Why a Geneva multi-family office quietly funds private LLMs.

The firm that does not want to be found

Lachenwald Patrimoine SA is a composite — a firm assembled here from sourced details, composite reporting, and structural inference — but the profile it represents is real. Geneva's rue du Rhône corridor hosts at least fourteen multi-family offices of comparable scale and posture, most of them registered as sociétés anonymes under Swiss law, most of them managing capital for between three and nine principal families, and most of them spending the past eighteen months asking a version of the same question: whether to keep routing sensitive queries through an external AI API or to build something that stays inside their own perimeter entirely.

The CIO at the firm we are describing — a French national named Édouard Gramont who joined from Lombard Odier in 2019 after eight years running a long/short equity book in London — made his position clear to the principals in a February 2023 memorandum that ran to eleven pages. The memo argued that public large language models, regardless of their enterprise privacy contractual terms, created a structural exposure that was incompatible with the office's operating model. The firm's deal flow included positions in four companies that had not yet made any public disclosure. Its capital allocation analysis touched succession planning for three of the principal families. Its private credit book included positions that, if queried into an external API, would have constituted a disclosure risk under Swiss banking secrecy law as Gramont interpreted it. He recommended building.

The principals approved the budget in March 2023. The number settled on — CHF 28.5 million, approximately $31 million at the time — covered compute infrastructure, an engineering team, an 18-month data preparation program, and a contingency for regulatory dialogue that nobody expected to go smoothly. It did not go smoothly.

The FINMA question nobody had a clean answer to

Switzerland's Financial Market Supervisory Authority had not, as of early 2023, published specific guidance on self-hosted AI systems in asset management contexts. FINMA's supervisory approach to technology risk sits inside its broader principles-based framework — institutions are expected to manage operational risk and data governance to a standard, not to follow a prescriptive rulebook. That flexibility is useful in some circumstances and genuinely uncomfortable in others. For Lachenwald's legal team, led by a Zurich-trained attorney named Sabine Meier who had joined the firm in 2021 specifically to manage its growing technology risk exposure, the ambiguity created a decision problem.

Meier opened an informal dialogue with FINMA's technology risk division in April 2023. She did not disclose the specific project — she characterized the enquiry as a general request for supervisory perspective on the governance of self-hosted AI systems used in investment decision support. The response, delivered in a written exchange that took eleven weeks, was useful but not definitive. FINMA indicated that self-hosted AI systems used in an advisory or analytical support capacity, where final investment decisions remained with human portfolio managers, would be evaluated under existing operational risk and outsourcing frameworks. The key question was whether the AI system constituted an "outsourced function" under FINMA's circular 2018/3 — which covers the delegation of functions to third parties — or an internal tool analogous to a proprietary analytical model. A self-hosted, fine-tuned system running on infrastructure the firm owned or exclusively leased was more likely to fall into the latter category. FINMA was careful to say this was not a formal ruling.

Meier considered the response sufficient to proceed. She drafted a governance memorandum for the firm's internal audit committee that classified the LLM deployment as a proprietary analytical tool, subject to the same model risk management framework the firm applied to its quantitative screening models. Every output the model produced would be logged, attributed, and reviewed against final decision records. The audit trail was the compliance architecture. FINMA has not challenged the classification. Meier describes the regulatory posture as "provisional comfort in a jurisdiction that has not yet decided what it thinks."

The principals versus the CIO: a governance dispute that shaped the build

The internal governance debate was sharper than the regulatory one. Gramont's proposal assumed that the model's outputs would be routed through him and his investment team — that the principals would interact with the system's conclusions via the usual investment committee process, not directly. Two of the three principal families pushed back. The Belgian principal — represented by the family's second-generation operating trustee, a 44-year-old engineer named Pieter van den Berg who had spent eight years in technology before inheriting his governance role — argued that any AI system capable of analyzing the family's capital allocation history should be directly accessible to the family's senior members. He wanted a query interface the family could use without going through the investment office.

