Inside a Singapore family office quietly funding private LLMs.

The VCC structure that made the build possible

Singapore's Variable Capital Company structure — enacted under the Variable Capital Companies Act of 2018 and administered by the Monetary Authority of Singapore — was designed to give fund managers a more flexible corporate vehicle than the conventional private limited company. What it also created, as Meridian Crest's legal counsel recognised in early 2023, is an unusually clean governance envelope for a technology deployment of this type. The VCC's sub-fund architecture allows the office to ring-fence each principal family's assets in legally distinct sub-funds while maintaining a single management entity — and a single technology infrastructure — across all three. The private LLM does not sit in any one sub-fund. It sits at the VCC management company level, accessible to the investment team but structurally separated from the beneficiaries of each sub-fund.

This matters for the regulatory conversation that followed. When Meridian Crest's chief operating officer, Tan Wei Lin, began the firm's informal dialogue with MAS in early 2023, the sub-fund architecture gave the regulator a cleaner governance picture than it typically sees. The system was not processing any single family's assets — it was processing the firm's analytical workflow. Tan used this distinction explicitly in her first written submission. MAS Technology Risk department's response, which arrived in eight weeks rather than the eleven weeks that the Geneva counterpart firm waited for FINMA, indicated that the deployment would be assessed under MAS's Technology Risk Management Guidelines rather than under the outsourcing framework MAS Notice 1117. The distinction carried the same practical weight as the Swiss decision: a self-hosted system running on exclusively leased infrastructure, with final decisions retained by human portfolio managers, looked like an internal analytical tool, not a regulated outsourced function.

The MAS dialogue had one complication the Geneva situation did not. Singapore's Personal Data Protection Act 2012 — administered separately from MAS, by the Personal Data Protection Commission — requires that personal data not be transferred outside Singapore without explicit consent or an adequacy assessment. Three of the firm's principal families have members who are Singapore permanent residents; their personal financial information sat in the corpus the firm was building. Tan's legal team spent six weeks in late 2023 working through the PDPA's "legitimate interest" exception and the business improvement exception introduced in the 2020 amendments. The conclusion: the corpus could be used for training provided that no personal data of living individuals was included without consent, and that the training occurred on infrastructure physically located within Singapore. The colocation arrangement the firm's technology lead had already negotiated satisfied the second condition. Consent letters were sent to the relevant family members in December 2023. All responded affirmatively within ten days. The corpus was cleared for use.

The CIO and the build she designed from scratch

The person who designed the deployment is not a technology hire. Priya Nair, Meridian Crest's Chief Investment Officer, joined the firm in 2021 after twelve years at GIC Private Limited, Singapore's sovereign wealth fund, where she ran a multi-strategy allocation programme covering Southeast Asian private credit and growth equity. She holds a first-class degree in computer science from the National University of Singapore — a credential she rarely mentions and did not include in the speaking biography she submitted for the Singapore Family Office Network's 2023 annual conference. The build was, by her account, something she did not expect to be doing. "I expected to hire someone to figure this out," she told a peer in a conversation recounted to this reporter. "Then I looked at what we needed and decided that nobody I could hire would understand the investment side well enough to make the right call on the data side." She hired one person: a former quantitative researcher from DBS's treasury desk named Marcus Chia, who joined Meridian Crest in April 2023 as Head of Quantitative Infrastructure. They built the system together.

The base model they chose was DeepSeek-V2, fine-tuned on a proprietary corpus that Nair and Chia spent eight months assembling. The corpus covers eleven years of the three families' prior investment history — acquisition memos, board minutes of investee companies where the families held governance seats, LP reports from the external funds in which the families are invested, and a structured archive of the office's own internal research. No external market data vendor's proprietary feed was included; the licensing terms for most third-party data preclude inclusion in a model training corpus. The corpus Nair and Chia built is entirely first-party, entirely historical, and entirely internal. It runs to approximately 340 million tokens. The fine-tuning took eleven weeks and required an infrastructure arrangement that had to be negotiated specifically for Singapore's data sovereignty context.

