Inside a Dubai single-family office quietly funding private LLMs.

The DIFC structure and the governance envelope it created

The Dubai International Financial Centre's legal framework — a common-law jurisdiction operating within the UAE federal system, administered by the DIFC Courts and regulated for financial services by the DFSA — has become the preferred structuring address for South Asian diaspora capital migrating out of Mauritius and BVI holding arrangements toward a jurisdiction that offers both Western-style legal enforceability and proximity to Gulf sovereign deal flow. Arcadia is among roughly a dozen Indian-origin single-family offices that industry advisors in the DIFC estimate have crossed the $1 billion AUM threshold since 2018. It is, by the account of three advisors with indirect knowledge of its structure, the first of that cohort to have committed capital at this scale to a private AI build.

The DIFC Category 4 licence — which permits managing assets for a single client, in this case the Mehta family's principal holding vehicle — created a governance envelope that Arcadia's legal counsel, the Dubai office of a Magic Circle firm, used explicitly in its technology deployment assessment. A Category 4 firm is not a fund manager in the conventional sense. It manages a single client's assets under a discretionary mandate that is negotiated between the regulated entity and the client rather than specified in a public fund prospectus. The technology the firm deploys to execute that mandate is an internal operational matter, not a product offered to the market. This distinction — which the DFSA's consultation paper on AI in financial services, released in draft in October 2023, had not yet resolved for Category 4 entities — gave Arcadia's counsel a degree of interpretive latitude that a multi-family office or a licensed fund manager would not have had. The model, as structured, is an internal analytical tool operated by a regulated firm for its sole client. The DFSA has not been asked to approve it. The regulatory assessment, which consumed eleven weeks of legal review, concluded that it does not need to be.

The multi-jurisdictional asset reporting problem that drove the build is specific to families of this type. The Mehta portfolio spans six jurisdictions — UAE, UK, India, Mauritius, Singapore, and a residual Canadian holding that predates the family's Gulf relocation — across asset classes that include listed equities, direct real estate, private credit, co-investments in four regional private equity funds, and two operating businesses in the UAE logistics sector that the family has retained rather than monetised. Consolidated reporting across this structure, under pre-existing systems, required eleven days of reconciliation work each quarter, involving three external advisors, two prime brokers, and a custodian whose reporting systems were not designed to interoperate. Nikhil Mehta's stated motivation for the AI build, relayed by one advisor who attended the initial scoping conversation, was precise: "I want to know what we own, what it's worth, and what's changed — in one place, in one hour, by Friday morning."

The principal and the build he designed without an RFP

Nikhil Mehta is 44. He completed a master's in financial engineering at Carnegie Mellon in 2003, spent six years at Deutsche Bank's structured products desk in London, and joined his family's office full-time in 2009, the year the principal holding structure moved to Dubai. He has chaired Arcadia's investment committee since 2015. He is not a technologist by training, and the advisors who know him do not describe him as a technology enthusiast. He is, by their description, a decision-speed maximalist. The 11-day quarterly consolidation cycle was not, in his framing, a cost problem or a vendor management problem. It was a decision-velocity problem. Capital that cannot be accurately positioned cannot be intelligently deployed. The AI build was a reporting infrastructure decision presented as an investment decision, and Mehta made it the way he makes investment decisions: without an RFP, without a committee, and without asking for a second opinion on the principal question.

The person he hired to execute it is Preethi Subramaniam, who joined Arcadia in November 2022 as Head of Technology and Data after seven years at a London-based multi-family office that had built, and subsequently decommissioned, an earlier generation of AI-assisted reporting infrastructure. Subramaniam's experience with the decommissioned system — a fine-tuned BERT-family model built on a cloud inference endpoint that the office eventually abandoned after a regulatory review found it incompatible with its GDPR data-processing obligations — shaped every structural choice she made at Arcadia. The first was the insistence on fully on-premises infrastructure. The second was the choice to start from an open-weights base rather than a closed-API fine-tuning approach. The third was the decision to build the data corpus before buying the compute — a sequencing that added four months to the build timeline and saved, by Subramaniam's estimate, $4.2 million in wasted fine-tuning cycles on an insufficiently structured corpus.

The base model is Mistral 7B, fine-tuned on a proprietary corpus that Subramaniam and a three-person engineering team spent seven months assembling. The corpus covers 14 years of the Mehta portfolio's investment history: custodian statements, quarterly LP reports from the four PE fund investments, property valuation records across the UK and UAE real estate holdings, board minutes from the two UAE operating companies, and the family's own internal investment memos dating to 2009. The total corpus runs to approximately 210 million tokens — smaller than the Singapore or Riyadh builds, which reflects the portfolio's relative concentration rather than any failure of ambition. The fine-tuning completed in January 2024 and required 22 days of compute time on a cluster of 16 NVIDIA H100 SXM nodes colocated at the Equinix DX1 data centre in Dubai Internet City. The hardware is single-tenant. The colocation contract, which runs $980,000 annually, specifies a 24-hour notification requirement for any government data access request — a clause that Arcadia's counsel insisted on and Equinix accepted after two weeks of negotiation.

