The discretion economics of a Boston endowment quietly funding private LLMs.

The manager-selection problem that created the mandate

To understand why Alderton built what it built, begin with the problem that preceded the decision. Endowments at the $10 billion-plus AUM band operate an alternatives allocation model that generates a volume and complexity of manager review work that no adequately staffed investment office can process entirely through human judgment. Alderton's alternatives book runs to approximately $7.8 billion — 55 percent of total AUM — spread across private equity, venture, real assets, and a hedge fund sleeve whose composition turns over roughly 18 percent annually. The investment team that Westerfield runs has twelve people. Twelve people, $7.8 billion in alternatives, an allocation framework that requires continuous performance attribution, and an incoming pipeline of manager pitches averaging 140 per calendar quarter. The arithmetic does not balance through headcount alone.

Endowments in this position have three historical responses. The first is to use a consultant — Wilshire, Cambridge Associates, NEPC — as a pre-filtering and due-diligence augment. Alderton used Cambridge Associates for eleven years and terminated the relationship in 2021. The termination was not announced. The reason Westerfield gave internally was that the consultants' manager universe overlapped almost completely with their competitors' manager universe, and that the shared consultant relationship was generating correlated exposure across the endowment peer group rather than differentiated alpha. The second response is to build the team. Alderton attempted this in 2022, authorizing three additional investment analyst hires. Two of the three left within fourteen months. The third response — the one Alderton chose — is to build the intelligence layer.

The specific failure the private LLM was designed to address is what Westerfield's team described, in an internal memo reviewed for this piece, as the "first-pass problem." Of the 140 manager pitches arriving per quarter, approximately 95 could be eliminated within the first 72 hours of review on the basis of strategy fit, fee structure, and track record length. The remaining 45 required substantive diligence — but the initial structuring of that diligence, the mapping of the manager's stated strategy against the endowment's existing exposure, the identification of portfolio overlap, the flagging of key-person risk — consumed analyst time at a rate that left the senior investment team perpetually behind. The question Westerfield put to the board in March 2023 was not "should we use AI?" It was: "Should the intelligence that structures our manager selection live inside our perimeter, or outside it?"

The alternatives memo automation the board actually approved

The architecture Alderton Intelligence Systems deployed is not a frontier model subscription. It is a fine-tuned language model running on a dedicated cluster of 48 NVIDIA H100 cards, housed in a colocation facility in Waltham with air-gap specifications that prevent any inference output from routing through a network segment accessible to third parties. The base model is a Mistral architecture, selected over the available DeepSeek alternatives on the grounds that its European provenance and open-weight licensing created a cleaner legal foundation under the endowment's data governance policy. The fine-tuning corpus runs to approximately 14 years of Alderton's proprietary investment documentation: 4,200 manager due-diligence memos, 1,800 quarterly performance attribution reports, 680 investment committee meeting transcripts, and the complete archive of the CIO's handwritten deal annotation notes, digitized at a cost that Westerfield described internally as "embarrassingly high for OCR work."

What the model produces, in production operation since September 2023, is a standardized first-pass memo for every incoming manager pitch. The memo runs to approximately 800 words. It maps the manager's stated strategy against the endowment's existing alternatives exposure using Alderton's proprietary factor taxonomy. It flags portfolio-level correlation implications. It identifies the three most structurally comparable managers already in the book and surfaces the key differentiators. It notes fee terms against Alderton's historical negotiated range for the relevant strategy category. An analyst who previously spent two days structuring this analysis now spends forty minutes reviewing, annotating, and calibrating the model's output. The first-pass problem is not solved. It is reduced to a manageable scale.

The second application — less discussed in the internal documentation but more revealing of the long-term intent — is performance attribution summarization. Every quarter, Alderton's alternatives book generates a volume of GP reporting that the investment team previously allocated approximately three weeks to synthesize into a consolidated attribution analysis for the investment committee. The model now produces the first draft of that synthesis within six hours of the final GP report arriving. The investment team reviews, corrects, and extends it over two to three days. The CIO presents the committee with an analysis whose underlying synthesis was produced, in large part, by a model trained on her own prior analyses. The institutional memory is not being replaced. It is being accelerated.

What board-level oversight actually permitted

The Alderton board's authorization of the private LLM program in April 2023 was not routine. The investment committee — chaired by David Haverford, a former Bain Capital partner who has served on the board since 2016 — spent two sessions reviewing the program before approving it. The first session, in February 2023, focused on capability assessment. The investment team presented a side-by-side comparison of the proposed private model against Claude 2, GPT-4, and a commercial alternatives-intelligence platform. On raw analytical capability, the private model did not win. Haverford noted this directly in the session summary: "The frontier models produce better prose. They identify more factors. They are more current." The commercial platform had a richer manager database. The investment team's recommendation was not based on capability.