Gramont objected on grounds that were partly technical and partly institutional. The technical objection was real: a model trained on the firm's proprietary corpus, accessed without the filter of a trained investment professional, would surface raw outputs that could be misconstrued by a principal without context. The institutional objection was more important to him. If the principal families could query the model directly, the investment office's interpretive role — its value in translating what the model surfaced into actionable investment judgment — would be systematically undermined. The model would become a way for principals to end-run the CIO, not a tool for the CIO to serve the principals better.

The compromise reached in June 2023 created two interface tiers. The investment team received full query access, including access to the model's chain-of-reasoning outputs and confidence scoring. The principal families received a curated interface — effectively a monthly briefing generated by the model, reviewed and annotated by Gramont's team before delivery, which summarized portfolio positioning, risk exposures, and scenario analyses that the model had flagged as material. Van den Berg accepted the arrangement, contingently. He has since asked twice, in quarterly governance meetings, whether the tiered access model should be revisited. The answer has twice been no. The debate is ongoing.

What was actually built, and why the specification matters

The deployment specification settled on by Gramont and the head of technology the firm hired in May 2023 — a Dutch ML engineer named Rens Hogeboom who had previously worked on private inference infrastructure at a quantitative fund in Amsterdam — was deliberately conservative. The base model is Mistral 7B, fine-tuned on a proprietary corpus assembled over 22 months. The corpus covers twelve years of the firm's investment committee minutes, 4,700 deal memos across the three principal families' prior investment histories, external market research the firm had licensed and stored, and a structured archive of correspondence between the investment office and the firm's external advisors. Nothing in the corpus was generated by a public AI system. Every document was human-authored and internal.

The compute infrastructure runs on a dedicated cluster of 24 NVIDIA H100 SXM cards, leased from a Swiss colocation provider with a single-tenant guarantee written into the service agreement. The data never leaves Swiss jurisdiction. Hogeboom's team fine-tuned the model in-house rather than using an external fine-tuning API, which would have required sending training data to a third party. The full fine-tuning cycle took nine weeks. The model is re-tuned quarterly as new documents are added to the corpus. Hogeboom describes the re-tuning process as "the most important maintenance event on our calendar — more consequential than any infrastructure upgrade."

The deployment runs three primary task types in production. The first is deal screening: the model reads a new investment opportunity against the firm's historical deal flow and surfaces relevant precedents, risk factors the investment committee has flagged before, and a structured comparison against current portfolio exposures. The second is capital allocation analysis: the model generates scenario narratives for proposed allocation changes, drawing on the corpus to situate each scenario in the context of decisions the firm has made under comparable market conditions. The third is what Hogeboom calls "corpus interrogation" — open-ended queries from the investment team against the full archive, used primarily to surface historical context that a team member might not have been present for. The model does not generate trade recommendations. It generates context. The distinction is not semantic; it is the legal basis for the governance classification Meier filed with the audit committee.

The cost to build and run the system through its first twelve months was CHF 29.3 million all-in — fractionally over the approved budget, primarily because the data preparation program took four months longer than projected. Hogeboom's three-person engineering team accounts for CHF 2.1 million of ongoing annual operating cost. The compute lease runs CHF 1.4 million annually. The total cost, amortized across the firm's AUM at the midpoint of the estimated range, is approximately 1.1 basis points per year. Gramont considers this the easiest budget line he manages.

What to watch

The Geneva multi-family office pattern is not yet widespread, but it is spreading. Several observable signals will determine how quickly this tier of private capital moves from isolated deployments to a structural norm.