The ASEAN specificity that Geneva cannot replicate

The comparison to European single-family office deployments is not one Nair invites, but it is instructive. A Geneva multi-family office building a private LLM on a CHF 28 million budget is solving a specific problem: how to keep sensitive European capital allocation queries away from US-based inference endpoints, in a jurisdiction where banking secrecy law makes external API exposure legally uncomfortable. The corpus that underpins those deployments is European in provenance — French, German, Belgian, Swiss investment histories, overwhelmingly in English and French, overwhelmingly in a single legal and regulatory tradition. Nair's problem is structurally different. The three families whose histories Meridian Crest manages invested across six ASEAN jurisdictions over eleven years. Their deal memos run in English, Mandarin, Bahasa Indonesia, and Thai. Their regulatory exposures are governed by six different legal systems. Their co-investment networks involve relationships with state-linked entities in Malaysia, Indonesia, and Thailand — relationships where the framing of an analysis, not just its content, carries political weight.

The multilingual corpus requirement drove a significant portion of the build cost. Standard tokenisation for Mandarin and Thai requires different model configuration choices than a purely English corpus; Chia spent six weeks working through the tokeniser alignment problem specifically. The solution — using a multilingual tokeniser fine-tuned on a Southeast Asian language corpus before the proprietary data fine-tuning — added one full re-tuning cycle to the build timeline and SGD 620,000 to the compute cost. Nair considers it the most important non-obvious decision in the build. "If the model does not think in Bahasa and Thai as naturally as it thinks in English," she said in the same peer conversation, "it does not think about these families accurately. And if it does not think about them accurately, it is not a tool. It is a liability."

The geopolitical texture of the ASEAN context adds a layer that the European deployments do not face in the same form. Two of the three families have direct investment relationships with government-linked entities in Indonesia and Malaysia — relationships that, if the nature of the analysis being run on them became known to the counterparty governments, could create diplomatic friction. The families' senior members are acutely aware of this; one family's patriarch specifically raised the issue in the December 2023 consent conversation about the PDPA-covered data. Nair's response was to add a classification layer to the corpus that excludes any document referencing a named state-linked entity from the portions of the corpus accessible through the standard query interface. Those documents sit in a restricted sub-corpus accessible only through a separate interface that requires Nair's direct authorisation. Chia describes this as "the most consequential ten lines of code in the whole stack."

The compute arrangement that MAS did not object to

Singapore's data centre market is mature by regional standards but constrained in single-tenant GPU capacity at the scale Meridian Crest needed. The firm's infrastructure requirement — dedicated NVIDIA H100 SXM nodes, single-tenant guarantees, Singapore-sovereign jurisdiction, a colocation provider with financial services regulatory experience — narrowed the viable provider list to three names. Nair and Chia evaluated all three and chose Equinix SG4, the Tai Seng data centre that has become the default enterprise co-location address for MAS-regulated entities with data sovereignty requirements. The contract specifies a cluster of 40 H100 SXM nodes under a single-tenant arrangement, with a clause that requires Equinix to notify Meridian Crest within 24 hours of any government data access request. The compute lease runs SGD 2.1 million annually. Nair describes this as approximately one basis point of AUM. She has described it that way to all three principal families. None of them asked a follow-up question.

The total build cost came to SGD 23.4 million — SGD 1.4 million over the approved budget, driven entirely by the multilingual tokeniser work and the additional fine-tuning cycle it required. Chia's loaded annual compensation is SGD 680,000. The ongoing operating cost, including compute, Chia's team of two additional engineers hired in September 2024, and the quarterly re-tuning cycle, runs SGD 3.2 million annually. At Meridian Crest's current AUM of $3.7 billion, this is approximately 5.9 basis points per year — higher than the Geneva equivalent, and higher than Nair projected when she made the budget case to the board. She plans to bring it below 4 basis points through AUM growth rather than cost reduction. The three principal families are in active fundraising conversations with a fourth prospective member — a Malaysian real estate family whose patriarch has been a co-investor with the Indonesian family for nine years. If that family joins, the per-basis-point cost drops to approximately 4.2. The model's ASEAN corpus becomes more valuable with each family that adds to it.

