Should a Family Office Run Its Own AI Model? Local vs. Cloud Trade-Offs in 2026

Last updated: October 28, 2026.

Quick Answer

Whether a family office should run its own AI model comes down to how strict its no-third-party requirement is, which work it wants the AI to do, and who will look after the machine. A local model is the only option where no third party touches the data, at some cost in capability, speed, and the ability to serve several users at once. A frontier model inside the office's own cloud account keeps the strongest capability with nothing to operate, with the cloud operator still involved. A governed connection to a wealth platform such as Aleta works with either, so the choice can be revisited as models and needs change.

Key Takeaways

• A family office has four ways to run an AI model on its wealth data, rising in control and in effort: a consumer app, an enterprise tier from an AI provider, a frontier model inside the office's own cloud tenant, and a local model the office runs itself, on a workstation or on a server.

• A local model is categorically the most private option: no provider, no processor, no contract, and no usage data leaving the building. For some families that alone decides it.

• A local model costs capability, speed, and the ability to serve several users at once; it rarely costs more money. The largest open-weight models run slowly on office hardware, multi-step agent workflows and document extraction still favor hosted frontier models, and a workstation serves one user at a time.

• A frontier model deployed in the family office's own cloud tenant is near-local without the work. Amazon Bedrock, Claude in Microsoft Foundry, and Google Vertex AI run the model inside the office's own account, region, and encryption keys, with no training on its data. The cloud operator remains the one third party, and some deployments still send usage data and safety-flagged content to the model provider.

• Cost rarely decides this for a family office. Published 2026 analyses put the break-even for self-hosting at millions of tokens a day, far above what a family office uses. The decision is about privacy, capability, and who does the work.

• Whichever option a family office chooses, the connection to a wealth platform such as Aleta stays the same: inherited permissions, read by default with approval for changes, an audit log on every query, and revocation at any time. The choice is reversible.

Why Do Family Offices Ask About Local AI Models?

Family offices ask about local AI models because the first AI question in any family office is where the data goes, and a model in the building is the one answer that settles it completely.

Citi's 2026 report AI in the Family Office names data security and privacy as the primary barrier to broader AI adoption, and records that only 22% of family offices use AI for operational tasks or investment analysis. The same report finds that 57% cite a lack of internal expertise as the biggest barrier. The two findings pull in opposite directions: the family offices most worried about data are often the ones least equipped to run their own AI. The way through that tension is to understand what each option gives and what it takes.

The instinct behind the question is sound. A family office should decide where its wealth data is processed and by whom. This article lays out the four ways to do that, what each one gives up, and what each one requires, so the office can decide on its own terms. The data foundation all four depend on is the subject of our article on the state of AI in family offices.

What Are the Four Ways to Run an AI Model on Wealth Data?

A family office can run an AI model on its wealth data in four ways, and they differ in who operates the model, where prompts and answers go, and how much of the work falls on the office.

A few terms first. An AI model here means a large language model (LLM), the kind of model behind Claude, ChatGPT, and Gemini. The term LLM has stuck even though today's models also read images and documents and increasingly act as agents. A hosted model runs on someone else's computers and answers over the internet, so the family office never handles the model itself. A frontier model is one of the most capable models available at any moment, such as the latest Claude, GPT, or Gemini, and frontier models are only available hosted. An open-weight model is a model whose files its maker has published for anyone to download and run, such as OpenAI's gpt-oss, Google's Gemma, or Alibaba's Qwen. A cloud tenant is the family office's own account with a cloud provider such as Microsoft Azure, Amazon Web Services, or Google Cloud, often the same account that already holds its email and files. A workstation is a single powerful desktop computer used by one or two people; a server is a machine built to serve a whole team at once and to stay on around the clock. Running a model on either, inside the office's own premises, is what people mean by on-premises or on-prem AI.

With those in place, the four options are these. A consumer app is a general-purpose AI assistant on an individual account, where the office controls nothing beyond the account's own settings. It is also the account staff open on their own, which is the subject of our article on shadow AI in family offices. An enterprise tier is the same provider's model, hosted on the provider's own systems, under a written agreement that excludes training and limits retention. A cloud tenant deployment runs a frontier model inside the family office's own cloud account instead of on the provider's systems, so prompts and answers stay within an environment the office already administers. A local model is an open-weight model the office downloads onto its own hardware and runs there with a small desktop app; the questions are answered on that machine and nothing is sent anywhere. It comes in two sizes of ambition: a workstation for one or two users, and a server for a team.

Table 1. Ways a family office can run an AI model on wealth data

Cost figures are orders of magnitude drawn from published 2026 pricing, linked where each figure appears below.

What Does a Local Model Give You That No Hosted Option Can?

A local model is the only option in which no third party processes the family office's data, and everything else it offers follows from that.