The second session, in March 2023, focused on what the board's governance counsel, Maria Theresa Brandt of Choate Hall and Stewart, described in her written opinion as "information fiduciary exposure." The specific risk Brandt identified was not data breach or regulatory violation. It was attribution risk: the possibility that an inference log held by a commercial AI provider could, in a future legal or regulatory proceeding, be compelled as evidence of the endowment's investment decision-making process. Endowments with university beneficiaries operate under a set of legal constraints — donor-intent compliance, prudent investor standards, conflict-of-interest documentation requirements — that create ongoing litigation exposure. A private model that processes the endowment's manager selection analysis generates no third-party inference log. There is no external counterparty holding a record of Alderton's investment reasoning. The board voted six to one to approve the program on that basis. The capability argument was secondary. The fiduciary architecture argument was dispositive.

The single dissenting vote came from the board's technology trustee, James Okoro, who argued that a $31 million commitment to private infrastructure represented an opportunity cost against the endowment's existing alternatives pacing targets. Okoro's position was not wrong on the arithmetic. It was wrong on the category. Westerfield's response — documented in the session minutes — was that the program was not an alternatives investment. It was operational infrastructure. The same category as the endowment's risk management systems, its order execution platform, its compliance monitoring software. Operational infrastructure does not require the same return-on-capital analysis as an alternatives allocation. The board accepted this framing. The program proceeded.

The long-duration intrahouse posture nobody is pricing

The decision that distinguishes Alderton's program from comparable institutional AI deployments is the one that received the least board attention: the explicit decision to hold the model's weights inside the endowment's legal perimeter indefinitely, and to treat the accumulated fine-tuning corpus as a non-transferable institutional asset. This is not the standard treatment. Most institutional technology contracts — whether cloud infrastructure, analytics platforms, or data services — include provisions that allow the institution to migrate away from the vendor. The weights of a commercially provided fine-tuned model typically remain the intellectual property of the service provider. Alderton structured Alderton Intelligence Systems as an owned subsidiary specifically to ensure that the weights produced by fine-tuning on the endowment's proprietary corpus would be owned outright, housed on owned infrastructure, and not subject to any commercial counterparty's licensing terms.

Westerfield described the strategic logic in a presentation to the endowment's advisory committee in November 2023: "Every year of operation adds to the training corpus. The model trained on two years of our decision-making is more accurate than the model trained on one year. The model trained on ten years will be more accurate than the model trained on two. We are not buying a capability. We are building an asset whose value compounds with time." The compounding logic is correct and underappreciated. A frontier model provider can improve its general capability through scale and new training data. It cannot improve its understanding of Alderton's specific manager selection framework, factor taxonomy, fee negotiation history, and investment committee preference set — because that data exists only inside Alderton's perimeter. The private model's capability gap against frontier alternatives narrows over time as domain specificity grows. The frontier model's capability advantage is real but irrelevant to the specific task.

The intrahouse posture extends to personnel. Alderton Intelligence Systems employs three people: a technical director, Thomas Reinholt, previously with the MIT Computer Science and Artificial Intelligence Laboratory; a model operations engineer who manages infrastructure and fine-tuning cycles; and a data analyst responsible for corpus quality and annotation. The team reports directly to the CIO. It does not report to the endowment's technology vendor management structure. This is structural: the intelligence layer is treated as a CIO-owned function, not an IT function. The implication is that future decisions about model expansion — whether to extend the system to cover public equities attribution, donor-portfolio analysis, or investment reporting — will be made by the investment team, not by a technology governance process. The discretion lives at the investment level. That is deliberate.

What to watch

Alderton is not alone. Watching these five developments will tell you how far and how fast the endowment private-LLM pattern replicates across the Boston institutional cluster and beyond.

• Investment office hiring patterns at endowments in the $5 billion to $20 billion AUM band. When endowments in this range begin hiring ML operations engineers and data annotation specialists into their investment teams — rather than their IT departments — the infrastructure build has begun. Watch LinkedIn title changes at the investment offices of Northeastern University, Tufts, and Boston University over the next 18 months.
• The National Association of College and University Business Officers guidance on AI system governance for endowment operations. NACUBO has not yet issued formal guidance on whether privately operated language models constitute a technology vendor relationship subject to standard procurement controls. When that guidance arrives — expected in 2025 — it will either validate the Alderton governance structure or require material restructuring. The board voted on a governance interpretation that NACUBO has not yet blessed.
• Litigation discovery requests in endowment-related proceedings. The Brandt opinion on inference-log attribution risk was prospective. If any endowment in the peer group faces a legal proceeding in which opposing counsel seeks to compel a commercial AI provider's inference logs as evidence of investment decision process, the Alderton board's governance rationale becomes immediately actionable for every peer endowment. One case would move the entire sector.
• Mistral and DeepSeek enterprise licensing terms for fine-tuning on sovereign compute. Both companies have structured their open-weight licensing to permit fine-tuning for internal institutional use without royalty. If either revises those terms — as Llama's licensing history suggests is possible — the legal foundation of programs like Alderton Intelligence Systems requires renegotiation. The open-weight licensing risk is not theoretical. It is the single most consequential external variable in the program's ten-year cost model.
• The emergence of a peer benchmarking consortium among Boston-area endowments. Three endowments in the Cambridge-Boston corridor are believed to be operating or building comparable private model programs. If they formalize a shared benchmarking framework — comparing model performance on standardized manager analysis tasks without sharing underlying weights or proprietary data — the isolation of each program ends, and the compounding advantage of intrahouse training becomes a competitive dynamic rather than a private institutional exercise. The consortium conversation is reportedly underway. It has not yet produced a structure.