Decision velocity and the principal's architecture of control

The speed differential between a principal-driven single-family office and an institution — even a Gulf sovereign institution — is the defining characteristic of how Arcadia made this decision. The Mehta family's governance structure contains four decision tiers: immediate operating decisions, which Nikhil makes alone; capital deployment decisions below $25 million, which require his sign-off and notification to his father, Rajesh Mehta, the family's 72-year-old patriarch, within 48 hours; decisions between $25 million and $100 million, which require a conversation between Nikhil and Rajesh before execution; and decisions above $100 million, which require a family council of three — Nikhil, Rajesh, and Nikhil's elder sister Deepa Mehta Shah, who manages the family's philanthropic and succession planning from London. The AI build, at $31 million, sat in the third tier. The family council conversation lasted 90 minutes. The decision was made.

This velocity is not available to an Abu Dhabi sovereign vehicle. The Abu Dhabi Investment Authority's capital deployment decisions above $50 million require approval from a committee that meets on a fixed schedule; infrastructure decisions of this type carry additional review requirements from ADIA's technology governance board. The Public Investment Fund's equivalent process — as described by three people with knowledge of it — involves a minimum of four approval stages for technology infrastructure commitments above $80 million. The difference is not about capability or ambition. It is about accountability architecture. A sovereign fund is accountable to a government, which is accountable to a citizenry, which requires process. A single-family office is accountable to one principal, who is accountable to one family. The process is the conversation. The conversation took 90 minutes.

Subramaniam's query interface reflects this accountability architecture. There are three access levels. The first — available to Arcadia's four investment professionals — runs queries against the full corpus through a natural-language interface and returns consolidated portfolio positions, valuation summaries, and scenario analyses. The second, available only to Nikhil, adds a succession planning module that models the portfolio's structure across three generational transfer scenarios under UAE inheritance law, UK inheritance tax frameworks, and Indian succession law — each of which applies differently to different tranches of the family's assets. The third, available to Subramaniam herself, is the administrative interface that governs corpus updates, re-tuning schedules, and audit log review. Rajesh Mehta has not been given access to any of the three levels. This was his preference, not a restriction imposed by his son. "My father wants answers," Nikhil told a peer in a conversation relayed to this reporter. "He does not want to learn a new interface."

What the Abu Dhabi sovereign approach does differently — and why

The contrast between Arcadia's build and the UAE's sovereign-led AI infrastructure programmes is instructive in both directions. The Mohamed bin Zayed University of Artificial Intelligence — the world's first graduate university dedicated solely to AI, founded in Abu Dhabi in 2019 — has produced the Falcon family of open-weights models, including Falcon-40B and the subsequent Falcon-180B, under the Technology Innovation Institute's stewardship. G42, the Abu Dhabi AI and cloud computing group chaired by Sheikh Tahnoun bin Zayed, has built inference and cloud infrastructure at a scale that dwarfs anything a single family office can contemplate: the Falcon models run on G42's Core42 infrastructure, which operates GPU clusters that industry estimates place at between 10,000 and 18,000 H100-class nodes. ADIA's reported exploration of AI tools for investment analysis has proceeded through enterprise agreements with Western AI providers rather than a sovereign build of the Riyadh type. The Emirati approach, in other words, has been plural: sovereign model development at TII, sovereign cloud infrastructure at G42, and pragmatic enterprise procurement at the investment authority level.

None of this is directly available to Arcadia, and Subramaniam did not attempt to use it. G42's infrastructure is available to enterprise customers, but its governance structure — and the relationship between G42 and UAE national security institutions that US Congressional scrutiny in 2024 made internationally visible — created a data sovereignty concern that ran in the opposite direction to the one Arcadia was trying to solve. The family's concern is not that its data might be accessed by a Western government. It is that its data might be accessible to any government, including the UAE's own. Equinix's DIFC-adjacent colocation facility, under a single-tenant contract with a foreign company subject to US data centre governance norms, offered a cleaner answer to that concern than any domestically sovereign alternative. The choice is not a critique of the Abu Dhabi sovereign approach. It is a consequence of the specific threat model that a DIFC-resident Indian-origin family with multi-jurisdictional exposure is managing — a threat model that a sovereign fund, which is itself a government entity, does not face and cannot address.