What to watch

Meridian Crest's deployment is eleven months old. The signals below will determine whether this becomes a Singapore standard or remains an outlier.

• MAS guidance on AI in family office contexts. The Technology Risk Management Guidelines that govern Meridian Crest's deployment were written for banks and insurers, not for VCC management companies. MAS has indicated to at least two industry bodies that updated guidance specific to family offices and single-family offices is under review. If that guidance arrives before mid-2025 and explicitly addresses self-hosted AI systems, it will function as a compliance template that accelerates the deployment curve for the 50-plus single-family offices that MAS-regulated advisors estimate are currently in the evaluation phase. If the guidance is delayed or ambiguous, those offices will remain in the holding pattern they are now in.
• The fourth family's VCC admission. If the Malaysian real estate family completes its due diligence and joins Meridian Crest's VCC structure, the corpus expansion and cost-per-basis-point improvement will become visible in Tan Wei Lin's next annual operating report to the board. A VCC admission is a public filing event in Singapore. The structure of the new sub-fund will be legible to advisors watching the space. Watch for a new sub-fund registration at the Singapore VCC register in Q2 or Q3 2025.
• Marcus Chia's hiring activity. Chia's two September 2024 engineering hires were posted through a recruitment agency that specialises in quantitative finance technology roles, not through general AI talent platforms. The role descriptions required experience with multilingual NLP and with financial document parsing in Southeast Asian languages — a combination that narrows the candidate pool to fewer than 200 practitioners across Singapore, Hong Kong, and Sydney. If Chia posts a third role with similar specifications, the firm's intent to scale the engineering function is declared. That signal has not yet appeared.
• The Personal Data Protection Commission's enforcement posture on AI training data. Meridian Crest's consent-letter approach to the PDPA's personal data restriction is a reasonable interpretation of the 2020 amendments, but it has not been tested against a formal PDPC complaint or enforcement action. If the PDPC publishes an advisory or enforcement decision that addresses the use of personal financial data in proprietary model training — even in an unrelated context — Nair's legal team will review the corpus-access architecture. A restrictive PDPC interpretation could require corpus restructuring that sets the re-tuning schedule back by two quarters.
• The first visible output from a competing Singapore SFO. Meridian Crest is not, by the estimates of three advisors working in Singapore's family office market, the only single-family office in the city-state to have cleared a build budget for a private LLM. Two others are believed to be in the infrastructure evaluation phase as of January 2024. If one of those offices brings a system to production, the confidentiality norms of the sector mean it will not announce the fact — but it will become visible through the same indirect signals: specialist engineering hires in multilingual NLP, dedicated GPU capacity reservations at MAS-regulated colocation providers, and a quieting in the frequency with which that office's investment team asks external AI vendors for enterprise pricing.

Frequently asked

How does Singapore's MAS regulatory framework compare to FINMA's for a family office deploying a private LLM?

MAS operates under a principles-based technology risk framework — the Technology Risk Management Guidelines — that, like FINMA's approach, evaluates self-hosted AI systems against existing operational risk and outsourcing standards rather than AI-specific rules. The practical outcome for family offices is similar: a self-hosted system on exclusively leased domestic infrastructure, with final decisions retained by human portfolio managers, is more likely to be classified as an internal analytical tool than as a regulated outsourced function. The key difference is Singapore's Personal Data Protection Act, administered separately from MAS, which adds a data residency and consent requirement that has no precise Swiss equivalent. Offices building in Singapore must resolve the PDPA question as a separate regulatory step. Geneva offices face no equivalent distinct data protection authority in the private financial context.

Why does the Variable Capital Company structure matter for a private LLM deployment?

The VCC's sub-fund architecture creates a governance separation that is useful in the regulatory conversation. The LLM sits at the management company level — not inside any one family's sub-fund — which means it processes the firm's analytical workflow rather than any single beneficiary's assets. This framing gives the regulator a cleaner picture of what the system does and makes the internal-tool classification more defensible. The structure also simplifies the PDPA consent architecture: the management company holds the data processing relationship with each family's members, which means a single consent framework applies across all sub-funds rather than requiring separate consent instruments for each family.