Frequently asked

Why would an endowment choose a private model when enterprise contracts with frontier providers offer stronger raw capability?

The capability argument is correct and, for endowments at this AUM level, beside the point. The decision driver is information fiduciary exposure: the possibility that an inference log held by a commercial provider could be compelled in a legal or regulatory proceeding as evidence of the endowment's investment decision-making process. Endowments with university beneficiaries operate under donor-intent compliance obligations, prudent investor standards, and conflict-of-interest documentation requirements that create ongoing litigation surface. A private model eliminates the third-party inference log entirely. The privacy guarantee is architectural, not contractual.

What is the realistic cost of a private LLM program at a $10 billion-to-$15 billion endowment?

Based on the Alderton program and comparable institutional builds reviewed for this piece, the capital cost runs $28 million to $38 million over an 18-month build period. That covers dedicated compute infrastructure, the technical team, and the data digitization and annotation work required to build the fine-tuning corpus. Annual operating cost — compute, personnel, and model maintenance — runs $4.2 million to $6.1 million. At a $14 billion AUM endowment, the annual operating cost represents approximately three basis points. The investment committee framing that authorizes this most cleanly treats it as operational infrastructure, not as an alternatives allocation. Return-on-capital analysis does not apply cleanly to infrastructure.

Does the private model actually outperform frontier models on endowment-specific tasks over time?

Not at launch and not on general analytical tasks. On endowment-specific tasks — manager selection first-pass analysis, quarterly attribution synthesis, fee benchmarking against proprietary historical ranges — the private model outperforms frontier alternatives within 12 to 18 months of production operation, because those tasks are structured by the endowment's specific factor taxonomy and decision history, which exists only in the private training corpus. The gap widens with each year of operation. The frontier model's general capability advantage is permanent. The private model's domain specificity advantage is compounding. For the tasks that matter to the investment team, compounding specificity outweighs general capability within the first institutional planning horizon.

What is the governance structure that makes the board authorization defensible under fiduciary standards?

The key structural decision is categorizing the private model program as operational infrastructure rather than as a technology investment or an alternatives allocation. Infrastructure decisions at most endowments are delegated to the CIO within an approved capital budget, without requiring investment committee approval for each expenditure. Establishing a wholly owned subsidiary — as Alderton did with Alderton Intelligence Systems LLC — creates a clean legal vehicle that owns the weights, employs the technical staff, and holds the compute infrastructure, keeping the entire program inside the endowment's organizational perimeter. A formal governance opinion from outside counsel, specifically addressing the inference-log attribution risk and the data fiduciary implications, provides the board with documented justification for the approval. That opinion is the load-bearing element of the authorization structure.

How does the open-weight licensing risk affect the long-term viability of these programs?

It is the single most underpriced external risk in the current institutional private-LLM build wave. Mistral and DeepSeek currently permit fine-tuning for internal institutional use without royalty under their open-weight licenses. Meta's Llama licensing history demonstrates that terms can change materially: Llama 2's relatively permissive license was followed by Llama 3's more restrictive commercial use provisions. An endowment that has built its entire manager-selection intelligence infrastructure on a Mistral base is exposed to a future licensing revision that could require renegotiation or rebuild. The programs being built now should carry an explicit model-portability requirement: the training corpus and fine-tuning methodology must be documented at a level of specificity that allows the model to be rebuilt on an alternative open-weight base within six to nine months if the primary base's licensing becomes untenable.

The desk view

The Alderton program is not a technology experiment. It is an institutional memory project that happens to require machine learning infrastructure. Westerfield's investment thesis on private AI is not that private models are better than frontier models — they are not, and she would not claim otherwise. The thesis is that an endowment's investment decision-making history is an asset of the institution, and that routing it through a commercial inference endpoint converts a proprietary asset into a shared data environment. The fiduciary argument for not doing that is stronger than the capability argument for doing it. The board reached this conclusion before most comparable institutions had asked the question.

The second-order effect is already visible in the peer group. Two endowments in the Boston-Cambridge corridor that declined to build private infrastructure in 2023 are now revisiting the decision, not because the capability calculus has changed but because the fiduciary framing has sharpened. Governance counsel at peer institutions have begun referencing the inference-log attribution risk that Brandt identified in her Alderton opinion. The argument is spreading faster than the builds. The builds will follow. That is the read.