The Falcon model family entered Subramaniam's base model evaluation and was rejected on a different ground: corpus language fit. Falcon's strongest pretraining coverage is in English and French. The Mehta portfolio's historical documentation runs predominantly in English, but approximately 18 per cent of the corpus — older acquisition memos, correspondence with Indian counterparties, records from the family's Ahmedabad operating period — is in Gujarati, Hindi, or a code-switched English-Hindi register that Subramaniam describes as "the language my father-in-law uses when he doesn't want his advisors to follow the conversation." Mistral's multilingual pretraining, while not specifically optimised for South Asian languages, outperformed Falcon on the Gujarati and Hindi document types in Subramaniam's benchmark evaluation. The margin was not large. It was sufficient.

What to watch

Arcadia's system is four months into partial production. The signals below will determine whether this becomes a template for the cohort of Indian-origin single-family offices in the DIFC — or remains an outlier built by an unusually technically capable COO with an unusually velocity-driven principal.

• The DFSA's Category 4 AI guidance. The DFSA consultation paper on AI in financial services, released in draft in October 2023, had not, as of the date of this report, issued final guidance that specifically addresses self-hosted AI deployments by Category 4 firms. If that guidance arrives before the end of 2024 and explicitly classifies internal AI tools operated for a single client as outside the outsourcing framework — the outcome that Arcadia's legal assessment anticipated — it will function as a compliance template for the eleven-plus DIFC single-family offices currently estimated to be in technology evaluation phases. A more restrictive reading would require Arcadia to conduct a formal DFSA notification and potentially restructure its data governance architecture.
• Subramaniam's second engineering hire. Arcadia's current engineering team is three people: Subramaniam and two engineers hired from the UAE's technology labour market in late 2022. The corpus's quarterly re-tuning cycle is manageable at this team size for the current corpus volume. If the succession planning module expands — as Nikhil Mehta has indicated to Deepa Mehta Shah it will — the re-tuning cadence will need to increase. A third engineering hire with specific experience in legal document NLP, targeted at the UAE and UK succession law corpus expansion, would signal that the system is being scaled beyond its original reporting mandate. Watch for a role posting through Arcadia's retained search firm, which sources with knowledge of the office identify as a regional boutique rather than a global executive search house.
• The Indian tax disclosure trigger. India's Finance Act 2022 and the subsequent FEMA amendment notifications have increased the disclosure obligations that Indian-origin non-residents with Indian-source income face when reporting to the Indian Income Tax Department. Two of the Mehta family's UAE operating businesses have revenue streams partially attributable to Indian counterparties. If the Indian tax authorities issue a notice requiring disclosure of the consolidated portfolio structure — a development that three advisors describe as a non-trivial possibility within the next 18 months — the AI system's multi-jurisdictional reporting function will face its first external compliance demand. How the system responds to that demand, and whether the corpus architecture is sufficiently structured to support it, will be a significant operational test.
• Rajesh Mehta's succession timeline. The family's patriarch is 72. The succession planning module that forms the second access tier of Arcadia's system was built, in Subramaniam's description, "for a conversation that hasn't started yet." When that conversation starts — triggered either by Rajesh's own initiative or by a health event — the system's succession scenario outputs will move from analytical exercises to documents that inform actual legal and financial decisions. The quality of those outputs, evaluated against the advice of the family's London-based legal counsel, will determine whether the succession module becomes the model's primary value driver or its most consequential liability.
• The DIFC cohort's movement. Industry advisors tracking the South Asian diaspora single-family office segment in Dubai estimate that three to five offices in the $1 billion to $3 billion AUM range are in active evaluation of private AI builds as of the first quarter of 2024. Arcadia's build is not known to any of them — the confidentiality norms of the sector and the absence of any public disclosure have kept it invisible. But the underlying drivers — multi-jurisdictional consolidation complexity, decision velocity requirements, and the data exposure concerns that public API usage creates — are identical across the cohort. If any of those offices brings a system to production in the next 12 months, the signal will appear in the same indirect indicators: specialist NLP engineering hires in the UAE technology market, dedicated GPU capacity reservations at DIFC-adjacent colocation providers, and a quieting in the frequency with which those offices engage external AI vendor sales teams.

Frequently asked

Why would a Dubai single-family office build a private LLM rather than use an enterprise agreement with a frontier provider?

The capability argument runs against building. Frontier models from Anthropic and OpenAI outperform a fine-tuned Mistral 7B on almost every general task. The driver for a family like the Mehtas is data exposure. Enterprise privacy agreements with public AI providers specify that prompt data will not be used for training — but the logs exist on infrastructure controlled by a third party, subject to that party's legal obligations in its home jurisdiction. For a family managing assets across six jurisdictions with material tax and succession sensitivities, a prompt log that exists outside the family's control is an unacceptable exposure surface regardless of the contractual protections around it. The private build eliminates the surface. No frontier model agreement can do that.