There is no provider and therefore no contract to negotiate, no data processing agreement to read, and no terms that can change at renewal. Nothing leaves the building: not the prompts, not the answers, and not the usage data or safety-flagged content that hosted deployments typically send back to the provider. The jurisdiction question disappears, which matters to families with cross-border exposure, litigation sensitivity, or a policy against any foreign cloud provider touching family information. And the model stays exactly as the office installed it until the office decides otherwise; hosted models are retired and replaced on the provider's schedule, which is convenient until a workflow that worked last quarter behaves differently this one.

None of these are technical advantages, so technical objections do not answer them. For a family whose principal wants a categorical answer to the question of where the data goes, a local model is the only option that gives one.

What Does a Local Model Cost You?

For a family office, a local model costs capability, speed, and the ability to serve several users at once more than it costs money.

Capability first. Open-weight models have closed much of the gap for everyday work, and a 2026 self-hosting guide describes the choice for summaries, drafting, and question answering as a cost, latency, and control decision rather than a capability gap. The work a family office most wants automated is still the work that favors hosted frontier models: multi-step agent workflows, in which an AI assistant queries positions, checks commitments, reconciles two sources, and drafts an investment committee note (our article on real AI prompts family offices run shows several), and document extraction, in which a 40-page private equity capital account statement has to yield the right figures for the right entity. Independent evaluations still place the hosted frontier models at the top on reasoning depth, tool use reliability, and reading images and documents, and the pace of improvement favors them too.

Speed second. The open-weight models that come closest to frontier capability are large, and on office hardware they are slow. A workstation with 192 GB of memory can run a compressed version of a 120 billion parameter model, but at 2 to 5 tokens per second, a few words a second, which is fine for a summary and tedious for an agent making twenty tool calls. A workstation also serves one person at a time and has no backup; when it is off, the family office's AI is off.

Ownership third. The security of the machine, its patching, its model updates, and its place in the next audit are the family office's own. On a workstation this is the care any device holding wealth data already needs. On a server that serves a team it is a job, and the job has to belong to someone.

What Does a Hosted Frontier Model Give You, and What Does It Cost You?

A hosted frontier model gives a family office the strongest capability with nothing to operate, and it asks for a measure of trust that a local model never does.

What it gives is concrete: the best reasoning and tool use available, fast answers, as many users at once as the office has, models that improve every few months without anyone in the office lifting a finger, and a vendor with a contract and a support desk to hold accountable. For agent workflows on reconciled wealth data, this is where the results are best today.

What it costs is also concrete: someone else processes the data. On an enterprise tier that someone is the AI provider, under a contract that excludes training and limits retention. In the family office's own cloud tenant the arrangement is stronger. Claude in Microsoft Foundry became generally available in June 2026, and for deployments hosted on Azure, prompts and answers are processed on Azure infrastructure and stored in the Azure region the office selects, Anthropic acts as an independent data processor, and content flagged by safety systems can be sent to Anthropic for review, according to Microsoft's documentation. Amazon Bedrock states that inputs and outputs are not shared with model providers, are not used to train any model, and are stored in the AWS region where Bedrock is used. Google states that Vertex AI does not use customer data to train its models without permission, and zero data retention can be reached by turning off caching and requesting an abuse monitoring exception. These are strong commitments from cloud providers a family office may already trust with its email, files, and login system. They are still commitments from a third party, and even usage data that leaves the environment is data leaving the environment.

The family office also inherits the provider's schedule. Models are retired and replaced, prices per token change, and a workflow tuned to one model version may need adjusting when that version retires. None of this rules hosted models out, and it belongs on the same page as their advantages.

Table 2. Benefits and drawbacks of each way a family office can run an AI model

What Does Running a Local Model Actually Require?

Running a local model requires very different things at the workstation level and at the server level, and most of the difficulty attributed to local AI belongs to the second.

At the workstation level, the setup is an afternoon for a competent IT partner. The office buys a high-end workstation or a Mac Studio, installs a runtime such as LM Studio, Ollama, or Open WebUI (a desktop app that loads the model file and gives the user a chat window), and downloads a model sized to the machine's memory. OpenAI's gpt-oss-20b runs in 16 GB of memory and gpt-oss-120b fits on a single 80 GB graphics card, and the Gemma, Qwen, and Mistral families ship sizes from a few billion parameters upward. The runtime also acts as the MCP client, the piece of software that connects the model to other systems over the Model Context Protocol (MCP), the open standard that lets an AI model query a system such as a wealth platform; our MCP explainer covers it in plain language. LM Studio has supported MCP since June 2025 and Open WebUI since version 0.6.31, so connecting a local model to a wealth platform such as Aleta is only a matter of configuration. The ongoing work is the care any device holding wealth data needs: patches, backups, access control, and a decision every quarter or two about whether to move to a newer model.