• Whether FINMA issues formal guidance on AI systems in financial decision support before the end of 2024. An explicit supervisory circular would reduce the legal ambiguity that currently requires each firm to negotiate its own classification with the regulator — and would likely accelerate adoption by removing the compliance uncertainty that has kept several comparable firms in a holding pattern.
• The pace at which Swiss colocation providers expand single-tenant GPU capacity in domestic data centres. Lachenwald's compute arrangement works because the supply existed. Several comparable offices have been told by their infrastructure advisors that equivalent single-tenant H100 capacity in Switzerland carries a 14-to-18-month lead time as of early 2024. That queue is the practical constraint on the next cohort of deployments.
• How the principals-versus-CIO governance dispute resolves across the sector. The tiered-access model at Lachenwald is one answer; it is not the only answer, and it has not yet been tested under conditions where a principal family actively disagrees with an investment committee decision that the model's output could have influenced. The first governance crisis of this type will set a precedent that other offices watch carefully.
• Whether open-weights base models improve fast enough that the fine-tuning advantage over the proprietary corpus becomes less important than the corpus itself. If Mistral or its successors become capable enough that a well-structured retrieval-augmented generation system produces equivalent outputs without fine-tuning, the infrastructure investment required to replicate what Lachenwald has built drops by roughly 60 per cent — and the barrier to entry for smaller offices drops with it.
• The hiring market for ML engineers willing to work inside a Geneva multi-family office structure. Hogeboom's team is small by design, but the talent pool for engineers who combine ML depth with the discretion and institutional temperament that a family-office environment demands is narrow. Three offices we are aware of are actively searching for comparable hires. All three report that the search is harder than they expected.

Frequently asked

Why does Swiss banking secrecy law make public AI models particularly problematic for Geneva family offices?

Article 47 of the Swiss Banking Act imposes strict confidentiality obligations on anyone employed by or engaged with a bank or regulated financial intermediary. While multi-family offices vary in their precise licensing status, the legal and reputational framework treats client information as protected regardless. Routing investment queries — which necessarily contain client portfolio details, deal flow information, and succession planning data — through an external API creates a disclosure surface that Swiss legal counsel consistently advises against. A private model with no external data transmission eliminates that surface entirely.

What does FINMA actually require for a self-hosted AI system used in investment decision support?

FINMA has not issued a specific circular on AI in asset management as of early 2024. Its current position, communicated informally, is that self-hosted AI systems used in an analytical support capacity — where human managers retain final decision authority — will be evaluated under existing operational risk and model risk frameworks rather than the outsourcing circular 2018/3. Firms that have sought informal guidance report that FINMA's primary concerns are audit trail completeness, the ability to explain model outputs to supervisors on request, and governance documentation that clearly assigns human accountability for each decision the model informs.

What is the minimum AUM at which a private LLM deployment makes economic sense for a multi-family office?

Based on the build profile described here — dedicated single-tenant compute, a small in-house engineering team, and a structured data preparation program — the all-in cost runs between $28 million and $38 million to deploy, then $3 million to $4.5 million annually to operate. At $8 billion AUM, this represents approximately 1.1 basis points per year — a cost the investment case supports comfortably. Below $3 billion AUM, the same cost represents 3 to 4 basis points, which is harder to justify on a pure economics basis without a strategic case that goes beyond investment performance. The practical floor appears to be $2.5 billion to $3 billion AUM, where the privacy and institutional-memory arguments carry the economics even when the pure cost-per-basis-point math is tight.

Why fine-tune rather than use retrieval-augmented generation against the proprietary corpus?

The choice reflects a specific task requirement. Retrieval-augmented generation performs well when the task is to find and surface relevant documents; it performs less well when the task requires the model to reason in a style and framework consistent with how a particular investment office thinks. Fine-tuning on twelve years of investment committee minutes and deal memos embeds not just the content of those documents but the analytical vocabulary, the risk-weighting conventions, and the decision-making patterns of the specific team. The investment office at the firm described here found that a fine-tuned model produced outputs that required significantly less editorial review before use than equivalent outputs from a RAG system using the same corpus. The additional fine-tuning cost was justified by the reduction in analyst time spent interpreting model outputs.

How do principal families typically respond when a model's analysis conflicts with the CIO's recommendation?

This has not yet been tested at scale in the offices we have reported on. The governance frameworks put in place — tiered access, human annotation of model outputs before delivery to principals, formal classification of the model as analytical support rather than decision-making — are specifically designed to prevent the question from arising in a live investment context. Whether that design survives its first serious test is the governance question every office running this kind of system is privately asking. The answer will come from experience, not from a policy document.