Why did Meridian Crest choose DeepSeek-V2 over Mistral or Llama as the base model?

The multilingual corpus requirement drove the choice. DeepSeek-V2's pretraining dataset has stronger coverage of Chinese-language and Southeast Asian text than Mistral's European-language-weighted corpus or Meta's Llama-3 base, which, while multilingual, has weaker Bahasa Indonesia and Thai coverage relative to Mandarin and English. For a corpus that runs in four languages — with Mandarin and Thai representing approximately 35 per cent of the documents by token count — starting from a base with better pretraining coverage in those languages reduces the fine-tuning burden and improves output quality on the multilingual task types the office runs most frequently. The tradeoff is that DeepSeek's open-weights licensing, which prohibits use for training models deployed to more than 100 million users, required a specific legal review that the Mistral and Llama alternatives would not have. The review concluded that a single-firm deployment with fewer than 50 users was unambiguously within the license terms.

What is the practical difference between the Singapore and Geneva family office deployments in terms of what the model actually does?

Both deployments use the model primarily for deal screening, capital allocation scenario analysis, and corpus interrogation — querying the historical archive of the firm's own investment decisions. The operational difference is the governance context in which those outputs are used. The Geneva deployment runs against a single legal and regulatory tradition; its outputs land in an investment committee that operates under Swiss law and in two languages. The Singapore deployment runs against six ASEAN legal systems, four languages, and investment relationships with state-linked entities whose political context shapes how analysis is framed, not merely what it contains. The restricted sub-corpus architecture at Meridian Crest — which requires direct CIO authorisation to access documents involving named government-linked counterparties — has no equivalent in the Geneva build. It is not a technology difference. It is a consequence of the region.

At what AUM does this type of deployment make economic sense for a Singapore single-family office?

Meridian Crest's all-in cost runs to approximately 5.9 basis points of AUM annually at the current $3.7 billion level — higher than the Geneva equivalent primarily because of the multilingual corpus engineering overhead and the additional re-tuning cycle it required. The practical threshold for a Singapore deployment to be cost-defensible on a pure investment case is approximately $2 billion AUM, where equivalent costs represent around 8 basis points — still within the range that a serious privacy or institutional-memory argument can support. Below $1.5 billion, the economics require the strategic case to do a great deal of work, and most advisors in the market are counselling offices at that level toward a retrieval-augmented generation approach on a shared infrastructure before committing to the full fine-tuning build.

The desk view

Eleven months into production, Meridian Crest's model runs on 60 per cent of the deal screening reviews the investment team executes — up from 28 per cent in the first three months of deployment. Nair's usage threshold is deliberate: she considers any coverage below 50 per cent a sign that the system has not earned the team's trust on enough task types, and any coverage above 80 per cent a sign that the team has stopped applying independent judgment before relying on the output. The 60 per cent figure is, in her view, the right calibration for a system that is eleven months old and trained on eleven years of history. It means the model is useful on the tasks it was built for and absent on the tasks where human judgment is still the only adequate instrument.

The comparison to Geneva is instructive not because the two deployments are converging but because they are diverging. The multi-family office on the Quai du Mont-Blanc built a system that is excellent at making one family of problems — European capital allocation, succession, private credit — instantly queryable against a coherent historical archive. Meridian Crest is building something structurally different: a system that learns to reason across jurisdictions, across languages, and across investment relationships whose political context is not stable. The Geneva office is managing a European family's European history. Meridian Crest is managing three families' ASEAN histories across four languages and six regulatory environments. The model is not a more complex version of the same tool. It is a different instrument for a different region, with different risks baked into every layer of the build. Nair's bet is that the instrument, properly maintained, is worth the complexity. The corpus grows by approximately 290 documents per month. The fourth family is still in due diligence. The quarterly re-tuning cycle ran on schedule in January. The model is working.