At the server level, the family office is building infrastructure. A single H100, the data center graphics card most AI models run on, rents for about $2 to $3 an hour on specialized GPU clouds, roughly $1,500 to $3,000 a month for one always-on card, and costs about $25,000 to $30,000 to buy. Dedicated servers with two to four H100s run $4,000 to $10,000 a month. The largest open models need several servers working together, which is why most organizations rent those through a provider. Published analyses put the all-in cost of a production self-hosted deployment at three to five times the raw GPU price once engineering time is counted, and a DevOps engineer to run it averages $145,000 a year in the US. This tier fits a family office with a team of users, a technical owner, and a mandate; it is rarely a way to save money.

On money, one comparison settles most debates. Published 2026 analyses put the break-even for self-hosting against premium frontier pricing at roughly 2 to 3 million tokens a day, and far higher against hosted endpoints that serve the same open-weight models. A token is roughly three quarters of a word, so the lowest of those figures is well over a million words every day. A family office asking a few hundred questions a week is far below that line. A family office that runs its own model is buying privacy and independence, and recording that reason in its own decision memo keeps anyone from mistaking it for a cost decision later.

Can a Family Office Get Near-Local Privacy With a Cloud AI Model?

Yes. A frontier model deployed inside the family office's own cloud tenant gives near-local privacy without the office operating anything.

In this setup the family office owns what matters operationally: its own cloud account, its region, its encryption keys, its network path, its logs, its login system, and its invoice. The model runs inside that boundary. The office's existing security team, or its IT partner, administers it with tools it already knows, and the cloud operator, which already hosts the office's email, files, and identity, is the one third party that remains.

Amazon Bedrock is the clearest example of the architecture. AWS runs each model provider's models in separate deployment accounts, one per provider per region, that the providers themselves cannot access, so the model provider sees neither prompts nor answers nor logs. AWS states that inputs and outputs are not shared with model providers or used to train models, that content is stored in the region where Bedrock is used, that traffic can run over a private network path instead of the public internet, and that the service is in scope for SOC and ISO and is FedRAMP High authorized in AWS GovCloud. Claude in Microsoft Foundry is the Azure counterpart: generally available since June 2026, with prompts and answers processed and stored on Azure for Azure-hosted deployments and Anthropic acting as an independent data processor. Google Vertex AI states that customer data is not used for training without permission, with zero data retention reachable through configuration. Which of the three fits a family office is usually decided by where the office already lives.

One Aleta client runs Claude inside its own Azure environment and connects it to the platform through Aleta's two-way MCP layer. The office kept its cloud agreement, its region, and its freedom to change models, and it changed nothing on the platform side: the same permissions, the same audit log, the same approval step before anything is written. Another family office could make a different choice for good reasons, run a local model on a workstation, and connect it through the same layer with the same controls.

Two footnotes keep the picture accurate. The cloud operator is still a data processor, so this is near-local rather than local, and a family office with a strict no-third-party rule will still look at a local model. And some deployments send usage data and content flagged by safety systems back to the model provider; Microsoft's own documentation says so for Azure-hosted Claude models. Reading the data handling page of the specific service before deciding pays off, because the residual differences between the three live there.

How Does Aleta Handle Local and Cloud AI Models?

A governed connection to a wealth platform such as Aleta stays the same whichever way the model runs, which is what makes the model decision safe to get wrong and easy to revisit.

Aleta is a total wealth platform that delivers consolidated wealth reporting for family offices, built on an open, AI-native architecture. Its two-way MCP layer is provider-agnostic: a local model, a model in the family office's cloud tenant, and an enterprise AI assistant all connect the same way and inherit the same four controls. The AI assistant works within the user's permissions, reads by default and stops for approval before anything is written, is logged on every query, and can be revoked by the office at any time. Client data is never used to train AI models. The platform is SOC 2 Type II certified, and its data foundation, more than 100 custodian and bank integrations reconciled daily by the Data Management Team, is what any model, local or hosted, reads from.

Because the platform side is constant, the model decision is reversible. A family office can start on enterprise terms, move a workload into its cloud tenant when residency matters, run a local model for the work it wants entirely in-house, and change any of those later, all against the same connection and the same data. Owning the AI strategy means being able to make that move, and the platform's job is to make it uneventful. The two-sided model behind those controls is set out in our article on whether it is safe to connect an AI assistant to family office data, and the questions to put to each vendor are in our guide to AI vendor security due diligence questions for family offices.

How Should a Family Office Choose Between a Local and a Cloud AI Model?

The useful questions concern the family office more than the technology.

• What is the actual constraint? A regulatory or residency requirement, a family policy, a jurisdiction concern, or a preference. Name it, because each points to a different option, and a preference can be revisited while a mandate cannot.

• Which work matters? For summaries, drafts, and questions on documents, a local model does the job well today. For agent workflows and document extraction on private markets, the hosted frontier models still do it better.

• Who owns it? A workstation needs an IT partner. A server needs a person. A cloud tenant needs the team that already runs the office's cloud. An enterprise tier needs a lawyer to read the contract.

• Which data class goes where? One policy, applied by data class, can put general drafting on one tier and wealth data on another.

• What stays constant? The governed connection: permissions, logging, and approval live on the wealth platform side (Aleta's, for example) and do not move when the model does.