections. Those protections govern what the provider can do with the institution's data; they do not transfer physical custody of the model, the training data, or the inference infrastructure. Al-Ihtiyati's build transfers all three. The model weights are owned by the composite vehicle. The training corpus is owned by the parent institutions under the data governance framework. The inference hardware is owned by Al-Ihtiyati and operated by Al-Ihtiyati staff. No Western AI provider's terms of service, however protective, can replicate this arrangement. The capability gap between Al-Ihtiyati's model and a frontier provider's model is real and acknowledged internally. The governance gap runs in the opposite direction.

What is the vehicle's relationship to the UAE's National AI Strategy, and does that relationship create public accountability?

Al-Ihtiyati aligns with the National AI Strategy 2031's call for Arabic-language AI capability developed on UAE sovereign infrastructure. The Office of Artificial Intelligence was consulted informally during the vehicle's formation. That consultation does not create a formal accountability relationship. Al-Ihtiyati is a private commercial vehicle owned by three sovereign institutions — not a government programme subject to ministerial oversight or public reporting requirements. The National AI Strategy provides political legitimation for the capital allocation, not regulatory supervision of it. Whether that distinction survives closer public scrutiny as the vehicle's build and operational costs accumulate is a question that the parent institutions' communications teams have, as yet, not been required to answer.

What does the $280M–$340M committed capital actually buy, relative to what a Western AI startup would build with the same amount?

At this capital level, a Western AI startup would likely produce a frontier-competitive general-purpose model — less capable than OpenAI's most advanced systems, more capable than most open-weights alternatives. Al-Ihtiyati is buying something different: a domain-specific model with sovereign provenance, built on a proprietary corpus that no external party holds and no external model has been trained on. The general capability trade-off is accepted deliberately. What Al-Ihtiyati acquires with $280M-$340M that a Western startup cannot is the 15-year investment documentation archive of three Abu Dhabi sovereign institutions, annotated by specialists in Gulf sovereign finance communication, running on infrastructure that is under Abu Dhabi physical custody. The value of that asset is not expressed in benchmark scores. It is expressed in the confidence of the investment committee chair who can query the model on a live deal without wondering where the query log is stored.

Is there a path by which Al-Ihtiyati becomes publicly visible?

Two paths exist. The first is deliberate: the parent institutions decide that the vehicle's existence, as a demonstration of UAE AI sovereignty in practice, is worth the visibility costs. This path would likely begin with a reference in a speech by an ADIA or Mubadala executive at a forum where such disclosures carry strategic signal value — the Abu Dhabi AI Forum, the World Government Summit, or a bilateral meeting at which UAE-US technology relations are the explicit agenda. The second path is inadvertent: a disclosure through a procurement filing, a staff LinkedIn update, a vendor's client reference list, or a regulatory document that names the vehicle in a context its founders did not anticipate. The vehicle's founders are aware of both paths. The discretion they have maintained for 24 months is evidence that they consider the cost of inadvertent disclosure higher than the benefit of deliberate visibility. That calculus will change — it always does.