How does the DIFC's regulatory framework affect a private LLM deployment?

The DFSA's AI consultation framework, as of early 2024, has not issued final guidance specifically addressing self-hosted AI tools operated by Category 4 licensees. The operative regulatory question is whether a private LLM used exclusively for internal analytical functions constitutes a technology outsourcing arrangement under the DFSA's outsourcing rules — which would require formal notification and governance documentation — or an internal technology deployment, which does not. Arcadia's legal assessment concluded that a fully on-premises system, operating on infrastructure the firm has contracted for directly, with no third-party access to model outputs, sits outside the outsourcing framework. That assessment is legally defensible under current guidance but has not been tested against a DFSA review. The forthcoming final guidance will resolve this question for the DIFC cohort.

How does a principal-driven single-family office make this decision faster than a sovereign wealth fund?

The accountability architecture is different at every layer. A sovereign fund answers to a government, which answers to a political constituency, which requires documented process, committee review, and procurement governance that typically adds six to eighteen months to any technology infrastructure decision above $20 million. A single-family office answers to one principal. The Mehta family's decision process for a $31 million commitment involved one 90-minute conversation between three family members. That conversation is the governance. It is also the audit trail. The speed is not an accident of temperament — it is a structural property of a single-principal accountability architecture that a sovereign fund, by its nature, cannot replicate.

What is the AUM threshold below which this type of build stops making financial sense for a DIFC family office?

Arcadia's all-in build cost of $31 million represents approximately 65 basis points of its $4.8 billion AUM — a one-time cost that the family's investment committee treated as infrastructure capital rather than operating expense. The annual operating cost, including colocation, engineering team, and quarterly re-tuning, runs approximately $3.1 million, or 6.5 basis points annually. At $1 billion AUM, equivalent costs would represent roughly 31 basis points annually — still defensible on a privacy and institutional-memory argument but requiring a strategic case rather than a pure investment case to justify. Industry advisors working with the DIFC single-family office cohort place the practical threshold at $1.5 billion to $2 billion AUM, below which a retrieval-augmented generation architecture on shared infrastructure is the more economically rational path.

How does the Indian-origin diaspora context shape this build differently from a European or East Asian family office?

Three features are structurally specific. First, the multi-jurisdictional inheritance law complexity: Indian succession law, UK inheritance tax, UAE inheritance frameworks, and Mauritius trust law each apply to different asset tranches differently, and the interaction effects are not well-handled by any off-the-shelf wealth management system. Second, the language corpus problem: South Asian diaspora families with Ahmedabad, Mumbai, or Hyderabad origin periods have historical documentation in Gujarati, Marathi, Telugu, or Hindi that European or East Asian family offices do not. Third, the Indian regulatory extraterritoriality risk: India's FEMA and income tax frameworks create ongoing disclosure obligations for Indian-origin non-residents that do not apply to families of European or East Asian origin. Each of these features shaped a specific element of Arcadia's build — the succession module, the multilingual corpus architecture, and the compliance reporting function respectively. Together they make the system less a generic wealth management tool and more an instrument calibrated to a very specific family's very specific legal and cultural context.

The desk view

Four months into partial production, Arcadia's system is handling the consolidated asset reporting function that drove the build decision. The 11-day quarterly reconciliation cycle is now four hours. Nikhil Mehta receives a Friday morning summary — positions, valuations, variance against prior quarter — generated by the system and reviewed by one analyst before delivery. The deal-screening function went live in March 2024 and is currently processing approximately 40 per cent of the inbound private equity co-investment opportunities that cross Arcadia's desk. The succession planning module is built and tested but has not been used in a live family conversation. Rajesh Mehta remains unaware of its existence.

The comparison to Abu Dhabi's sovereign approach is one that Subramaniam resists but that the data supports. ADIA's investment in AI infrastructure — through enterprise agreements, through G42's ecosystem, and through MBUAI's research output — operates at a scale and with a public-good mandate that a $4.8 billion family office cannot and does not attempt to match. What Arcadia has built is not a contribution to UAE AI capability. It is a proprietary instrument for one family's specific decision-making architecture, calibrated to the family's legal exposure, language history, and accountability structure. The Gulf sovereign AI programmes are building infrastructure for populations. Arcadia built infrastructure for three people. The instrument is different because the principal is different. That distinction — between sovereign AI and private AI, between public mandate and family discretion — is the organising logic of every decision that Subramaniam made over the 14 months it took to build the system. The DIFC cohort is watching. The advisors who serve it are watching. The system, for now, is working.